164 awaiting the owner

Genuine fractional-allocation machinery with real mathematical interest — the ʿawl and radd adjustments are a beautiful over/under-subscription problem. But it is religious law, and the register's standing rule 1 (structure, never sacred content) holds it at flagged. The integrator's inclination is (a); the agent flagged rather than deciding. The kinship territory (T04) later shipped inheritance on generic parties and left this flag untouched; T07's evidence rules and T15's usury-workaround class both cite it as the precedent.

• (a) Ship with a generic-parties framing and no religious attribution, keeping only the arithmetic.
• (b) Decline, and record the decline so it is not reopened.

One catalogue entry (an assembler pass to give it a C-ID against the current 958), then a normal builder slot. Zero if declined.

Two T07 flags on the farāʾiḍ precedent. Witness-count corroboration is religious law and its mechanic is already taken by S03-08 (Quorum), so nothing is lost by holding. The Sanhedrin rule — a unanimous guilty verdict acquits, a screening rule that distrusts unanimity as evidence of a broken process — is genuinely interesting and could ship generic and unattributed if you wish.

• (a) Hold both (the default; nothing is lost on the corroboration side).
• (b) Ship the unanimity-distrust mechanic generic and unattributed, no religious framing.
• (c) Decline both and record.

(b) is one small entry — a verdict/certificate type on an existing docket grammar; the D19 cap applies.

The operations are sound and everything on the board is a token, but this is the one territory where the theme is people not being saved. sorted_once (C-744) shipped in wave 5 mechanics-first under the register tone gate; C-729 and C-734 are unbuilt. The word "expectant" is the sharpest instance. The mechanics do not depend on the framing.

• (a) Ship as written, dry and abstract, trusting the Phosphor register.
• (b) Reskin as a training exercise — tags on manikins, a documented real practice — keeping every number and draining the mortality.
• (c) Hold the family.

(a) two builder slots; (b) the same plus a copy pass on sorted_once; (c) zero.

The 1854 anomalies are public history and the entry is a genuinely great deduction. brewery_spared shipped in wave 5 with abstract-map copy and boards that never claim to be Soho. The reveal copy still cites real deaths (616; Susannah Eley by name). The territory's recommendation: ship, with the reveal written as history — but the copy deserves your eye, not an implementer's default.

• (a) Keep the reveal as history with sources, named deaths included.
• (b) Keep the history, drop the named individual, keep the count.
• (c) Keep the reveal at the level of "the brewery, the workhouse, the Hampstead case" with no counts.

A copy edit in the reveal bank plus a golden move for that type (CACHE_VERSION bump rides the next wave).

The operations are sound, no burial is depicted, and the failure verdict is "plan refused". But it is the weather file's one entry where the theme is a lethal natural hazard people actually die in. Admitted to the catalogue and held off every build list since wave 8b until this flag returns. Problem types attributed to Statham et al. 2018; terrain synthetic.

• (a) Ship as written, dry and abstract.
• (b) Reskin to a generic "thaw hazard" board keeping every number; the problem-type attribution moves to the reveal.
• (c) Hold.

One builder slot either way; (b) costs a copy pass on the rule card.

Documented as a cowrie race game dedicated to the goddess Pattini and played in a festival season. A race mechanic with no distinct operation anyway (the C-151 family), so it was not proposed; recorded OWNER so it is not reopened as a gap. The realistic outcome is a recorded decline.

• (a) Record as DECLINED (religious context plus zero new axes).
• (b) Keep at OWNER for a future cultural-advisor conversation.

Zero either way; a register edit.

The instruments are historically real and the boards are pure register arithmetic, but the theme is payouts that grow on deaths. paid_living shipped in wave 10 under option (a) with the flag riding on its brief: exits are "lapsed shares", no person imagery, reveals carry the history (Tonti 1653). C-906 Wrong Under Every Table is unbuilt and additionally states that flat-rate annuities lost money on the young.

• (a) Confirm (a): ship as written.
• (b) Reskin both to a generic "membership fund" and move the tontine entirely to reveal copy.

(a) zero; (b) a copy and golden pass on paid_living plus C-906's brief.

The player only ever breaks the corner and no manipulation is instructed. The question is what the reveal leads with. Admitted and held off the build list (wave 9b onward) until this returns — the C-887 device at its lighter weight.

• (a) Lead the reveal with the market-design lesson: why deliverable-supply rules exist.
• (b) Lead with corner history as colour (any anecdote pinned to a documented episode or omitted).

One builder slot; the choice is a reveal-bank ordering.

T15's standing rule (no tickers, no venue branding, no manipulation as skill; binds reveals and banks) was adopted at assembly and explicitly applies to six already-shipped market entries at reskin time: C-162, C-224, C-226, C-835, C-836, C-839. This is a confirmation request: the rule was adopted by the assembler and the owner has not seen it.

• (a) Confirm as adopted.
• (b) Confirm for new builds only; leave shipped types as they are.
• (c) Amend the rule's wording.

(a) an audit pass over six renderers' banks at some wave close; (b) zero.

Sixteen-cord divination is living religious practice for millions; the binary combinatorics are published mathematics. The integrator's recommendation is (b): the mechanic survives intact and the risk of trivialising a living religion drops to near zero. T12 later reaffirmed the flag untouched.

• (a) Name Ifá with an accurate, respectful attribution line.
• (b) Ship the combinatorial structure attributed generally to West African divination systems.
• (c) Decline.

An attribution line in a reveal; the entry's mechanics are unaffected.

If used, the reveal must plainly state the Shona-tradition builders. This corrects a well-documented historical erasure, so the wording matters and should not be left to an implementer. T09 (built environment) recorded the flag untouched because its subject is masonry and someone may look there for it.

• (a) Approve a one-sentence reveal line drafted for your review.
• (b) Decline the setting entirely; the masonry mechanics ship under other attributions.

One sentence of copy, reviewed.

Formally group-theoretic with a century of published literature, unlike songlines. S03's recommendation, endorsed by the integrator: ship S03-64 synthetically first (invented section names carrying the real group structure) and hold S03-63 for community review. Neither ever ships framed as "an Aboriginal kinship puzzle" — the structure is the content, the culture is not the skin.

• (a) Confirm: S03-64 synthetic now, S03-63 held.
• (b) Hold both.
• (c) Decline both.

(a) a normal builder slot on synthetic tokens.

The available sources mix astronomical and divinatory content and could not be separated to shipping precision (D54/D58). The Chinese xiu entry ships on published astronomical widths alone. T14 (weather and navigation) recorded its silence on this "so it reads as deference, not oversight".

• (a) Keep at OWNER; a future research pass starts from academic astronomy sources only.
• (b) Record as DECLINED.

(a) a research slot in a later territory; (b) a register edit.

Attribution is generic to documented Andean reciprocal-labour practice; the arithmetic model is invented and printed. The register asks for the same light naming review it applied to the reo Māori names before ship; the mechanic survives unattributed if naming fails review.

• (a) Approve naming ayni with a named source (e.g. Allen) cited in the reveal.
• (b) Ship unattributed as a generic reciprocal-labour ledger.

A citation check; no mechanics touched.

Two Cycle 3 carries. The reo Māori names are publicly taught and citable; a light cultural-advisor review of the naming was recommended before ship. The yupana and Acolhua glyph reconstructions are genuinely disputed; if used, the reveal must say so rather than assert. Related Cycle 4 stances already fixed and not reopened here: whakapapa synthetic genealogies only; rongorongo synthetic corpus and no reading ever claimed; nitus figures under generic names; pua kumbu loom geometry only.

• (a) Approve the names as written with the existing citations.
• (b) Ask for an advisor review before either type is built.

(a) zero; (b) an outside conversation the agents cannot have.

T08's rule (no real drug facts, doses, criteria or thresholds in play, ever) is stricter than R08's illustrative-values compromise and intentionally so; T16 carried it whole into assessment material. Any loosening for any entry is the owner's call. This is a confirmation request so the rule is yours and not the assembler's.

• (a) Confirm as the standing rule.
• (b) Permit R08's illustrative-values stance for named entries on request.

Zero.

pass_audit (shipped wave 8) is framed as auditor's literacy — the record audits a device, never a person — and the owner may prefer a further reframe as "test the tester" QA. C-818 ships the 1969 draft lottery as reveal history only; if you prefer, the reveal can lead with the 1970 repair (double randomisation) rather than the 1969 failure.

• (a) Keep both as assembled.
• (b) Reframe pass_audit as QA; lead C-818 with the repair.

Copy edits only.

Declined at T10 on tone (the T08 casualty precedent): every load in that file is an object. The arithmetic already ships in C-790/C-801 with scenery. The register notes the owner may revisit with a reskin if wanted.

• (a) Leave declined.
• (b) Reopen as a scenery-only load-path entry if it earns two axes against C-790/C-801.

(b) an assembler check first; likely zero axes.

C-004 was built in wave 1 as the station-chain simulator (follow_chain) per the brief, not as R05-33's cycle-following prisoners. The cycle log asks whether the original entry remains open for a second build or is closed by the one that shipped.

• (a) Close C-004 as built.
• (b) Keep the prisoners reading open as a separate entry (needs its own D39 case).

A catalogue note.

The optics territory's one OWNER-light flag. The entry asks the player to assign screen frequencies to regions of an abstract tinted panel so that a human reader at reading distance sees uniform tone and a printed sampling reader (a copier's grid) resolves a beat that spells a word like VOID — the arithmetic of moiré, emitted to a reader that must misread correctly. The class of technique is textbook and patent-public (void pantographs, band-moiré authentication; Volpe US 5,487,567, EPFL WO 2006/006063); the parameters of any live document are not, and the entry never approaches them — no banknote, cheque, certificate, ticket, licence or identity document is shown, named, implied or described, and every frequency and grid is the board's own. The question is purely one of framing: whether a puzzle whose subject is copy evidence sits comfortably in this game, or whether the same numbers should be dressed as prepress moiré avoidance. Admitted to the catalogue as C-1000★ and held off every build list until this returns (the C-887/C-911 device).

• (a) Ship as written — the reader is a copier, the demanded mark is VOID, the reveal names security printing as openly published technique (the assembler's and the territory's recommendation).
• (b) Reskin the reader as a halftone press, keeping every number, so the subject becomes prepress moiré avoidance rather than copy evidence.
• (c) Hold.

One builder slot either way; (b) costs a copy pass on the rule card and reveal. Under either option the build brief carries two standing instructions: Scrambled Indicia is a registered trademark and is described generically if at all; the intaglio latent image is named only to say it is a different physics.

D: has been critically full since wave 2 (~2 GB free at the time). Caches were cleared or offloaded to O:\rclone-offload\; the user's Downloads folder is the largest remaining cold item and only the owner can say whether it is cold.

• (a) rclone move it to O:\rclone-offload\Downloads.
• (b) Leave it; agents keep working around the pressure.

One command; a few minutes on the network.

core/**, services/**, data/**, app/components/** and three stragglers are never imported at runtime. Their tests stay green and mean nothing. Deleting them removes a standing trap for new agents and shortens the jest run; keeping them costs nothing but attention.

• (a) Delete, with the dead tests, in one commit.
• (b) Move to an attic/ directory outside tsc and jest scope.
• (c) Keep.

(a) an hour and a tsc/jest run; risk is low because nothing imports it.

Every APK so far signs with the debug keystore — fine for personal installs, not for distribution. A release keystore is a one-time act the owner has to perform (or authorise) because the key is the identity of the app on a store.

• (a) Generate a release keystore now, store it outside the repo (O: is fine), wire signingConfigs.release.
• (b) Defer until a store listing is actually planned.

Thirty minutes; a secret to keep.

The timer counts up, so haptics.timerWarning has shipped reserved and wired nowhere since Cycle 2 (the King: buzzing at elapsed 0:30 states no fact the user caused). A countdown mode would activate it. It cuts against the product's stance on pressure, and D35/D37 already cap clocked play at five types with an untimed briefing.

• (a) Build as an opt-in Settings mode for Sprint days only, with the briefing pattern.
• (b) Leave the haptic dormant; strike the idea.

(a) medium — Settings, PuzzleScreen, one haptic call, motion-law review.

Home shows only today. A seven-day tide table (the causeway floods once a day) would show the coming week's shape — which session templates fall on which days and which identity families are due — without naming puzzles. The week gets a rhythm the player can feel, and Cipher Days and Twin Days become something looked forward to. Trade-off: it spends surprise. The player who likes not knowing loses a little; the player who plans a commute gains a lot.

• (a) Templates only (Deep Dive, Sprint, Mixed Trio, Cipher Day glyphs on a seven-cell register).
• (b) Templates plus the identity family of each day's lead type.
• (c) Do not build; keep today opaque until it opens.

Small. SessionGenerator is already deterministic per date; this is a read-only Home register on the Board slab. No generator change, no CACHE_VERSION.

A mastery vibe without a streak. Reflection's constellation already ramps each type through four acquaintance steps. An inventory page lists every type as an instrument in the station's log — first met, times met, median, best — set in mono, sorted by family, with unmet instruments shown as dark glass. Trade-off: an inventory is a checklist by another name, and a checklist is a completion pressure the product has avoided. The mitigation is the register: ○ unmet, ● met, nothing counts down, nothing is "missing".

• (a) Build as a fifth screen off Reflection.
• (b) Fold into the constellation as a tap-through on a star.
• (c) Do not build.

Medium. Reads BenchRecordsStore only; one screen, one slab, a FlatList of 190 rows.

With 190 types the standing session logic meets a new type about one day in three. An opt-in "expedition" preference biases the day's type selection toward types the player has never met, until the roster is exhausted, then switches itself off. Determinism is preserved the way adaptive difficulty is: local state alters the route, never the answers (D18). Trade-off: two players on expedition days no longer share a session, which weakens the "same date, same puzzle for everyone" conversation the product otherwise supports.

• (a) Opt-in Settings switch, default off, next to adaptive difficulty.
• (b) A one-tap "expedition" from Home that applies to tomorrow only.
• (c) Do not build; let the arc do the exploring.

Small to medium. A selection bias in SessionGenerator over BenchRecordsStore's met set; a Settings row; a Sepia template.

Sepia's 48 templates observe outcomes (records, pace, family spread). The warmer observation is about method: "you take rung 3 twice as often on stream boards as on grids", "your first tap on a docket is the certificate, nine times in twelve". Specific, unflattering, true — exactly the register the vision asked for. Trade-off: it needs new store aggregates (rung use by grammar family, first-tap class), and the hallucinated-insight guardrail means each template needs a sample floor, so the lines are rare for months.

• (a) Two method templates in polish 5, with floors of 12 sessions each.
• (b) Wait until the inventory page (IDEA-025) exists and put the observations there instead of in Sepia's mouth.
• (c) Do not build.

Medium. PerformanceStore aggregates, two templates, the five bible gates.

No images, no emoji grid, no streak. One mono line the player can copy: COGNITIVE:HUB │ 2026-09-19 │ MIXED TRIO │ 3/3 │ AT PAR plus the day's named tools ("the crossing-pair certificate"). It shares what was learned, not how well. Trade-off: any share invites comparison, which the product's you-vs-you stance resists; it must be opt-in, found under the results card, never prompted.

• (a) Build via app/share/ with the tools line.
• (b) Build without the tools line (state only).
• (c) Do not build; sharing stays out.

Small. expo-clipboard and expo-sharing are already dependencies; the string derives from PUZZLE_TYPE_INFO.

The product is sound-free by law, and stream types (one card per tap, then a question) have no felt tempo. A silent metronome — the Board rail advances one ▪ per stream step, the current step in brackets — gives the stream a visible pulse and tells the player how far along they are without a number. Trade-off: motion law. The rail step must be an instantaneous state change, never a fade or a slide, and the D19 grammar census would count the rail as shared surface.

• (a) Build into kit/stream.ts's renderer contract so all stream types get it.
• (b) Trial on the halt-stream pair (settled_before, starter_ten) only.
• (c) Do not build.

Small. One rail register; a motion test.

A difficulty ladder the player climbs by choice, not by hidden bias. On Deep Dive days the briefing offers difficulty 3, 4 or 5 of the day's type with the par printed; the seed is (date, type, difficulty) so everyone who picks 5 gets the same 5. Valid answers of any one instance never change (D18); the player selects a route. Trade-off: it makes the adaptive-difficulty setting mostly redundant, and the constellation's acquaintance ramp has to decide whether a level-3 solve counts the same as a level-5.

• (a) Build for Deep Dive only, default to the session's rolled difficulty.
• (b) Build for all templates.
• (c) Do not build; keep difficulty the machine's.

Medium. Briefing UI, a seed change confined to the chosen path, records keyed by difficulty. CACHE_VERSION bump.

Reflection's constellation groups stars by the eight product categories, which D21 calls a lying map: nine of the original 31 were one narrow ability. Grouping by identity family (~90 families, ~35 nameable) would make the constellation tell the truth the visual grammar already tells — types that think alike sit together. Trade-off: 90 clusters are harder to read than 8; a two-level view (eight outstations on the horizon, families within each) keeps the world's fiction and the catalogue's truth.

• (a) Two-level: category outstation → family clusters within.
• (b) Family-only, flat.
• (c) Leave as is.

Medium. identityOf already yields the family; a layout pass on the constellation renderer and a depth-ramp check.

D11 requires contrast pairs (normal vs misère, EV vs Kelly, abstract vs social Wason) in the same session, but no template is built for it. A Twin Day template — two variants of one mechanic, the reveal naming the contrast — teaches the one thing a single puzzle cannot. Trade-off: a fifth template changes the date-hash streams, so every future session shifts for every player (a CACHE_VERSION bump handles installs, but the "same date since launch" continuity is spent once).

• (a) Add the template; accept the one-time shift.
• (b) Implement Twin Days inside Mixed Trio (two of the three slots) without a new template.
• (c) Do not build.

Medium. Template wiring, catalogue tagging of pairs, a reveal card.

The owner's Cycle 6 phrase is "well-made assets". The product refuses image assets (L4 deliberately empty). The Phosphor answer is a small set of hand-built instruments — barometer, tide gauge, lamp, balance, drum — composed from Views on the four-step solid ramp, each one a Board header ornament earned by an identity family, never chosen per type (D44/D45). Twelve instruments, each a few dozen Views, zero APK bytes. Trade-off: the grammar has no seam for taste, so the instrument-to-family mapping must be a checked-in derived table, and the amber cap (≤3 per screen) means the instruments are lamp-lit carbon, never glowing.

• (a) Commission an art pass: twelve instruments derived by family, golden-pinned like identity.
• (b) Six instruments for the six largest families only.
• (c) Do not build; the Board slab is the asset.

High for this codebase — a visual-grammar spec amendment, a solid-stage art pass, view-budget tests per instrument.

A mastery vibe needs one honest mark, not badges. After a type has been solved clean (no rung 3–4, at par) three times, its Board earns a lampBloom chip on the top edge — the 1px hard specular the theme already caps at three per screen. It is the only visible mastery mark in the app, it is tiny, and it reads as "this instrument is calibrated". Trade-off: it is a badge, and the argument that it is not is the cap and the material — if that argument fails, the idea fails.

• (a) Build as specified, three clean solves.
• (b) Build, but only visible on the inventory page (IDEA-025), never on the puzzle board.
• (c) Do not build.

Small. One derived flag from BenchRecordsStore; one chip in Board.tsx.

Adjudication 01 made the eight decommissioned outstations on the horizon the eight categories, and nothing renders them. A horizon strip on Home — eight low silhouettes in weatherFar/weatherHaze at the top of the hero, each with a lamp that brightens as the player meets more of that category's types — is the exploration vibe in one glance and derives from the code, not decoration. Trade-off: categories are a lying map (D21), so the lamps measure breadth of acquaintance, not skill, and the strip has to say nothing else. Amber budget: the lamps are wellLit, and only the station the player is at today is amber.

• (a) Build on the Home hero.
• (b) Build on Reflection above the constellation.
• (c) Do not build.

Medium. Eight silhouettes as Views, a met-count per category, a reduced-motion-safe brighten (instant state change).

The daily ritual ends on a results card and nothing is kept unless the player writes a Thought Spark. A station keeps a log. One composed line per session — 2026-09-19 · MIXED TRIO · 3/3 · the crossing-pair certificate · 0:42 under median — appended automatically to the journal as fact, never as praise, gives the ritual a written trace the player can scroll back through. Trade-off: it dilutes SparkJournalStore, which is currently the player's own words only; the log may need its own store and its own tab so the two are never confused.

• (a) A separate StationLogStore with its own card on Reflection.
• (b) Append to the journal with a ▫ marker distinguishing composed lines.
• (c) Do not build.

Small to medium. A store, a composer from existing outcome data, a list.

Three fully argued directions in DIRECTIONS.md. Each gives the app real drawn assets, a place to explore, and mastery made visible; they differ in what they spend the boldness on. The vision author's recommendation, marked as such: C + window. C's exploration is true rather than themed; the window is the single most delightful thing on the table and composes with it.

• (a) A — The Engraved Plate. Machined instrument panel in carbon: knurled bezels, dials,
• (b) B — The Tank. The Home recess is a window onto the tidal flat in three depth planes
• (c) C — The Chart. The 190 types plotted as soundings on a sea chart whose layout is
• (d) C + window. C, then B's Home window and drawn Sepia as a fourth wave. 4–5 waves.

as above; APK delta 0 (C) to +0.5–0.9 MB (A, B, C+window) before font subsetting.

"Vibe" is where ambient motion usually gets reached for. The law (UI_SPEC §9, SPEC_MOTION §1.5, D47) forbids anything the user did not cause, and it is also the app's performance strategy. The hand is already a lawful cause (attitude parallax, D49). The vision author's recommendation, marked as such: 0, plus 1 if B or C+window is chosen. Against 2: small value, large precedent.

• (a) 0 — No amendment. Motion comes from the hand (parallax), utterances (cloud), and
• (b) 1 — Widen D49 by one container. Home's window may lean with the device like the solid
• (c) 2 — A bounded "tide" clause. One element per screen may take a hard step (no tween)

see docs/cycle6/DIRECTIONS.md

Sepia ships as Variant A — 24 flat cells, no body. CHARACTER_BIBLE §7 specced a faceted body (C) as the target; it was gated on Skia, which was removed. The sample ASSET_SAMPLES/sepia_profile.svg shows what a drawn, faceted, horizontal, eye-only Sepia looks like in the palette. The vision author's recommendation, marked as such: the SVG body if C6-04 admits SVG; otherwise keep A. The Views version is not worth its ugliness.

• (a) Keep A. The strip. Zero cost, zero mascot risk, and the strongest "the guide is a
• (b) Add the faceted body (Bible variant C, as SVG polygons). Six poses as vertex tables,
• (c) Faceted body as ~40 Views (no SVG). Same figure, worse arms and tail; ~1.3 waves.

0 / 1 / 1.3 waves.

SVG was refused because Skia superseded it (ADJUDICATION_02); Skia was then removed on size (ADJUDICATION_04). The reason is gone; the consequence — no vector assets at all — remains. Expo 54 pins react-native-svg@15.12.1; Android side is Kotlin/Java with no large native library. Estimated +0.4–0.8 MB APK, to be measured at the same gate that killed Skia, and offset by subsetting the ten font files (−0.6–0.8 MB). Separately, UI_SPEC §4 bans icon fonts; a 24-glyph hand-drawn chart-mark typeface is the cheapest way to draw marks (one Text each, zero native code). Correction at filing: initTilt() is already called at app/App.tsx:62; the vision's H12f claim is stale, so the sensor path is wired. The vision author's recommendation, marked as such: 3. Fonts are this brand; a typeface of marks is native to it. SVG is the honest tool for a figure, and its cost is payable with font subsetting.

• (a) 1 — Admit SVG under containment (app/ui/assets/** + app/character/** only; never
• (b) 2 — Admit the chart-mark typeface only as a named exception to §4 (one importer,
• (c) 3 — Both.
• (d) 4 — Neither. Assets are Views and build-output PNGs. Directions A and B become

0.3 wave (containment tests + measurement) for 1; 0.2 wave for 2.

32.79 MB today; +~0.16 MB per 12-type wave. With zero assets the 34 MB gate is reached in 7–8 more waves. Cycle 6 must either buy headroom or re-set the gate. Both are measurable; neither is a taste call, which is why it is here — it decides how much art the cycle can afford. The vision author's recommendation, marked as such: 3. Headroom is cheap; a gate that is silently approaching is worse than a gate that was consciously moved.

• (a) 1 — Keep 34 MB; buy headroom now. Subset fonts (−0.6–0.8 MB), verify shrinkResources
• (b) 2 — Re-set the gate to 38 MB soft / 40 MB kill with the same measure-every-wave
• (c) 3 — Both, plus an ABI check: confirm the shipped APK is arm64-v8a only (BUILDING.md

see docs/cycle6/DIRECTIONS.md

Cycle 2 made the session "sound-free" by law; the haptic map is the tactile channel. "Fun" sometimes means a click. expo-audio@~1.1.1 is available; three tones (key, commit, verdict) can be generated procedurally by a script (<40 KB, no recorded assets). The vision author's recommendation, marked as such: 0. The haptics already say everything the tones would; the module costs more than it returns.

• (a) 0 — Stay sound-free. Zero cost. The audience plays on a silent commute.
• (b) 1 — Opt-in instrument set, default off. Three synthesised tones mirroring

see docs/cycle6/DIRECTIONS.md

The haptic map has seven entries, one of them (timerWarning) dormant. Mastery events in Cycle 6 — a depth step (SOUNDED → CHARTED), a seal struck — are rare and true, and a distinct pattern for each is the cheapest way to make mastery tactile without a pixel. The vision author's recommendation, marked as such: 1. Two entries, both caused by the player, both rare.

• (a) 0 — No additions.
• (b) 1 — Two patterns: haptics.sounding (Light impact, then Medium after 90 ms — a lead

see docs/cycle6/DIRECTIONS.md

The four-step ramp (unmet / met / timed / rated) needs a name in copy, a mark, and a home. Never rank, level, streak, XP, tier, mastered. Separately, PuzzleScreen.tsx still renders a STREAK block on the results card (timer-off sessions) and three milestone strings ("Day 30. That's a habit now." / Day 100 / Day 365), and UI_SPEC §5.1 still specs STREAK:12 — against CLAUDE.md's "no streaks" three times over. Cycle 6's vocabulary has to resolve that contradiction one way or the other. The vision author's recommendation, marked as such: the direction's lexicon, and 1 for the streak — the LOG says "you have been here 311 times" without ever saying "in a row".

- (a)

see docs/cycle6/DIRECTIONS.md

The Reflection constellation is 190 chips in a wrapped grid; it stopped being navigable somewhere past 60. Whatever replaces it must let a player arrive at any type by tap and see its record. The question under the taste question: is the layout derived from the grammar (D44 extended — types that think alike sit together, no hashing, golden-pinned) or curated by hand (someone decides which room Futoshiki is in)? The vision author's recommendation, marked as such: derived, always — a curated map of 190 types is a maintenance debt that grows twelve entries a wave and cannot be defended when two agents disagree. Structure: chart.

- (a)

structure per C6-01; derived layout ~0.3 wave for the table + golden.

Two Cycle 3 retirements were accepted and never executed: the tagline DAILY CALIBRATION FOR WORKING MINDS (→ OUTSTATION IX in the status line) and the four nav pips ● ▪ ○ ▫ (→ TODAY / LOG / DRIFT / BENCH with a 2 px amber rule under the active word). Both are ~40 lines each and make Home read as the station before any asset lands. They are here because they change words the owner sees every day. The vision author's recommendation, marked as such: 1. --- ## Machine-readable block (for the dashboard's Owner Inbox)

- (a) 1 — Ship both in wave 0. · 2 — Ship the tagline only (keep pips). · 3 — Keep as is and strike the retirements from ART_DIRECTION so the doc stops lying.

0.1 wave.

Puzzles where the answer is a gesture the tap cannot express — hold a key for a duration, press two keys at once so two thumbs tap 3 against 2 and the shared period is the clock, or drag a slider to a felt position. The atlas finds that every adjustment task already has a step-tap twin (▴▾ keys, the method of adjustment), so drag, flick and pinch are never required; hold and chord are the two gestures with no twin at all, and they are what the bimanual-coordination cluster needs.

• (a) Admit hold and chord only; drag, flick and pinch stay out because the step-tap twin covers them.
• (b) Admit the full set in a stated order (hold, chord, drag, flick, pinch), each under its own motion-law review.
• (c) Admit hold only, as a tap with a duration; chords stay out.
• (d) Decline — keep the product tap-only; the bimanual cluster stays closed and is recorded as such.

Gates atlas rows 60, 71, 73, 88 and seed P39 (gap rank 14, "bimanual chords", zero coverage today). A gesture-vocabulary amendment to SPEC_MOTION; hold is one Key state, chord is a second pointer on the same surface. Zero if declined.

Puzzles where you hold the phone at the answer — set a horizon inside a canted cockpit frame, turn the device to bring a shape into a pose the eye says is reachable and the body discovers is a mirror at 180°, or walk a route in your head and tilt toward where you started. D2 already admits attitude-as-answer with a tap twin inside a ±25° cone, snapped to a face. The atlas asks three narrower things: may one continuous attitude sample at commit be the answer (not a snapped face); may the tap twin exercise a different faculty (vector arithmetic for path integration) rather than a degraded version of the same one; and is AttitudeService (D2a, specced, never built) to be built so sensor, step-tap twin and Jest fixture are one stream.

• (a) Yes to all three: continuous sample at commit inside the cone, a different-faculty twin permitted, build AttitudeService.
• (b) Build AttitudeService and admit the continuous sample, but the twin must be the same faculty degraded, never a different one.
• (c) Build AttitudeService only; play stays face-snapped as D2 has it.
• (d) Decline — attitude parallax stays the only sensor use; the cluster stays closed.

Gates atlas rows 22–23, 40, 52, 79–81, 90 and seeds P3–P4, P21, P44 (gap rank 3, "attitude as answer": C-394 in backlog, parallax renders only). AttitudeService is the one build in this list that is already specced; app/ui/tilt.ts remains the sole sensor importer. Zero if declined.

A family of puzzles where you tap at exactly the right moment — intercept a marble at the gap, continue a beat after it stops, reproduce an interval, release on the mark — scored in millisecond bands and replayable. D35 caps the real-clock family at five and calls it full because a clock "punishes a jostle"; but the five capped entries are speed-strategy (do many things fast), not timing (do one thing at the right instant). The atlas asks whether precision timing is a separate Gp family with its own cap, inheriting D37's brief-then-timed split, and, separately, whether a ten-second timed window inside an otherwise untimed puzzle counts as a real-clock entry at all. This is the largest zero-coverage cluster in the atlas.

• (a) Rule precision timing a separate family with its own cap and the D37 split; the five-entry real-clock cap stays for rate.
• (b) As (a), and additionally rule that a bounded timed window inside an untimed puzzle is not a real-clock entry.
• (c) Rule that a timed window inside an untimed puzzle is admissible but precision timing is not a separate family (it draws on the existing cap of five).
• (d) Decline — timing is the same animal as rate; the family stays capped at five and full, and the cluster is recorded closed.

Gates atlas rows 10–11, 30, 35, 42, 53–64, 67, 72–75, 148–153, 160–161 and seeds P6–P10, P28, P30, P37, P39, P42 — gap rank 1, "motor timing as a family", 0 shipped (C-478 uses ratios, no clock). The atlas notes UI-thread timestamps already suffice, so the cost is a doctrine amendment to D35 plus a cap, not new native code. Zero if declined.

Puzzles whose content is what your eye does with motion or a flash — count dots shown for a quarter-second and then masked, say whether two frames flicker as one thing moving or two, track four of eight identical dots for five seconds, catch the second of two changes 300 ms apart. SPEC_MOTION §1.5 forbids uncaused motion; D9 rules a tap-launched simulation caused. The atlas asks how far "caused by the tap" stretches: to a one-shot exposure of at most 300 ms plus mask; to a two-frame loop that runs until the player taps stop; to several seconds of tap-launched motion. This is the second-largest zero-coverage cluster and the one every grade-B seed in §3 waits on.

• (a) Admit the one-shot brief exposure (≤300 ms, masked) only.
• (b) Admit (a) plus the looping two-frame display that runs until the player stops it.
• (c) Admit all three, including several seconds of tap-launched motion.
• (d) Decline — D9 stays at tap-launched simulation of a modelled system; perceptual dynamics and brief exposure stay closed.

Gates atlas rows 6–9, 19–22, 34, 36–37, 48–50, 55–56, 63–64, 91, 120, 158–159, 162, 164, 200 and every grade-B seed in §3 — gap rank 2 ("perceptual dynamics", rows 8, 19–21, 48–50, 63–64, 159) and rank 4 ("brief exposure", rows 6–7, 9, 34, 36, 164), both at zero coverage. A SPEC_MOTION §1.5 amendment; each admitted tier needs the reduce-motion answer in IDEA-058 and, for (a), the shutter rule in IDEA-054. Zero if declined.

Puzzles played partly through the fingertips — count pulses in the dark to learn a route's turn counts, or compare two rhythms the phone taps into your palm. Today the haptic map is a fixed feedback vocabulary of five words with set meanings (keyDown, commit, land, correct, wrong, plus the dormant timerWarning), and app/ui/haptics.ts is the only importer. Letting a type use pulses as content changes what a buzz means. The hardware caveat is real: one actuator, and Android amplitude varies by device, so counting is safer than rhythm comparison.

• (a) Admit pulses as content with a mandatory visual twin for accessibility, count and rhythm both.
• (b) Admit counting only (pulse tallies), not rhythm or amplitude comparison, given device variance; visual twin mandatory.
• (c) Decline — the haptic map stays a feedback vocabulary; the cluster is recorded closed.

Gates atlas rows 82–83, 91 and seed P36 (gap rank 11, "haptics as information", zero coverage). A haptics.ts amendment for content patterns, one visual twin per type, and a settings interaction with the existing haptics toggle. Zero if declined.

Phosphor has no hue: amber on carbon, lightness only. That admits lightness-constancy and luminance-ladder puzzles and excludes anything that needs colour — ordering hues, or naming the ink colour of a word that spells a different colour. The atlas assumes the palette law holds absolutely and asks whether a single sanctioned test hue, used only inside perceptual types and held as one token in theme.ts, would be tolerable.

• (a) Admit one sanctioned test hue as a theme.ts token, usable only by perceptual types under a grep rule.
• (b) Decline — the palette law holds absolutely; hue ordering and colour Stroop stay out (the atlas's working assumption).

Gates atlas rows 26–27 and 149 (three rows; no gap-report rank of its own). One token, one grep-rule amendment to the hex-outside-theme check. Zero if declined.

If brief exposures are admitted (IDEA-051 option a or later), a glance away on the train can cost the whole flash. The atlas's proposed resolution under D33/D34 is that the exposure is a shutter — replayable at a printed par cost, never gone — so a jostle costs par, not the daily. The question is whether this discharges D35's "punishes a jostle" concern for brief-exposure types, and whether the replay should cost anything at all.

• (a) Confirm: the exposure is a shutter, replayable at a stated par cost.
• (b) Replay at zero cost for the first replay, par cost thereafter.
• (c) Decline — brief exposure is not admitted on any terms (this closes IDEA-051 option a as well).

Rides on IDEA-051; the rows it protects are gap rank 4 ("brief exposure", rows 6–7, 9, 34, 36, 164). A par-cost rule in the hint-rung scaffold and a replay control on the Board. Zero if declined.

Puzzles that measure your own bias — set two lines to equal length under the Müller-Lyer arrows and the board reports how far your setting sat from equal, as a fact about the log. D22 forbids promising transfer; Cycle 6 §2.2 says you versus you. The question is whether a published literature mean is admissible as context beside the player's own number, or whether any external reference, even a citation, breaks the you-versus-you stance.

• (a) Admit the literature mean as one sourced line of context beside the player's number.
• (b) Player's own number only, today and over time; no external reference of any kind.
• (c) Decline the measurement copy entirely — illusion rows score the physical judgement only and never report the felt value (the IDEA-059 path A).

Gates atlas rows 25, 31, 33, 89, 134, 201–203 and seeds P26, P41 (gap rank 7, "perceptual calibration", zero coverage; step-tap adjustment already exists as Key). A D22 wording rule for "your measurement today" copy; no mechanics change.

Puzzles that quietly reward having been here yesterday — a layout that recurs so the search gets faster if you saw it before, a tempo you are asked to remember from a previous session, a sequence you learned without knowing it. A type whose difficulty depends on attendance is a streak by another name. The atlas asks whether to admit the cluster under a rule that absence is never penalised (the recurrence simply helps if you were there) or to exclude it.

• (a) Admit with the rule that absence never penalises: par is set for the cold player, the recurrence is a bonus that reads as familiarity, and nothing is ever counted "in a row".
• (b) Decline — exclude the cluster; the product's "no streaks" stance admits no attendance-dependent content.

Gates atlas rows 35, 58, 72, 109 and seeds P8, P33 (gap rank 15, "cross-day faculties", zero coverage). A session-layer hook that carries a small per-type state across days (the meta layer already spans stages on Cipher Day). Zero if declined.

Puzzles about how we read bodies — say which way a dozen moving dots are walking, or which of two glyph figures is looking at the other. Biological motion and several social quirks want a figure on the board. The atlas assumes synthetic point-light or glyph figures are admissible and anything face-like (face register, the Thatcher inversion, gaze direction from eyes) is out, and asks for that boundary to be confirmed rather than inherited.

• (a) Confirm: point-light and glyph figures in; anything face-like out.
• (b) Widen: admit schematic faces (register, Thatcher, gaze) drawn in the palette, as glyph faces.
• (c) Decline — no figures at all; the rows stay closed.

Gates atlas rows 21, 179, 185, 206–207 (five rows; no gap-report rank of its own). Zero for confirmation; (b) is a representation-register entry. The biological-motion row also depends on IDEA-051.

Cycle 6 §6.7 says every surface renders as a correct still. A perception puzzle whose content is motion has no still form. The atlas proposes that the scheduler simply skips such types under reduce-motion (the way WeeklyArc already biases selection) rather than offering a degraded still twin, and asks for confirmation or a twin requirement.

• (a) Confirm: the scheduler skips motion-content types under reduce-motion; the daily is drawn from the rest of the roster.
• (b) Require a still twin for every motion-content type before it ships.
• (c) Decline — no motion-content types at all (folds into IDEA-051 option d).

Applies to every row IDEA-051 opens, chiefly gap rank 2 ("perceptual dynamics", rows 8, 19–21, 48–50, 63–64, 159). (a) is a selection rule in SessionGenerator reading app/ui/reducedMotion.ts; (b) multiplies each such type's build cost by roughly two. Zero if declined.

House rule 2 requires an answer the generator knows. For every illusion, the physical truth is known and the felt value belongs to the player. A puzzle may score the physical judgement (which line is really longer) or report the felt value as a measurement (IDEA-055), but a type that scored the player on having the illusion — "guess how wrong you are" — would break house rule 2 and D22 both. The atlas asks for this boundary to be confirmed so nobody builds it.

• (a) Confirm the boundary: physical judgement is scored, the felt value is only ever reported, and "having the illusion" is never a scored answer.
• (b) Hold the illusion shelf until IDEA-055 settles, then confirm.
• (c) Decline the shelf — no illusion rows at all; gap rank 7 is recorded closed.

Zero for confirmation: one sentence added to the house rules. It governs the same rows as IDEA-055 (25, 31, 33, 89, 134, 201–203, P26, P41; gap rank 7, "perceptual calibration").

The illusion bench's Ponzo board — judge which of two marks is physically longer when converging lines say otherwise — landed on a lime kiln in U01's spin and a dockside crane in U06's. U01's designer notes the level crossing is "the better picture": converging rails are Ponzo, barrier arms are Müller-Lyer fins, one apparatus carries both illusions. The kiln is the wheel's honest answer; the crossing was reached by declining a fusion seed (P26) that landed on it. Mechanics identical under all three.

• (a) Kiln — the wheel's answer as spun.
• (b) Dockside crane — U06's spin, already the skin of the merged entry.
• (c) Level crossing — the picture that carries both illusions; a rails-and-barriers board.

Zero — a skin choice on one level of one entry. The entry builds under any answer; the assembler recommends (b) by default so nothing waits.

The intuitive-physics boards reveal a documented common belief — "half of adults think the ball keeps curving", "34% reasoning, 95% imagining". U04 argues that a proportion of a named study's sample is a fact about that study, not a claim about the player, and that "half of McCloskey's undergraduates" is more honest than "many people". Cycle 6 §2.2's you-versus-you stance may exclude any external reference all the same. This is adjacent to IDEA-055 but distinct: IDEA-055 asks whether a population figure may sit beside the player's own number; this asks whether a study's proportion may appear in a reveal that prints nothing about the player at all.

• (a) Allow published proportions in reveals, attributed to study and sample, never to "people".
• (b) Name the phenomenon and its source, print no number.
• (c) Neither — reveals describe what the board did and stop.

Zero; every U04 entry is written to survive (b). Governs reveal copy on C-1014, C-1015, C-1029★, C-1064, C-1066, C-1077 and the water_level fold; VERIFY rows in the register quote the figures exactly if (a).

Twenty-four intuitive-physics boards share one shape — a forecast slot, a RUN key, a trace ghost, a still end-state — on one renderer kit. U04 proposes that the shape be a WeeklyArc.affinity tag with a one-per-week cap, so that a player never meets two forecast boards in a week regardless of how different the physics underneath is; five sub-families (impetus-run, statics-forecast, level-scale, count-reveal, calibrate-first) are already adopted as never-two-in-a-week. The top-level cap is a product-pacing decision: it makes the physics cluster rare on purpose, and it is the D80 remedy for fourteen of thirty boards sharing one nominate-item surface.

• (a) Cap predict-run at one per week across all 24 (U04's proposal).
• (b) Cap the five sub-families only (already adopted); no top-level cap — up to five forecast boards a week.
• (c) Cap at two per week as a middle setting.

One affinity tag and one cap in WeeklyArc; zero generator work. Under (a) the cluster surfaces about fifty times a year; under (b) up to five times a week.

A sounding is the felt-value readout of a perception or timing entry — set the fins-in shaft until it looks equal and the board files how far off you were, as a fact about the log. Part Thirteen admits soundings only as unscored riders on their scored hosts (C-1053, C-1059, C-1090, the U03 calibration lines), never as entries, and their copy waits on IDEA-055. Where they appear is a separate question. U06 offers three placements and recommends the first; U03's timing reveals and U05's log lines assume the same shape without naming it.

• (a) An optional phase offered after the scored type on the day it is seated — skippable, never required to finish a daily, its readout kept on the Bench.
• (b) A Cipher-day texture — soundings appear only inside the meta-finale's trials.
• (c) A Bench station — soundings live in Settings/Bench and never inside a daily.
• (d) No soundings — the scored boards ship, the felt value is never taken.

(a) is a post-commit optional phase on the Board plus a BenchRecordsStore row per sounding; (b) rides Cipher Day's existing stage layer; (c) is a Bench surface only. Zero under (d). Every host builds under any answer.

Two files independently designed the same unscored surface: tap the centre of twelve small pins and the Bench prints your mean landing offset and spread — "your rate card today". U06 and U03 both note that the offset vector is exactly what a per-device tap correction would consume, and both recommend it as a later Settings item rather than a puzzle. Turning it on would shift every renderer's hit-test by the measured offset. Fitts's law is adopted as a design constant for target sizes and is never a score.

• (a) Ship the Bench sounding only; print the rate card, correct nothing.
• (b) Ship the sounding and an opt-in "correct my taps" setting that applies the measured offset to hit-testing across the app.
• (c) Decline the sounding — the Bench stays a records surface.

(a) one Bench station and a BenchRecordsStore row. (b) additionally touches every renderer's hit-test through Key — one shared change, but a global one, and it interacts with the reduce-motion and haptics toggles' settings surface. Zero under (c).

Part Thirteen carries thirteen held entries whose interaction grammar does not exist in the product — tap-in-time (7), tap-at-moment (1), chord (2), hold-release (1), set-attitude (2). U03 put the finding in one line: "the new verbs are the rulings." Ruling yes on gestures (IDEA-048), attitude (IDEA-049) or precision timing (IDEA-050) is not only admitting a faculty; it is admitting a grammar family into the D74 census, and every family in the census carries a one-per-week cap and a Key state. The census counts the five apart today and asks that, when any of the three rulings is made, the cap and the Key state be decided in the same breath rather than discovered at build.

• (a) Confirm: any yes on IDEA-048/049/050 names the verb, its cap and its Key state together (U03's proposed ruling sentences are in its §7).
• (b) Rule the faculties first and let the assembler propose caps at the next census.
• (c) No new verbs — every held entry ships only as its step-tap twin or not at all (this is the "decline" branch of all three rulings taken together).

Zero now. Under (a), three sentences of doctrine per ruling. Under (c), the seventeen Gp/Gk/Gh/Ga-U8 seats stay catalogue-only.

app/ui/grammar/facts.ts holds nineteen ability codes; the six faculty files tag entries with seven it does not hold. Part Thirteen seats those entries under the files' codes (relabelling would hide exactly the duplication D21 exists to expose), so today no builder can declare their facts. This is a law question because the Ability union feeds identityOf and the identity golden (tests/__golden__/identities.json). U01 proposes a zero-identity-impact ruling for the visual three: every Gv ability already rules arc, so Gv-IL, Gv-LE and Gv-CF enter RULING_OF_ABILITY as arc and no existing identity moves. The non-visual four (Gp, Gk, Gh, Ga-U8) have no glyph; U03 notes Gt's is "the least-wrong" precedent (C-478) but does not claim it. U04's Gv-MA was declined at assembly (mental animation is Gv-Vz).

• (a) Add all seven now; the Gv three rule arc; the non-visual four reuse Gt's glyph until a ruling on IDEA-050/052 gives them their own — zero identity impact, golden untouched.
• (b) Add the Gv three only (they have eleven buildable entries today); defer the four until their rulings land.
• (c) Add all seven and give the non-visual four a new glyph now — the identity golden regenerates deliberately.
• (d) Do not extend; tag the eleven Gv-IL/LE/CF entries Gv-CS and record the loss in the census.

(a)/(b) a union extension and RULING_OF_ABILITY rows with no golden change; (c) a deliberate golden:regen; (d) zero code and a permanent census blind spot.

U02's wheel landed three boards on themes the designer calls thin — the spool (C-1032, a rotation between two poses), the clerk (C-1120, a modelled reader with a memory span), the drum (C-1038★, one pose for two observers) — and offers a respin under U02-NN-r1 if the owner prefers; the recommendation is to ship as written. Separately, C-1034 Whose Left Was It is a railway signalling board; the file forbids any reference to a real accident, company or line and offers a harbour-pilot port/starboard skin if a non-railway board is preferred.

• (a) Ship all four as written.
• (b) Respin the three thin themes; keep the signal box.
• (c) Respin the three and take the harbour-pilot skin for C-1034.

Zero — skins only; mechanics identical under every option.

Eleven illusion-bench entries put the bias inside a printed reader the player builds against — a vial reader, a crane driver who hears the first weight, a tower operator whose report leans on the last. U06 writes them as instrument design, not deception, and notes that where a reader is person-shaped it is a printed rule with no mind. The question is whether the themes stay on the board or every reader is labelled THE READER with the trade confined to the reveal. Recommendation: ship with the theme.

• (a) Ship with the trade themes; person-shaped readers stay printed rules.
• (b) Name every reader THE READER on the board; the theme lives in the reveal only.

Zero; a labelling choice on C-1079…C-1091.

The T20 litter and carrying-pole boards put bearers under the poles — Every Bearer Under the Bar sets hinges so each bearer's share lands in band, Swap Without Setting Down rotates a crew through stations, and Over the Wheel's level-4 litter skin puts four under two poles. Every bearer is a token with a capacity: no face, no body, no name, no caste or hereditary occupation (the kahar as a caste term is not used), and the Chinese rank-by-bearer-count fact is reveal history only. The file's question is whether a human should be implied at all, or whether the tokens should read as posts a pole rests on. Recommendation: (a) — the practice is the point.

• (a) Ship as written — crew tokens with capacities on a pole; the reveal names palanquin, doli, kago and the whippletree as documented practice.
• (b) Reskin the bearers as stations (posts a pole rests on) so no human is implied; every number, band and hinge is unchanged, and the litter attribution moves wholly into the reveal.

Zero — a token glyph and three reveal sentences; mechanics identical under both.

Stand It, Then Wear It attributes the pack frame to the Korean jige, attested by name from a 1690 book, with the tumpline (mecapal, namlo) as the strap form. The source also records that American soldiers dubbed the frame the "A-frame" and that it carried supplies and ammunition in 1950–53. The file omits the line by default on the T17 gunnery-tone precedent, applied lightly: nothing on the board is a weapon, but the sentence is military colour a farming tool does not need. The question is only whether the reveal may carry it as one labelled historical sentence.

• (a) Omit — the reveal stops at the 1690 attestation and the tumpline forms (the default).
• (b) Include one sentence, labelled as history, with the source cited; no board text changes.

Zero — one reveal sentence either way.

You asked for "look around a 3D world and identify things". The spike found the APK at 33.13 MB against the 34 MB gate (0.87 MB headroom). Skia (+9.7 MB) and WebGPU (+17 MB) are out on size alone. The live choices are a View-drawn diorama (free, snap-look only, about 50 objects), expo-gl with a small in-house renderer (+0.9–1.0 MB), or expo-gl + three.js (+1.8–2.2 MB). A font-loading fix already queued (the APK ships 33 font files; the app loads 10) should recover about 1.8 MB, which puts the in-house renderer back under 34 MB with room to spare. Whatever you pick, the first explorable-room puzzle (look_room) is built renderer-independent: the scene is pure data and the verifier runs without a GPU, so the renderer is a one-file swap. Recommendation: (b).

• (a) Diorama only — the existing layered Views, 0 MB, snap look-around, no free camera.
• (b) Fonts fix, then expo-gl with an in-house WebGL2 renderer (600–900 lines) — free camera, low-poly rooms, stays under the 34 MB gate.
• (c) expo-gl + three.js (+ react-three-fiber) with the APK gate re-set to 38 MB soft / 40 MB kill — the richest scenes and tooling, gate moves.
• (d) Filament (native PBR 3D) — only if a measured build fits inside 40 MB; no web support; highest build risk on this machine.

(a) nothing; (b) about one agent-wave after the fonts fix; (c) one wave plus a gate change; (d) a measured spike before anything else. A Phase-2 measured build (sizes and cold start for each variant) is scripted and runs when the build machine is free.

Reading React Native 0.81's Android source, the spike found that it drops the skew part of a View's transform matrix. The app's solid faces use exactly that skew, so on your phone the dimetric boxes may be drawing as flat rectangles rather than slanted faces. The web build renders them correctly, which is why no gate has caught it. This comes from reading source code and has not been seen on a device. One look at any board with solid blocks (or the SolidLab screen) on the phone settles it. If it is real, the fix costs no APK bytes (two nested Views per face) but needs a ruling on whether a computed rotation inside a face breaks the no-rotateZ design law.

• (a) Check on the phone and report what the blocks look like; the fix follows if they are flat.
• (b) Rule now that computed rotations inside solid faces are allowed (the law keeps forbidding decorative rotation), so the fix can ship without waiting for the check.
• (c) Leave it until the 3D renderer choice (IDEA-071) replaces View solids anyway.

A glance at the phone; the fix is about half an agent-wave and halves the View ceiling for solids.

Part Fifteen admits 22 washing-day boards — the copper and the reused water, the rinse split three ways, the mangle, the line and its shadow, the sad irons, the laundry mark. No board depicts a washer; every board moves tubs, garments, pegs and irons. But the history is plain in the sources: laundry was overwhelmingly women's work, often servants', and in some times and places enslaved people's ("The job of doing the laundry was reserved for women"; the Great Wash as work "often undertaken by poor servants and washerwomen"). The question is only what the reveals say. The file recommends (a): leaving the labour out makes washing day sound like a hobby. The Catalan idiom fer safareig ("do the laundry", meaning gossip) is not used under either option, because it trades on a gendered stereotype.

• (a) Reveals state it once, plainly, as history, where the practice is named (the recommendation).
• (b) Reveals omit labour history entirely and describe only the process.

Zero — a sentence in a handful of reveals; every board is identical under both options.

C-1169 Return Every Bundle is a memory board: learn which mark belongs to which house, sort the washing, then return every item from memory. Its reveal names the laundry mark and the duplicate-ticket system that the Chinese hand laundries of the United States used — a documented, neutral practice, and the reason tickets existed (memory alone fails at return). The history around those laundries also includes discriminatory licensing ordinances († Yick Wo v. Hopkins, 1886, recalled, not yet verified). The file recommends (a): the practice is the point, and the legal history deserves more than one reveal line can give it. The flag was raised on T21-12, which was cut at assembly; the reveal and this question moved to C-1169.

• (a) Name the ticket system and stop (the recommendation).
• (b) Name it with one line of the legal history, cited once verified.
• (c) Omit Chinese laundries and describe laundry marks generically.

Zero — one or two reveal sentences; the board is identical under every option.

The open-air washing ghat at Mahalaxmi, "constructed in 1890" with "rows of open-air concrete wash pens, each fitted with its own flogging stone", is the obvious real-world picture for C-1157 Rinse Above, Wash Below and C-1149 Tub, Then Mangle. It is also a living, organised workforce, and the place name contains the word dhobi, which names a scheduled caste and which the register already bars as a label for people. The dabbawala precedent allows naming a living workforce's practice accurately without reproducing its internal system. The file recommends (a).

• (a) Never name it (the recommendation).
• (b) Name the site as a place in one reveal, with its construction date and nothing about the people.

Zero — one reveal sentence or none.

You spent two minutes chopping virtual wood and said it could be a puzzle. It is now two: C-1170 Read the Round (read the end grain — checks, knots, the other face, the rings — and deliver the stove order in the fewest strikes) and C-1179★ Split for the Stack (the same round, but the pieces must also build a woodpile with square ends). Both passed the duplication test on their own and sit first and second on Part Sixteen's build list because you asked for them. The question is whether they stand alone or anchor a small family with a shared look — the stack that grows beside the block during a puzzle, and related yard and workshop boards later.

• (a) Two boards only; no family look.
• (b) A "yard work" family: the woodpile boards share one visual identity and the growing stack, and later yard or workshop boards join it where they pass dedup.
• (c) Read the Round alone; Split for the Stack waits until the first has been played on a phone.

(a) nothing extra; (b) a shared board asset and a family tag in the scheduler (never two in one week); (c) nothing now.

Every U07 board ships today as taps. Twenty of them also have a richer form that waits on a ruling you have not yet made: fifteen want gestures (IDEA-048 — swipe the axe, drag the mower, flick the stone, twist the sofa), five want timing (IDEA-050 — the top-of-swing band on the axe, the live swing, the live ripple walk), and a few want longer motion (IDEA-051c). Your two minutes of chopping suggest the gesture is part of the pleasure. This is not the ruling itself; it asks which boards you would want to feel first, so the rulings can be judged on real examples. A gesture never changes what is being decided; where timing is added, the score gains a motor term and the file says so. The stone-skipping flick is the one case where the decision moves into the hand, and it would be judged as a new puzzle if timing opens.

• (a) Gestures first — rule IDEA-048 with the woodpile, broom and mower as the test boards.
• (b) Timing first — rule IDEA-050 with the axe's swing band and the live footsteps game (cut today, revived if timing opens) as the test boards.
• (c) Both together, on the woodpile only, as one experiment.
• (d) Neither for now; the tap forms are enough.

(a)–(c) a ruling plus a build of the chosen riders (no new C-IDs; each rider lands on an existing entry); (d) nothing.

C-1172 is the pleasure of clearing a full row, made into a planning puzzle: the whole stream of pieces is printed in a fixed order, nothing falls on a timer, there is no hold and no randomness, and you choose each piece's column and turn so the rows clear. It uses no name, no familiar piece colours and not the famous seven-piece set — its pieces are its own mixed shapes. It still looks like a famous commercial game's genre. It is admitted to the catalogue but held off the build list until you answer.

• (a) Comfortable — build it as written.
• (b) Build it with a different dress (the file's own spin is courses of stones in a fish weir, where a watertight course "clears").
• (c) Too close — leave it in the catalogue unbuilt.

(a) nothing extra; (b) a skin, same mechanics; (c) nothing.

You asked for a "silly morning games" feel. The voice law forbids exclamation marks and shouting, not silliness. The file proposes deadpan premises stated flatly — "The hammers are hiding again." "Today's order: four logs and some kindling, for a stove that is fussy." — and one-word verdicts that name the act (Split. Swept. Stacked. Pressed.). The question is how far that goes.

• (a) Deadpan premises and act-verdicts on the enjoyment boards only.
• (b) The same, extended to briefings across the whole roster where a board has a physical act.
• (c) Keep premises strictly informational; silliness lives only in art and motion.

Copy only — a pass over briefing strings; no mechanics change.

The cuttlefish guide could carry one dry line per activity family, never cheering: "Checks are splits that started without you." "The pan was always going to win." The file offers it as a small source of the morning vibe; the question is whether Sepia should speak that often.

• (a) One line per activity family, shown once when the family first appears.
• (b) One line per family, rotating on each visit.
• (c) Quieter — Sepia stays as she is.

Copy and one app/character/ hook; no mechanics.

You asked for "a sense of progress … if I want to improve this part, I can do more puzzles here". Cycle 6 wave C1 is building that as the Chart — the per-area progress model, the Chart screen with a "practice here" run, and today's leg on Home. U07 offers a second metaphor for the same data: a small cupboard with one shelf per ability area (spatial fills with split logs, reasoning with stamped cards, memory with jars), where finishing a puzzle adds one object to its area's shelf. Shelves never lose objects, never show a missed day and never compare you with anyone; they count puzzles finished, not "ability gained". This is an option to sit alongside the Chart, not a replacement for it.

• (a) Chart only, as C1 is building it.
• (b) Chart plus the shelf as a small view of the same per-area counts (on Home or on the record card), reading from C1's progress model.
• (c) The shelf inside the Chart: each bank's objects shown at its station.
• (d) Revisit after C1 ships and has been used for a week.

(a) nothing; (b) one small view and object art per area, no new data model; (c) art plus layout inside the Chart; (d) nothing now.

The day's puzzles could appear on Home as a row of bubbles; finishing a puzzle pops one (a single frame and a light tick). Leftover bubbles can be popped for nothing — a fidget, not a score. The risk is plain: a row that fills day by day starts to feel like a streak, which the product refuses. The question is whether the idea is worth that risk.

• (a) Try it with guards: today's puzzles only, no history, nothing carried to tomorrow.
• (b) Only the free-popping fidget, with no link to puzzles at all.
• (c) Decline.

One Home component and a haptic call on the existing selection word.

You asked to "look around a 3D world and identify things". Today's renderer can draw a cut-away room of about twenty cube-built objects and show it from each of four walls, like looking into a dollhouse. Part Sixteen admits five boards built that way — count the hammers from every wall, spot what moved on the stall after you turned, put the room back to match a photo, find five things behind, under and inside, and walk a warehouse so every aisle has been seen. None waits on anything. Standing inside the room and turning your head needs the real 3D renderer that IDEA-071 is deciding. This asks whether the dollhouse view is enough of what you meant, or whether the first-person form is the point.

• (a) The dollhouse view is enough; build the diorama boards as they are.
• (b) Build them now, and add first-person skins later if IDEA-071 brings a renderer.
• (c) First-person is the point; hold the diorama boards until the renderer exists.

(a)/(b) nothing beyond the planned builds (the phone check in IDEA-072 applies to every solid-stage board); (c) the diorama boards wait on IDEA-071's outcome.

C-111 Stroke Order asks you to work out, from which stroke crosses which, the order the strokes were drawn in. Lifting sticks from a pick-up-sticks pile without disturbing the others is the same operation: a stick can only come off once nothing lies on it. The file calls it "the most enjoyable skin that operation has ever had". The mechanics are identical; only the dress changes.

• (a) Adopt pick-up sticks as C-111's skin when it is built.
• (b) Offer both dresses, strokes and sticks, as alternate banks.
• (c) Keep strokes only.

A skin on a backlog entry; nothing until C-111 is scheduled.

The U-series files pick each board's setting by spinning a wheel of 114 themes, so that settings are varied rather than chosen by habit. In U07 the activity itself is what you asked for, and nineteen of the thirty-one spun proposals landed on settings that do not fit the activity (a farrier's shoe for mowing stripes, a Jacquard card chain for moving furniture). The file dressed those boards in the activity and recorded the spin. The question is which you prefer for enjoyment-axis boards from now on.

• (a) Keep the wheel's awkwardness where it can be made to fit, as serendipity.
• (b) Let the activity always dress itself on enjoyment-axis boards; spin only the props.
• (c) Respin the thin ones.

Zero — skins only; mechanics identical under every option.

The Chart's four depths are UNCHARTED → SIGHTED (solved once) → SOUNDED (solved without a hint) → CHARTED (three hint-free solves). Direction C as written made SOUNDED a timed hint-free solve and CHARTED "a median exists" — but the timer is off by default, so most players could never deepen a station, which breaks the Chart's own line "Nothing here is locked". The builder decoupled it: hint-free solves count whether timed or not, and times appear as records beside the depth rather than gating it. This is live now; it needs your sign-off or reversal.

• (a) Keep as built — depth counts hint-free solves, timed or not; times are records only.
• (b) Restore the spec — SOUNDED and CHARTED require timed solves; untimed players stop at SIGHTED.
• (c) Middle — SOUNDED counts any hint-free solve; CHARTED requires one timed solve among the three.

Zero to small — a rule in app/progress/mastery.ts and its tests.

You asked for "a sense of progress … if I want to improve this part, I can do more puzzles here". The Chart gives exactly that per area, and deliberately never shows a whole-chart percentage (a test enforces it), so it cannot become a completion bar. Even so, a map of dots that light up can pull toward grinding. Worth a look on your phone once the next APK lands (screenshots of the web build were sent 2026-09-24). The next UI wave (C2) is adding depth, distinct marks for the four states, per-station record cards and small one-time moments when a station deepens.

• (a) Keep per-area counts only, no overall figure (as built).
• (b) Also show a single quiet overall line ("41 of 229 stations sounded") in the Chart header.
• (c) Soften further — hide counts on the map itself and show them only inside the area card.

Zero to small.

Part Seventeen admits 20 tea-and-coffee boards and merges a round-table turntable skin into the shipped slow_way_round (serve a list of requests at the least turning). Two of the practices behind them carry a documented courtesy of rank: Arabic coffee is served first to "the oldest or most significant guest", and the yum cha turntable is turned so "the oldest person at the table has the opportunity to have the first serving". No board uses it as a mechanic — every board fixes a direction of service instead — so the question is only whether reveals mention it. The file recommends (a): a ranking of people is not a puzzle mechanic in this product, and a reveal that mentions it invites the reading that it is.

• (a) Boards fix a direction of service and reveals do not mention precedence (the recommendation).
• (b) Reveals mention the courtesy once, as history; boards unchanged.

Zero — at most one reveal sentence on the slow_way_round turntable skin; every board is identical under both options.

C-1218 Kopi-O Kosong is a memory board: learn the coffee-shop order code of Singapore and Malaysia in five slots (milk, sugar, strength, temperature, base), then name the drinks that come up and build the drinks that are called, with the code hidden. Kopi O is black, C is evaporated milk, kosong is no sugar, siew dai less, gao thick, poh thin, peng iced — real Hokkien, Cantonese and Malay words, verified against their source and stable for decades. It is the catalogue's first memory board whose associations compose, and a player who learns the five slots rather than the drinks pays no penalty. The file recommends (a): the real code is the point, and it is secular, everyday and verified. The source's note that C may come from a brand name is never repeated under either option.

• (a) Ship the real code, with each word's source language in the reveal (the recommendation).
• (b) A synthetic code of the board's own, with the kopi counter named in the reveal only.

Zero either way — a ten-row curated bank under (a), a generated one under (b); the mechanics are identical.

C-1217 The Guest Tea asks you to learn a few sample teas as visual profiles, then match a row of cups to them — one of which is a tea you never sampled, built to look like one you did. Its real ancestor is tōcha, a medieval Japanese contest of identifying teas by taste, which was also a gambling game with prizes and was banned in 1336. The wager is declined outright in the register under the standing gambling rule. What remains open is only the name. The file recommends (a): the structure is generic blind tasting, and naming the contest while leaving out what it was mostly known for would be an odd half-history.

• (a) Blind tasting described generically; no name (the recommendation).
• (b) Name tōcha in the reveal as a medieval Japanese tea-identifying contest, without its wagering (the source would be fetched and quoted first).

Zero — one reveal sentence or none; the board is identical under both options.

Tea and coffee reached most of the cultures in Part Seventeen — the samovar, the Turkish double pot, the kopi counter, South Indian filter coffee, the British milk-first argument — through trade and plantation economies with a history of coerced and indentured labour. The boards are about serving: pots, cups, glasses and ice, with no estate, region of production or cultivar named anywhere. The question is whether reveals that name a practice's origin should say one plain line about how the leaf or bean got there. The file recommends (a): the territory is serving, not growing, and one line cannot carry that history honestly. Latte art's source also names a café and a barista; neither is used under either option.

• (a) No reveal mentions production history (the recommendation).
• (b) One plain line of history where a practice's origin is named, cited once verified.

Zero — a sentence in a few reveals or none; every board is identical under both options.

Part Eighteen admits 19 mending boards. One of them, C-1237 Twenty-Six Inches Is Five Sizes, now carries a second bank folded in from the research file: a drawer of screws to fit to a limited roll of drivers, learnt by head mark (slot, plain cross, cross with four ticks, cross with a dot, square, six-lobe). The trap is structural: the plain cross driver fits the ticked and dotted screws and strips them when tight. Most people know these drives by trade or inventor names — Phillips, Pozidriv, Robertson, Torx. The file recommends (a): the puzzle is the visible mark, not the name, and generic names keep trademarks and people's names off the board.

• (a) Generic names on the board and in the reveal alike — cross, ticked cross, dotted cross, square, six-lobe, slot (the recommendation).
• (b) Generic names on the board; the familiar trade and inventor names in the reveal, as history.

Zero — one reveal paragraph differs; the board is identical under both options.

C-1237 Twenty-Six Inches Is Five Sizes is a memory board built on Sheldon Brown's observation that tyre sizes alias: tyres called "26 inch" sit on five different rims (559, 571, 584, 590 and 597 mm), one rim has several names ("700C", "28 inch", "29 inch" are all 622), and "mathematically equal" decimal and fractional sizes do not fit each other. You learn the table, then fit a bench of tyres to rims with the table hidden. Its second bank uses the Japanese JIS cross-recess screw, often marked with a small dot, which a plain cross driver will damage. Both vocabularies are real, verified against their sources and stable for decades. The file recommends (a): the aliasing is the point, and a made-up system would lose the reason the board exists.

• (a) Ship the real ISO, inch and French tyre names and the JIS mark, with the sources named in the reveal (the recommendation).
• (b) A synthetic sizing system of the board's own, with the real one named in the reveal only.

Zero either way — two six-row curated banks under (a), generated ones under (b); the mechanics are identical. The maker-coded row in the source's table is not used under either option.

The research territory takes its title from Make-Do and Mend, a Board of Trade booklet of 1943 issued under clothes rationing, with ways to mend, darn and patch garments. The one board that would name it is the everted-sleeve level folded into C-1140 Which Face Shows the Flower (place the patch on the flat sleeve so it covers the wear seen inside out). No board depends on it. The file recommends (a): one plain line of history where the booklet is named, nothing more — it is the documented source of the phrase, and it is neither sentimental nor a claim about anyone.

• (a) One plain line of history where the booklet is named, nothing more (the recommendation).
• (b) No wartime reference; "make do and mend" read as an everyday phrase.

Zero — one reveal sentence or none; every board is identical under both options.

Household mending, like laundry in the washing-day boards, was mostly women's unpaid work, and boro — the patched cloth that Japanese sashiko reinforced — was the necessity of poor rural households. Part Eighteen's boards move cloth, thread and stitches and never depict a mender; the question is only whether a reveal that names darning or boro should say one plain line of social history. The file recommends (a): describe the practices and do not gender or romanticise them — one line cannot carry that history honestly, and poverty should not be made quaint.

• (a) Reveals describe the practices, without gendering or romanticising them (the recommendation).
• (b) One plain line of social history where boro or darning is named, cited once verified.

Zero — a sentence in a few reveals or none; every board is identical under both options.

C-1236 Every Screw Back in Its Hole is a visual-memory board: a device comes apart stage by stage, you lay each screw in a tray however you like, and then you put every screw back in its own hole, with near-identical screws one length step apart. Its setting is the community repair day, a movement that began as the Repair Café in Amsterdam on 18 October 2009 (Martine Postma). The name belongs to a foundation and may be a registered mark. The file recommends (a): "repair days" generically, no name — the board needs the bench, not the brand.

• (a) "Repair days" generically; no name (the recommendation).
• (b) Name the Repair Café in the reveal, with its founding date, as history.

Zero — one reveal sentence differs; the board is identical under both options. The movement's venue count is not used under either option.

Phosphor is amber on carbon with no other hues, so the chopping-wood round, the tea boards, the washing-day boards and the diorama rooms are all drawn in greys. After the polish pass the log reads as end grain (bark, rings, cracks, knots) and the rooms read as rooms, but the wood is grey wood and the tea is grey tea — the "silly morning games" warmth you asked for is harder to reach without colour. This is distinct from IDEA-053, which asks about a single test hue for perception puzzles. The question here is whether a small, fixed set of desaturated material tones (e.g. three wood tones, one tea, one linen) may appear inside puzzle boards only, never in chrome, never for meaning a colourblind player would lose.

• (a) Keep Phosphor absolute — boards stay greyscale; warmth comes from shape, motion and copy.
• (b) Admit a small material palette (≤6 desaturated tokens in theme.ts, boards only, never carrying information the grey form lacks), applied first to wood and tea.
• (c) Amber-tinted materials only — wood and tea drawn in dim amber ramps (accentDim family), no new hues; spends the amber budget the AmberSlice law caps.

(a) nothing; (b) about half an agent-wave plus a theme-law amendment and test updates; (c) about half a wave, but it competes with amber's "one voice" rule.

Part Nineteen admits 20 boards built on the figures of social dances — the square, the contra line, the Scottish reel, the Cuban rueda, the line dance, the ballroom. Almost every figure has two roles. They were historically called gents and ladies; many dance communities now use "leads and follows" or "larks and robins", partly because each pair has the same number of syllables. The boards themselves draw the roles as glyphs (● and ○), so this is only about the words in copy and reveals. The file recommends (a): glyphs on the board, "lead / follow" in copy, and larks and robins named once in a reveal as the practice it is.

• (a) Glyphs on the board (● lead, ○ follow); "lead / follow" in copy; larks and robins named once in a reveal (the recommendation).
• (b) "Larks / robins" throughout.
• (c) The historical terms in reveals only, never on a board.

Zero — a word choice in copy; every board is identical under all three options.

Four square-dance boards and the rueda board run on calls: do-si-do, allemande, promenade, right and left grand, ladies' chain, and in Cuba dame ("give me") and dame dos. One board, C-1257 Called From the Card, teaches twelve calls at intake and then asks you to dance a called sequence with the card hidden — its trap is that the do-si-do and the allemande look busy and move nobody. The names are real, public folk vocabulary; each call's effect is written by the board, and no society's definitions are quoted. The file recommends (a): retrieving the real names is that board's point, with "ladies' chain" drawn by role glyph rather than the word.

• (a) Ship the real call names, "ladies' chain" rendered by role glyph (the recommendation).
• (b) Synthetic call names on the board, the real ones named in the reveal.

Zero either way — one twelve-row bank under (a), a generated one under (b); the mechanics are identical.

This is IDEA-057 (figures without faces) applied to a whole territory of people moving together. Every T24 board draws dancers as tokens with a role glyph and a facing tick; the richer loop forms (the set dancing a call, held on IDEA-051(b)) would move them as point-light tokens. The question is whether a dance board may ever draw a body. The file recommends (a): tokens only — the figures read cleanly as tokens, and nothing in any board needs a body.

• (a) Tokens with role glyphs only; point-light motion on the loop riders (the recommendation).
• (b) Simple silhouettes, never faces.

(a) nothing beyond the planned renderers; (b) a silhouette glyph set and a review against IDEA-057.

Some social dances carry names from history. The Paul Jones — two circles moving opposite ways until the music stops, then you dance with whoever is in front of you — is named after a naval officer, and a favourite Irish céilí dance is named after a siege. The one admitted board that uses such a dance is C-1247 Dame, Dame Dos, whose level 4 is the two-circle mixer with a printed stop (never random). The file recommends (a): generic names on the board ("the two-circle mixer"), and the familiar name in one reveal line.

• (a) Generic names on boards; the familiar names in a reveal line (the recommendation).
• (b) No names at all.

Zero — one reveal sentence or none; the board is identical under both options.

Competitive ballroom ranks couples by the skating system, introduced in 1937: a place goes to the couple a majority of judges put there or better, with ties broken by summing that majority's marks. It avoids a known trap of averaging — marks of 1-1-1-1-7 average 2.2, yet no judge ranked that couple second. The file proposed a board (fill the last judge's column so the system gives a printed order) and recommended shipping it. The assembler cut it: it is the fourth in the file's own cut order, a competition rule rather than a social figure, and it repeats the operation of a catalogued ballot board and the shipped Same Table, Three Champions. Choosing (a) reopens it for a later pass; choosing (b) confirms the cut.

• (a) Ship it as a rule board with Dawson and 1937 in the reveal and no festival named (the file's recommendation; would reopen the cut entry).
• (b) Leave it cut.

(a) one L board at a later assembly, re-argued against C-723 and same_table; (b) nothing.

C-1243 Step When the Poles Open is built on tinikling: two bamboo poles tapped apart on two beats and brought together on the third, and a dancer who steps between them without being caught. It is widely called the national dance of the Philippines; the source lists that status as unofficial. The file recommends (a): "a Philippine folk dance from Leyte", with no national claim — accurate, and it avoids asserting an official status that does not exist.

• (a) "A Philippine folk dance from Leyte"; no national claim (the recommendation).
• (b) Mention the popular "national dance" label together with its unofficial status.

Zero — one reveal phrase; the board is identical under both options.

Part Twenty admits 23 boards built on how people grade produce by eye: the candler's egg grades AA, A and B; the tomato's breakers, turning and pink; Kenya's coffee screen grades AA, AB and PB; rice's head rice and second heads; the honey colour classes from water white to dark amber; the wool classer's band of three neighbouring quality numbers. These are real trade words from published standards. Every threshold on a board is declared the board's own copy, and the standard is named only in the reveal. One board, C-1280 The Grader's Words, teaches fourteen of these words and asks you to class drawn specimens with the card hidden, so the real words are its content. The file recommends (a): ship the real names, because retrieving them is that board's point, and they are stable, secular and public.

• (a) Ship the real grade names, thresholds declared the board's copy, the standard named in the reveal (the recommendation).
• (b) Synthetic grade names on the boards, the real ones named in the reveal.

Zero either way. Under (a) C-1280 carries a fourteen-row curated bank; under (b) it carries a generated one. The mechanics are identical.

Date growers name five stages of ripening: hababouk, kimri (green), khalal (full size, yellow or red), rutab (soft and brown from the tip to about halfway) and tamar (fully brown). The horticultural literature uses the transliterated Arabic names. Two admitted boards use them. C-1277★ Pick the Rutab asks you to pick every date at rutab over several visits as the strands ripen at different speeds. C-1280 The Grader's Words includes the date words in its bank, and its trap is calling a half-brown date tamar because it is brown. The board that showed one bunch from three sides was cut at assembly. The file recommends (a): transliterate the names on the board, as the literature does. No religious or festival association is used under either option.

• (a) Transliterated Arabic stage names on the board (the recommendation).
• (b) English stage descriptions on the board ("half brown", "fully brown"), with the Arabic names in the reveal.

Zero. Only a label changes; both boards are identical under both options.

Four people's methods sit under Part Twenty's boards. Uceda and Frías (1975) wrote the olive maturity index behind C-1261 Cut to the Stone and C-1264★ The Mill Takes the Mean. Charpentier described the size–weight illusion in 1891, which is the background of C-1267 Heavy for Its Size. Miller's 1956 "magical number seven" is behind C-1274 Seven Shades Without the Chart and C-1275★. Wilbur Scoville's 1912 panel test is behind C-1266 Three of Five Can Still Tell. None of them is a puzzle actor; the question is only whether a reveal names them. The file recommends (a): name them in reveals only, as the authors of methods. Unlike IDEA-061's sample proportions, no finding would be quoted as a measurement of the player.

• (a) Named in reveals only, as the authors of methods (the recommendation).
• (b) Unnamed; the method is described without its author.

Zero. Only a reveal sentence changes; every board is identical under both options.

A 1988 European quality standard required Class I cucumbers to be "practically straight": at most 10 mm of arc for every 10 cm of length. It was revoked in 2009, and it later became a stock example in political arguments about European rules. The file proposed a board built on it (T25-20: class each drawn fruit by the height of its bend against its length) and recommended shipping it with a neutral attribution. The assembler cut it on design grounds. It was the fourth entry in the file's own cut order, and it repeats the length-ratio judgement of C-1272 Head Rice with the same grammar and a closely related failure. Choosing (a) or (b) reopens the entry for a later pass. Choosing (c) confirms the cut.

• (a) Ship it, attributed as "a 1988 European quality standard, revoked in 2009" and nothing more (the file's recommendation; reopens the cut entry).
• (b) Ship it with the curvature rule as the board's own and no standard named.
• (c) Leave it cut (the assembler's ruling).

(a) or (b): one L board at a later assembly, re-argued against C-1272; (c) nothing.

"The Rivet Without a Twin" (rivet_twin) often left half the wing empty, because both rivets of a pair must sit on the panel and an axis far off-centre leaves only a thin strip where that is possible. The generator fix (commit f413762) keeps only axes that leave at least half the panel usable, and the rivets now spread well (empty panel down from ~40% to ~29% at level 1, and from ~25% to ~15% at levels 3–5, where a middle-line guess can no longer pass). The cost is at level 1 only: the two remaining axes both sit one notch off the middle, so a blind tap on a middle line now passes level 1 about half the time instead of a quarter. Levels 2–5 are as hard as before or harder.

• (a) Keep the fix as built — level 1 is a gentler on-ramp with a well-spread panel.
• (b) Restore the far-off-centre axes at level 1 only, accepting a panel that is about half empty on those days.
• (c) Keep the fix and add one more twinless-rivet decoy at level 1 so a middle-line guess still needs the right rivet.

(a) nothing; (b) a one-line generator change plus a cache bump; (c) about an hour of generator work plus a cache bump.

Five admitted boards are built on toy-brick models: C-1290 Only What Is New (put the new bricks where a step picture drawn from another side says they go), C-1292 Build It Again, Mirrored (build the mirror of a wing), C-1293 Where It Went Wrong (find the step where the model went off the sheet), and two fusions. People recognise a brick by its proportions: a plate a third of a brick high, studs at a fixed pitch. No brand, logo, set or system is named anywhere, and every proportion is declared the board's own. The question is how close to the familiar shape the drawings may sit. The file recommends (a): proportions close to the common ones, declared the board's own, no logo on any stud and no system named.

• (a) The board's own proportions, close to the familiar ones, no logo, no system named (the recommendation).
• (b) Deliberately different proportions, so no brick system is evoked at all.

Zero either way. Only the drawn proportions change; the puzzles and their answers are identical.

Real flat-pack and brick instruction sheets draw a person doing the step, and a second person when a step needs two. The T26 boards draw each step as parts, arrows, counts and turn marks only, and "two people" becomes a count mark. This is IDEA-057 (figures without faces) applied to a whole kind of board: C-1291 Every Dowel Finds a Hole, the brick boards, and C-1180★'s raise dial. The file recommends (a): no human figure. The steps read cleanly without one, and nothing on any board needs a body.

• (a) No human figure; "two people" drawn as a count mark (the recommendation, consistent with IDEA-057 and IDEA-100).
• (b) Faceless point figures on the sheets, if IDEA-057 is answered to allow them.

Zero under (a). Under (b), a small glyph-figure asset shared with IDEA-100's dance boards.

Two admitted boards stack three rolled snowballs. In C-1287 Roll It to Size you choose how far to roll each ball so they stand in size order, the upper ones light enough to lift, and the stack reaches a height. In C-1300★ Three Balls From One Lawn a rolled ball clears the snow behind it, so the second ball must find fresh ground. A snowman usually has a face, a hat and a carrot nose. Under IDEA-057 the boards draw none of that. The file recommends (a): three balls and no face, which still reads as a snowman. Option (b) drops the word "snowman" as well.

• (a) Three balls, no face, hat, scarf or carrot, called a snowman (the recommendation).
• (b) No snowman framing: "snowballs stacked".

Zero. Only a title and a reveal line change; both boards are identical under both options.

A holzhausen ("wood house") is a round woodpile: split logs laid in a ring, bark down at the base, bark up like shingles on the roof, with air drawn up through the middle. It is being added to the shipped board Split for the Stack (split_stack) as a level-5 ring layout. The ring changes the layout, not the puzzle. Because split_stack is already built, the addition will need a cache bump whenever someone builds it. The question here is only the name. The file recommends (a): describe it as "a round woodpile (holzhausen)" in the reveal and keep the board's own words on the board.

• (a) "A round woodpile (holzhausen)" in the reveal only (the recommendation).
• (b) The German word on the board itself.

Zero. Only a label changes. The ring layout itself is a separate, recorded CACHE_VERSION event for whoever builds it.

The file proposed a board on hanging a row of pictures: pick a hook for each frame so the centres fall on one line with equal gaps, allowing for how far each frame's wire drops below its hook. It mentioned a gallery habit of hanging picture centres at 57 inches from the floor, but could not verify it. The assembler cut the board on design grounds. It was third in the file's own cut order, it is a flat sum in a territory about building in 3D, and straightening a picture already sits on the calibration boards. The file recommended (a), which would bring the board back. Choosing (a) or (b) reopens it for a later pass; choosing (c) confirms the cut.

• (a) Bring the board back with its own centre line and no convention named (the file's recommendation; reopens the cut entry).
• (b) Bring it back and name the 57-inch convention once a source is verified (also reopens it).
• (c) Leave it cut (the assembler's ruling).

(a) or (b): one small board at a later assembly, re-argued against the calibration boards; (c) nothing.

The vibe pass gave Home a one-line day-book greeting that changes by weekday and weekly arc ("Monday opens the arc…"). It is written in the app's own voice, not Sepia's, because Sepia's character rules forbid a generic opener, cap her Home lines at three days in seven, and require every line to carry a verifiable number from your log. Making her the morning greeter would bring the cuttlefish forward as the face of the ritual — warmer and sillier, closer to the "morning games" feel — at the cost of loosening those rules.

• (a) Keep as built — the app greets; Sepia speaks only when she has a number to report.
• (b) Let Sepia deliver the daily greeting (one line, every day), relaxing the three-in-seven cap for Home only; she still never praises or scolds.
• (c) Alternate — Sepia on the days she has a number, the day-book line otherwise.

(a) nothing; (b)/(c) a small change in app/character/ plus new lines and tests.

Sharing is opt-in, but the Share keys send an image card first and fall back to text. DEC-006 says sharing, "if it ever exists, is text and opt-in". The vibe pass cleaned the share line (no emoji; the route like 3A-02 ▸ 1C-07 ▸ 4B-01 instead of a category that always read "Puzzle") but left the image-first behaviour alone because changing it is a product decision.

• (a) Text only, as DEC-006 states — drop the image card.
• (b) Keep image-first with text fallback, and amend DEC-006 to allow an opt-in image.
• (c) Offer both explicitly: SHARE TEXT and SHARE CARD.

(a)/(b) small; (c) small plus one Key on the results card.

The APK is 32.5 MB against the 34 MB gate, growing ~160 KB per wave — about eight waves of room left. The audit found that release code minification (R8) and resource shrinking were never switched on; together they cut the test build from 32.54 MB to 27.34 MB (measured, with the decoder removal included). The risk is that R8 can strip something the app loads by reflection, which shows up only as a crash at runtime — and no phone is attached here. A test build with R8 on is at O:\rclone-offload\cognitive-hub-v3.4.0-SIZE-TEST-r8.apk. If it opens, you can reach Home, the Chart, a practice run and a few puzzles, and nothing crashes, R8 becomes the default. (The unused GIF/WebP decoders — ~0.8 MB — are already removed from every build; that change has no runtime risk.)

• As of v3.28.0 (2026-09-30, information only): the shipped APK is 33,599,916 bytes, 0.40 MB (400,084 bytes) under the 34 MB gate (v3.27.0 was 33,480,732 bytes, 0.52 MB under). v3.28.0 carries waves 48 and 49, which together added about 119 KB. Recent waves add about 55–130 KB each, which leaves roughly 3–7 waves before the gate binds. The R8 test APK measured 27.34 MB, about 4.4 MB smaller. Nothing has been decided; this question and IDEA-117 remain yours.
• As of v3.29.0 (2026-09-30, information only): the shipped APK is 33,663,096 bytes, 0.34 MB (336,904 bytes) under the 34 MB gate (v3.28.0 was 33,599,916 bytes, 0.40 MB under). Wave 50 added about 63 KB. At recent rates of about 55–130 KB a wave, roughly 2–5 waves remain before the gate binds, fewer if a wave brings assets. The R8 test APK measured 27.34 MB, about 4.4 MB smaller. Nothing has been decided; this question and IDEA-117 remain yours.
• As of v3.30.0 (2026-10-01, information only): the shipped APK carries polish wave P1 and wave 51 and is 33,733,348 bytes, 0.27 MB (266,652 bytes) under the 34 MB gate (v3.29.0 was 33,663,096 bytes, 0.34 MB under). The two together added about 70 KB, nearly all of it wave 51's three boards; P1 added no Views or assets. At recent rates of about 55–130 KB a puzzle wave, roughly 2–4 puzzle waves remain before the gate binds, fewer if a wave brings assets. An agent tried to build a fresh R8 test APK at this close, but the change to the build copy was not permitted, so none was built: the only R8 test APK is still cognitive-hub-v3.4.0-SIZE-TEST-r8.apk (27.34 MB). Nothing has been decided; this question and IDEA-117 remain yours.
• As of v3.31.0 (2026-10-01, information only): the shipped APK carries wave 52 and polish wave P2 and is 33,783,332 bytes, 0.22 MB (216,668 bytes) under the 34 MB gate (v3.30.0 was 33,733,348 bytes, 0.27 MB under). The two together added about 50 KB. The headroom is shrinking: at this close's rate about four puzzle waves remain, and at the 130 KB high end of recent waves about one, so roughly 1–4, fewer if a wave brings assets. A fresh R8 test build is still not permitted from the agent side, so the only R8 test APK is still cognitive-hub-v3.4.0-SIZE-TEST-r8.apk (27.34 MB). Nothing has been decided; this question and IDEA-117 remain yours.
• As of v3.32.0 (2026-10-01, information only): the gate binds soon. The shipped APK carries polish wave P3 and wave 53 and is 33,841,592 bytes, 0.16 MB (158,408 bytes) under the 34 MB gate (v3.31.0 was 33,783,332 bytes, 0.22 MB under). The two together added about 58 KB. At this close's rate about 2–3 waves remain, and at the 130 KB high end of recent waves one. After that, a wave that adds boards cannot ship under the gate until this question or IDEA-117 is answered. A fresh R8 test build is still not permitted from the agent side, so the only R8 test APK is still cognitive-hub-v3.4.0-SIZE-TEST-r8.apk (27.34 MB, about 4.4 MB smaller). Nothing has been decided; this question and IDEA-117 remain yours.
• As of v3.33.0 (2026-10-01, information only): at most ONE more puzzle wave fits. The shipped APK carries polish wave P4 and wave 54 and is 33,889,988 bytes, 110,012 bytes (about 0.11 MB) under the 34 MB gate (v3.32.0 was 33,841,592 bytes, 0.16 MB under). The two together added 48,396 bytes. That leaves room for wave 55 at most. After wave 55, the size gate blocks further puzzle waves until this question or IDEA-117 is answered. Polish waves add little and can continue. A fresh R8 test build is still not permitted from the agent side, so the only R8 test APK is still cognitive-hub-v3.4.0-SIZE-TEST-r8.apk (27.34 MB, about 4.4 MB smaller). Nothing has been decided; this question and IDEA-117 remain yours.
• As of v3.34.0 (2026-10-01, information only): new puzzle waves are now PAUSED by the gate. The shipped APK carries wave 55 and polish wave P5 and is 33,932,644 bytes, 67,356 bytes (about 0.07 MB) under the 34 MB gate (v3.33.0 was 33,889,988 bytes, 110,012 bytes under). The two together added 42,656 bytes; P5 was about 2.4 KB of it. Recent closes have added about 40–60 KB each, which leaves no safe margin, so no new puzzle wave is planned until this question or IDEA-117 is answered. A very lean wave might squeeze in, but the margin is too thin to plan on. Polish waves continue. Answering (a) here would free about 4.4 MB, after a phone test of the R8 build. A fresh R8 test build is still not permitted from the agent side, so the only R8 test APK is still cognitive-hub-v3.4.0-SIZE-TEST-r8.apk (27.34 MB). Nothing has been decided; this question and IDEA-117 remain yours.
• Wave 48 (2026-09-30, information only): the APK is now also blocked by C: space. C: fell to about 0.97–1.0 GB, below the 1.2 GB build threshold, so v3.28.0 was not built and the latest APK stays v3.27.0 (the size above). Wave 48's three types are on main and will add their bytes to the next APK once C: recovers (see DIRECTIONS, owner action). Resolved at wave 49: the build copy moved to D:\j, and v3.28.0 shipped with waves 48 and 49 (the size above).

• (a) I tried the test APK and it works — make R8 and resource shrinking the default.
• (b) It crashed (tell me where) — keep R8 off and look for keep-rules.
• (c) Skip R8; take the other levers when needed (bundle compression −3.2 MB at a small cold-start cost; see IDEA-117).

A few minutes on the phone.

• As of v3.35.0 (2026-10-01, information only): still PAUSED. The shipped APK carries polish wave P6 alone and is 33,932,428 bytes, 67,572 bytes (about 0.07 MB) under the 34 MB gate, 216 bytes smaller than v3.34.0. Polish costs nothing measurable; puzzle waves stay paused until this question or IDEA-117 is answered. The only R8 test APK is still cognitive-hub-v3.4.0-SIZE-TEST-r8.apk (27.34 MB). Nothing has been decided; this question and IDEA-117 remain yours.
• As of v3.36.0 (2026-10-01, information only): still PAUSED. The shipped APK carries polish wave P7 alone and is 33,934,640 bytes, 65,360 bytes (about 0.07 MB) under the 34 MB gate, 2,212 bytes more than v3.35.0. Polish costs a few kilobytes a pass at most; puzzle waves stay paused until this question or IDEA-117 is answered. The only R8 test APK is still cognitive-hub-v3.4.0-SIZE-TEST-r8.apk (27.34 MB). Nothing has been decided; this question and IDEA-117 remain yours.
• As of v3.37.0 (2026-10-01, information only): still PAUSED. The shipped APK carries polish wave P8 alone and is 33,935,328 bytes, 64,672 bytes (about 0.06 MB) under the 34 MB gate, 688 bytes more than v3.36.0. Polish costs a few hundred bytes to a few kilobytes a pass; puzzle waves stay paused until this question or IDEA-117 is answered. The only R8 test APK is still cognitive-hub-v3.4.0-SIZE-TEST-r8.apk (27.34 MB). Nothing has been decided; this question and IDEA-117 remain yours.
• As of v3.38.0 (2026-10-01, information only): still PAUSED. The shipped APK carries polish wave P9 alone and is 33,936,512 bytes, 63,488 bytes (about 0.06 MB) under the 34 MB gate, 1,184 bytes more than v3.37.0. Polish costs a few hundred bytes to a few kilobytes a pass; puzzle waves stay paused until this question or IDEA-117 is answered. The only R8 test APK is still cognitive-hub-v3.4.0-SIZE-TEST-r8.apk (27.34 MB). Nothing has been decided; this question and IDEA-117 remain yours.
• As of v3.39.0 (2026-10-01, information only): still PAUSED. The shipped APK carries polish wave P10 alone and is 33,937,884 bytes, 62,116 bytes (about 0.06 MB) under the 34 MB gate, 1,372 bytes more than v3.38.0. P10 completed the flat-face sweep, and the polish backlog that needs no owner decision is now essentially exhausted, so the loop is waiting on this question or IDEA-117 before puzzle waves can resume; the next intact T27 trio is "The plate" (C-1264★ · C-1318 · C-1316). The only R8 test APK is still cognitive-hub-v3.4.0-SIZE-TEST-r8.apk (27.34 MB). Nothing has been decided; this question and IDEA-117 remain yours.
• As of v4.0.0 (2026-10-04, information only): the gate is lifted. You lifted the 34 MB gate on 2026-10-03 (DEC-068); size is now reported against a soft 40 MB budget, so this question no longer blocks puzzle waves. APK v4.0.0 (Cycle 7 part 1) is 35,690,089 bytes, 1,752,205 bytes more than v3.39.0, mostly react-native-svg, and 4,309,911 bytes under the soft budget. R8 is still worth a phone test as a size lever (about −4.4 MB measured on the old test build), and the only R8 test APK is still cognitive-hub-v3.4.0-SIZE-TEST-r8.apk. Nothing about R8 has been decided; the status line below is yours to change.

Native libraries are currently stored uncompressed inside the APK (16.6 MB of it). Setting useLegacyPackaging=true compresses them: the APK file drops by ~11.3 MB, but the installed footprint grows to ~38 MB because Android unpacks them. Every earlier size ruling — including Skia's refusal at 39.85 MB and the 3D renderer table — assumed the current uncompressed layout, so flipping this silently would change what the gate means. Separately, bundle compression would save 3.2 MB and halve per-wave growth at an estimated +50–150 ms cold start (not yet measured).

• (a) The gate measures the APK file (download) — legacy packaging becomes available, and so does real 3D within the budget.
• (b) The gate measures installed size — keep libraries uncompressed; use R8 and bundle compression instead.
• (c) Raise the gate itself (e.g. 38 MB soft / 45 MB kill) and keep the current layout.

A one-line gradle change either way; the decision is what matters.

• As of v4.0.0 (2026-10-04, information only): answered in effect by DEC-068. On 2026-10-03 you wrote that 34 MB does not have to be the cap; the gate is lifted and size is reported against a soft 40 MB budget, close to option (c), with the library layout unchanged. APK v4.0.0 is 35,690,089 bytes, 4,309,911 bytes under that budget. What the soft budget measures (the APK file, as every close has reported it) is unchanged. The status line below is unchanged; closing this question is yours.

The representation register says dancers are never drawn as figures: every dance board moves tokens with a role glyph (● ○). The wave-25 builder of reel_eight (C-1240) and grand_march (C-1239) drew each dancer instead as a small faceless solid on the diorama's hall floor: a body block, a head block, and on the reel a nose block showing which way the dancer faces. No role glyphs appear on the stage; numbers and letters sit on the rail and the plans. The blocks read as 3D and make facing visible at a glance, which fits the owner's 3D-polish direction. The register rule exists so the dance boards never become a cast of people. Faceless blocks sit close to that line (compare IDEA-057, figures without faces). The recommendation is (c), a hybrid that keeps the blocks' depth and puts the register's glyph back on the stage.

• (a) Keep the faceless blocks as shipped, and amend the register to allow faceless solid dancers on diorama floors.
• (b) Go back to glyph tokens (● ○) on the stage, as the register states, and show facing another way.
• (c) Hybrid: keep the blocks, but put the role glyph on each block's top face, so the stage reads as tokens with depth (the recommendation).

(a) a register amendment only; (b) or (c) a renderer change to the two boards (and to feet_went's floor, if it should match), with no change to generators or answers.

app/ui/Key.tsx sizes the outer Pressable from style, but the base View under the face has no flex: 1. A height set on a Key therefore never reaches its face, and a face with little content shrinks to its content height. That is why web_wet's empty squares showed as hairlines until the wave-25 polish. Renderers now work around it by giving their content a minHeight of the key height minus 4 (the base's 2 px side and the face's two 1 px borders). Adding flex: 1 to the base fixes it once for everyone, but every Key on all 301 boards could change height. This is an engineering-taste call with a visible effect, so it is filed rather than made.

• Polish wave P1 (2026-10-01, information only): the survey of the Cycle 1–4 boards found the quirk on most of them: empty sudoku, magic-square and futoshiki cells, nonogram and position-recall squares, face-down memory cards and the one-stroke dots (which had no face at all) drew as hairlines or vanished, and most Check … keys shrank to a strip. P1 applied the per-board minHeight workaround to 29 older renderers, with no change to Key.tsx. The question above is unchanged and remains yours.
• Polish wave P4 (2026-10-01, information only): the P4 tier showed the same quirk on eleven more boards (head_rice, mirror_wing, box_pile, channel_shadow, touch_frame, straight_line, which_wall, nest_inside, every_wheel, even_turns, more_glass): steppers, tray keys and primaries drew as strips about 20 px tall inside 48 px hit targets. P4 applied the same per-board minHeight workaround. Key.tsx is unchanged and the question above remains yours.
• Polish wave P5 (2026-10-01, information only): one more board: two_leadsmen's CAST THE LEAD primary drew as a strip, and took the same per-board workaround. Key.tsx is unchanged and the question above remains yours.

• (a) Add flex: 1 to Key's base, then run a screenshot sweep of all 301 boards and fix any that change, and later remove the per-renderer workarounds.
• (b) Keep the per-renderer minHeight workaround, written into the builder brief, and leave Key.tsx as it is.

(a) a one-line change plus a full screenshot sweep and whatever it turns up, likely a wave of polish time; (b) nothing now, and one extra style line per board that sizes keys by height.

• Polish wave P6 (2026-10-01, information only): a 360 px re-survey of the 147 Cycle 5 wave 11–40 boards no polish wave had touched found the quirk on about a hundred of them (a DOM check flagged 105, a few of them figure keys drawn short on purpose), most often on steppers, verb keys and primaries. P6 applied the per-board workaround to eleven (weir_facing, corners_meet, same_mistake, grade_burning, bearer_bar, grand_march, chalk_seam, smaller_hole, three_frames, shear_draws, first_pour); the rest remain. Key.tsx is unchanged and the question above remains yours.
• P6 close (2026-10-01, information only): the case for an answer. About a hundred boards from waves 11–40 still draw the flat face, and polish waves mend about ten a pass. Option (a) would cover all of them in one change (one line in Key.tsx and one screenshot sweep); option (b) means roughly ten more polish passes for these boards alone. P7 continues board by board until you choose. Nothing has been decided.
• Polish wave P7 (2026-10-01, information only): thirteen more boards from P6's next tier took the per-board workaround (turn_table, thirty_two, sheet_shades, reel_eight, behind_under, wash_find, hint_no_spoil, off_spiral, every_crumb, every_dowel, grocer_pyramid, plumb_batter, spread_thin). On grocer_pyramid the flat faces had also pushed BASE 2 and FROM THE RIGHT off the centre line of their keys. About ninety boards from waves 11–40 still draw the flat face. Key.tsx is unchanged and the question above remains yours.
• P7 close (2026-10-01, information only): the case for an answer, updated. About ninety boards from waves 11–40 still draw the flat face after P6 and P7 mended twenty-four between them. At about a dozen a pass that is seven or eight more polish passes for these boards alone; option (a) would cover all of them in one change. P8 continues board by board until you choose. Nothing has been decided.
• Polish wave P8 (2026-10-01, information only): a 360 px re-check of the boards P1–P7 had not touched found the DOM check's flat face on 80 of them, six of those the deliberately short figures (skirt_line, green_monday, two_leaves, cut_stone, three_of_five, stack_screens). P8 applied the per-board workaround to the twenty where it read worst (cure_disease, only_new, question_wears, banker_stand, developed_together, every_wall, question_broken, rank_cups, air_cell, labels_force, next_spring, open_drawer, read_round, strip_answer, forty_five, tube_show, split_stack, between_numbers, concede_deny, passed_average). The check now flags 54 boards besides the six figures, a few of them figure keys drawn short on purpose. At about twenty a pass that is two or three more polish passes; option (a) would still cover them in one change. Key.tsx is unchanged and the question above remains yours.
• Polish wave P9 (2026-10-01, information only): P9 applied the per-board workaround to 22 more boards (curb_cable, count_verb, any_ten, linen_first, eight_cups, own_rule_closes, slow_way_round, lawn_stripes, even_generations, file_before, sides_middle, three_irons, web_wet, whites_first, milk_last, small_rinses, soft_water, too_close, trace_lake, turn_cups, plus P8's skips question_wears and rank_cups). About 34 flagged boards remain, four of them (drift_wall, ethylene_bowl, pick_ladder, pearl_strand) likely figures drawn short on purpose, so P10 may be the last board-by-board pass. Option (a) would still cover them, and keep new boards from needing the workaround, in one change. Key.tsx is unchanged and the question above remains yours.
• Polish wave P10 (2026-10-01, information only): of the 34 boards still flagged after P9, four are figures drawn short on purpose (drift_wall, ethylene_bowl, pick_ladder, pearl_strand). P10 applied the per-board workaround to the other thirty. The flat-face sweep of waves 11–40 is now complete; the only boards still flagged are the deliberate figures and any_ten's slot ring (IDEA-141). Option (a) would now mainly remove about 140 per-board workarounds rather than fix visible faults. Key.tsx is unchanged and the question remains yours.

Vibe pass 2 gave every stage row on the results card one line comparing today's time with that stage's own bench: "RECORD · WAS 1:35", "−0:10 VS MEDIAN", "BEST 1:20", or "HINTED · OFF THE BENCH". A comparison is a taste call: it is you against you, but it still puts a number beside the day. It was built the conservative way. It shows only when the timer is on, only against your own bench, and a hinted stage is marked off the bench. With the timer off the card is unchanged. The recommendation is (a): keep it as built, since it appears only where the player chose to be timed.

• (a) Keep as built: only with the timer on, own bench only, hinted stages marked off the bench (the recommendation).
• (b) Always show it, comparing against earlier timed bests even on untimed days.
• (c) Never show it; the card carries no per-stage comparison.

(a) nothing; (b) a small change in the card and its copy tests; (c) removing the line and its tests.

The owner asked for "if I want to improve this part, I can do more puzzles here". The results card now lists every chart area worked today under WORKED TODAY, and the single PRACTICE HERE suggestion from vibe pass 1 moved directly under those rows. The alternative is a small "More here" key on every area row, so any area worked today can start a practice run. That is more choice on the card, at +4 Views per row, and it changes the shipped behaviour the owner has already seen. The recommendation is (a): keep one suggestion, because the card stays calm and the Chart already offers practice from any area.

• (a) Keep one PRACTICE HERE suggestion under WORKED TODAY, as built (the recommendation).
• (b) Add a small "More here" key to every WORKED TODAY area row (+4 Views each).

(a) nothing; (b) a small change in <WorkedAreas> and the results card, plus tests.

Each WORKED TODAY row stacks one solid block per stage played in that area today. The stack is rebuilt every morning, so it can never grow into an attendance tower (DEC-006). The area's lasting standing is on a flat bevelled gauge beside it: the sounded share, with the charted share inside it. The alternative puts the standing in 3D instead: a solid post per area that grows with its charted share, like the Home leg posts. That reads more like progress you can see, but a 3D object that only grows sits closer to the streak line. The recommendation is (a): keep the blocks for today and the gauge for standing.

• (a) One block per stage today, plus the flat gauge for standing, as built (the recommendation).
• (b) A solid post per area, grown by its charted share, like the Home leg posts.

(a) nothing; (b) a renderer change in <WorkedAreas> and its View costing, with no change to the log.

<Solid> has its lit top and left edges back, as an opt-in litEdges prop that costs up to 3 Views per cell. They are on for the boards with room: fold_net, anamorph, pagoda_weight, blind_weighings, hidden_count, paint_ledger, whites_first, soft_water, the Solid lab and WORKED TODAY. They are off for handedness, impossibility_proof, void_fit and sightlines, which would go over the 400-View budget, and for web_wet and three_irons, which have too little headroom. So some 3D boards look finished and some do not. The recommendation is (a): keep it opt-in until the look is seen on a phone.

• (a) Keep it opt-in on the boards with room, as built (the recommendation).
• (b) Build a cheaper outline-only variant for the four boards over budget.
• (c) Raise the View ceiling for those four boards so the full strips fit.

(a) nothing; (b) a new variant in app/ui/solid/pieces.ts with its own tests and costing; (c) a budget-law change plus an on-device frame-time check.

A lit strip takes the brighter of two colours: the plain average of the faces it joins, or its face lifted by EDGE_LIFT. The plain average is invisible at 1 px, so the lift does the work. EDGE_LIFT is 0.12 on the ramp, which is subtle: the edge reads as a highlight, not an outline. It is one constant in app/ui/solid/pieces.ts, so changing it is cheap, but it changes every lit board at once. The recommendation is (a): keep 0.12 until someone has seen it on a phone, where 1 px strips may read differently.

• (a) Keep 0.12, subtle (the recommendation).
• (b) Stronger, for example 0.2, so the edges read clearly at arm's length.
• (c) Softer, for example 0.08, closer to the plain average.

One constant, plus a screenshot check of the lit boards.

Wave 26 set a voice rule: figures that are not the player (chasers, dancers, customers) take no gendered pronoun. It has been applied to the wave-26 and wave-27 boards and to five older types (keep_tree, starter_ten, own_rule_closes, same_mistake, bearer_bar, plus hint_no_spoil's feedback). Many older generator texts still use he or she for their figures: the ferryman's "HE SAYS", the ship as "she", the Rulebook listener's "she/her", the characters in the Lateral scenarios, the drum reader, the lender and others. So the roster is inconsistent: a new board says "the rival", and an older one says "he". Sweeping the rest changes hashed text (hints, rules lines, reveals), so every swept type's golden entries move and a CACHE_VERSION bump is needed. Historical people and quoted texts stay as they are under any option: Zheng He, Samuel Pepys and a named caliph keep their pronouns. The recommendation is (a), done in one batch, so the goldens move once and are declared together.

• (a) Sweep every remaining story figure in one batch: move and declare the affected goldens and bump CACHE_VERSION once (the recommendation).
• (b) Keep the rule for new boards only, and leave the older texts as they are.
• (c) Drop the rule and let figures be gendered where it reads naturally, including on new boards.

(a) a text sweep across roughly a dozen generator files, a declared golden move per affected type, one cache bump, and a check that no historical or quoted pronoun was touched; (b) nothing now, and a lasting inconsistency across the roster; (c) nothing now, and the builder-brief rule is struck.

only_she ("Only She Finds Her") is a kinship puzzle: compose a chain of kin terms that one listener, under her stated kinship system, reads to exactly the target person, while an eavesdropper using a different system is left with several people. Kin terms carry sex (a mother's brother is not a father's sister), so the sex of the people in it is the puzzle's subject matter. The pronoun rule was written for figures whose sex does not matter to the puzzle, so it does not obviously apply here: the title names the listener and the person found by sex, which is what the puzzle is about. It was left out of the wave-27 sweep and is filed rather than decided. The recommendation is (a): keep the title and the content as they are.

• (a) Keep the title and the content as they are, and record only_she as an exception to the rule, since sex is the subject (the recommendation).
• (b) Reword the title to be pronoun-free (for example "Only One Finds Them"), keeping the kinship content that needs sex.

(a) nothing, and one line in the builder brief naming the exception; (b) a tagline and name change in PUZZLE_TYPE_INFO, and a golden move if any hashed text changes.

Research was paused at C-1302 by DEC-067, so the build could turn to 3D polish and a complete vibe. Since then the waves have built from the enjoyment parts (Parts Sixteen to Twenty-One: activities people choose for their own sake, tea and coffee, repair, folk dances, grading by eye and hand, assembly). Wave 32 finished Part Twenty, and the builders' shortlists now name only five open entries there (C-1221, C-1227 and C-1235 in Part Eighteen, C-1242 and C-1243 in Part Nineteen). Update (same day, after a full triage — docs/cycle6/BACKLOG_TRIAGE.md): of the 155 codes in C-1148…C-1302, 87 are built, 1 is held (C-1172, on IDEA-078), 11 are ★ fusions waiting on an unbuilt parent, and 56 are buildable now (51 unclaimed after wave 33's five) — roughly 17 more trios, with the 11 fusions unlocking as their parents land. So the enjoyment backlog is not nearly spent: it holds about 15–20 waves. Across the whole catalogue (C-001…C-1302) roughly 940–1,000 codes are unbuilt, many of them deferred, folded or declined. Earlier parts still hold many unbuilt entries, so the loop will not run dry, but they did not come from the enjoyment intakes the owner's 2026-09-24 direction asked for. DEC-067's own trigger (the A-grade shortlist dropping below about 60) is framed for the whole catalogue, not for the enjoyment parts, so it does not settle this. The question is what the waves after wave 33 should build. With the triage, the recommendation shifts: (b) restricted to the enjoyment parts (build the 51 triaged entries) with (c) interleaved, and commission (a) a new territory only when about five trios remain. The original recommendation — (a) now — assumed a backlog of five. This is filed for the owner, not decided.

• (a) Reopen research for a new enjoyment intake (T27): one territory, researched and assembled into a new addendum part, then built.
• (b) Keep building older unbuilt catalogue entries from the earlier parts, with research still paused.
• (c) Turn the waves to polish and vibe work: the Chart, the results card and a 3D pass, with no new types for a while.

(a) one research territory and one assembler pass (roughly the effort of one intake before wave 19), and a lifting of DEC-067's pause; (b) nothing new, and the waves drift away from the enjoyment direction; (c) no new types while it runs, and more of the polish queue and the vibe-pass questions (IDEA-114, IDEA-115, IDEA-120..124) answered on screen.

• Wave 49 (2026-09-30): this becomes binding after about one more wave. After wave 49, 11 enjoyment entries are buildable plus two unlocked fusions (C-1288★, C-1244★), but only three narrow abilities remain, and Gf-RG and Gf-RQ have one buildable entry each (C-1153, C-1206). So at most one D21-clean trio is left: C-1153 (Gf-RG) + C-1206 (Gf-RQ) + a Gq entry. After that trio every remaining entry is Gq, and the waves must take (a), (b) or (c). This is still yours to decide.
• Wave 50 (2026-09-30): this is now BINDING, and new puzzle waves are paused until you choose. Wave 50 built the last D21-clean trio (C-1153 Gf-RG, C-1206 Gf-RQ, C-1244★ Gq). What is left: 9 buildable entries, all Gq (C-1200..1203, C-1220..1223, C-1264★), plus two unlocked fusions, C-1288★ (Gq) and the newly unlocked C-1212★ (Gf-RG, unlocked when C-1206 landed). Two narrow abilities cannot fill a trio of three. docs/cycle6/BACKLOG_TRIAGE.md is now current: 141 built, 11 buildable, 2 fusions waiting (C-1204★, C-1205★), 1 held (C-1172). The options above stand; agents will not pick one.
• 2026-10-01 (information only): you answered "Do everything". Recording that answer here and in DECISIONS is left to you; the status line below is unchanged. What has run under it so far: (a) T27, the dawn chorus, Part Twenty-Two (C-1303…C-1328, 26 entries on sixteen abilities, ten suggested trios); (b) wave 51, tea time, the first trio to seat an older catalogue entry (C-1212★ every_pot, C-1202 waist_lighter, and C-454 fragments from Cycle 3), 379 → 382 types; (c) polish wave P1, the Key face-height workaround on 29 older boards. P1 and wave 51 shipped together in APK v3.30.0. Wave 51 built C-1212★, so T27's trio 4 needs a new Gf-RG seat.
• Wave 52 and polish P2 (2026-10-01, information only): (a) and (b) continued: wave 52 built T27's trio T1, the dawn chorus (C-1310 most_at_once, C-1303 twice_over, C-1317 through_leaves), the first types from the reopened research, 382 → 385. (c) continued: polish wave P2 fixed which_invisibility, the results card and seven clipped or flat boards. Both shipped in APK v3.31.0. The next intact trio in T27's shortlist is trio 2, "The hide"; trio 4 is still broken. The status line below is unchanged; recording the decision is yours.
• Polish P3 and wave 53 (2026-10-01, information only): (c) continued: polish wave P3 fixed P2's next tier (two_leadsmen, elder_to_whom, morse_fence), five dead primaries board by board, the ten worst boards of a 200-board survey and the status line, and filed IDEA-138. (a) and (b) continued: wave 53 built T27's trio T2, the hide (C-1320 from_hide, C-1307 phrase_bird, C-1312 singing_silent), 385 → 388. Both shipped in APK v3.32.0. The next intact trio in T27's shortlist is trio 3, "The field guide"; trio 4 is still broken. The status line below is unchanged; recording the decision is yours.
• Polish P4 and wave 54 (2026-10-01, information only): (c) continued: polish wave P4 re-shot from_hide on its final geometry (no change needed), fixed fourteen boards of P3's next tier, and filed IDEA-139. (a) and (b) continued: wave 54 built T27's trio T3, the field guide (C-1313 usually_winter, C-1314 mark_splits, C-1311 jay_buzzard), 388 → 391, and unlocked C-1327★. Both shipped in APK v3.33.0. The next intact trio in T27's shortlist is trio 5, "The marsh"; trio 4 is still broken. The APK now has room for at most one more puzzle wave (IDEA-116). The status line below is unchanged; recording the decision is yours.
• Wave 55 and polish P5 (2026-10-01, information only): (a) and (b) continued: wave 55 built T27's trio T5, the marsh (C-1323 two_counters, C-1309 other_branch, C-1306 higher_branch), 391 → 394; C-1328★ stays locked until C-1322 is built. (c) continued: polish wave P5 fixed two_leadsmen's CAST THE LEAD face, from_hide's call rows and ten boards from waves 41–54, and filed no new IDEA. Both shipped in APK v3.34.0, which is 67,356 bytes under the 34 MB gate, so new puzzle waves are paused until IDEA-116 or IDEA-117 is answered; polish continues. When the gate lifts, the next intact trio in T27's shortlist is trio 6, "The plate"; trio 4 is still broken. The status line below is unchanged; recording the decision is yours.
• Polish P6 (2026-10-01, information only): (c) continued alone: polish wave P6 moved pivot_landing's DOOR label, unclipped weir_facing and mended eleven boards from waves 11–40, and filed no new IDEA. It shipped in APK v3.35.0, 216 bytes smaller than v3.34.0. Puzzle waves stay paused by the size gate. The status line below is unchanged; recording the decision is yours.
• Polish P7 (2026-10-01, information only): (c) continued alone: polish wave P7 spaced behind_under's overlapping rail tags, held wash_find's TELL-TALE whole and mended twelve more boards from P6's next tier, and filed IDEA-140. It shipped in APK v3.36.0, 2,212 bytes more than v3.35.0 and 65,360 bytes under the gate. Puzzle waves stay paused by the size gate. The status line below is unchanged; recording the decision is yours.
• Polish P8 (2026-10-01, information only): (c) continued alone: polish wave P8 gave twenty more flagged boards their key faces, put cure_disease's slot claims and concede_deny's three verbs in one row each, wrapped between_numbers' steppers whole and held passed_average's DOUBLE-WEIGHTED whole, and filed no new IDEA. It shipped in APK v3.37.0, 688 bytes more than v3.36.0 and 64,672 bytes under the gate. Puzzle waves stay paused by the size gate. The status line below is unchanged; recording the decision is yours.
• Polish P9 (2026-10-01, information only): (c) continued alone: polish wave P9 made curb_cable's clipped d5 slots tappable again, unclipped linen_first's d5 pieces, held three hyphenated labels whole, split four lone-key rows evenly and gave 22 boards their key faces, and filed IDEA-141. It shipped in APK v3.38.0, 1,184 bytes more than v3.37.0 and 63,488 bytes under the gate. Puzzle waves stay paused by the size gate. The status line below is unchanged; recording the decision is yours.
• Polish P10 (2026-10-01, information only): (c) continued alone: polish wave P10 gave the last thirty flagged boards their key faces, completing the flat-face sweep, kept two_letters' d5 cards inside the board, split three boards' rows 2 + 2 and kept one_hundred's THEY DIVERGE from wrapping alone, and filed no new IDEA. It shipped in APK v3.39.0, 1,372 bytes more than v3.38.0 and 62,116 bytes under the gate. Polish that needs no owner decision is now essentially exhausted; puzzle waves stay paused by the size gate. The status line below is unchanged; recording the decision is yours.
• Cycle 7 part 1 (2026-10-04, information only): after P10 you directed the next work yourself. You lifted the size gate on 2026-10-03 (DEC-068) and set the art direction on 2026-10-04 (DEC-069): puzzly gadget worlds, one per ability area. Cycle 7 built the visual stack, Sepia's vector rig, the gadget hub with six world gates, the world frame and solve wavefront, and the first two world scenes on the Chart, and shipped in APK v4.0.0. No puzzle wave ran in it; puzzle waves may resume under DEC-068. The status line below is unchanged; recording the IDEA-127 decision is yours.

Seven boards in Part Twenty-Two are read off a sonogram: a picture of sound, time across and pitch up, the way field guides print songs beside the plates. Examples are splitting a thrush's song into its repeated phrases (C-1303), sorting two birds singing one song between them (C-1304), and grouping calls by the size of their drop, whatever note they start on (C-1306). They are built silent, so they work on a muted phone on a train. Real birders also learn by ear, and the obvious delight is hearing the song. The app is sound-free by law today, and IDEA-042 (an opt-in tone set) is unanswered. This question is narrower. Could a song board play the board's own synthesised song beside its sonogram, as an optional extra that the puzzle never depends on? The recommendation is (a) now, with (b) worth revisiting if IDEA-042 opens sound at all.

• (a) Silent only: the sonogram is the whole board (the recommendation).
• (b) An opt-in PLAY key on the song boards, off by default. It plays the board's own synthesised notes, never a recording of a real bird, and the puzzle stays fully solvable without it.
• (c) Audio-first song boards, with the sonogram as the accessibility twin.

(a) nothing. (b) a small synthesiser for the board's note shapes (no recorded assets) and a settings check, riding on IDEA-042's decision. (c) the same plus a hearing-free twin for every board, and the commute promise is weakened.

• Wave 52 (2026-10-01, information only): twice_over (C-1303), the first sonogram board, shipped silent under option (a): nothing plays, and the board is read off the strip alone. Still yours to decide.
• Wave 53 (2026-10-01, information only): the hide shipped silent too. phrase_bird (C-1307) draws its level 4–5 probes as sonograms, and singing_silent (C-1312) prints its song strips; nothing plays on either. Still yours to decide.
• Wave 54 (2026-10-01, information only): the field guide shipped silent as well. jay_buzzard (C-1311) prints its heard calls as log lines, the raptor's loudest by type face alone; nothing plays on any of the three boards. Still yours to decide.
• Wave 55 (2026-10-01, information only): the marsh shipped silent too. higher_branch (C-1306), named in the pitch above, groups calls by their drop on small printed sonograms; the sonogram is the song, and nothing plays on any of the three boards. Still yours to decide.

Every bird on the T27 boards is invented: its name, plate, song, range and field marks are the board's own. So the generator knows every answer, nothing needs checking against a real field guide, and no real bird is misdescribed. Real birding culture still appears in the reveals: the garden-count rule, the winter count begun in 1900, the chiffchaff named for its song in five languages. Some players would enjoy meeting real garden birds instead. That needs a curated, hand-verified bank of real species and marks, the kind of content this codebase has got wrong before. The recommendation is (a).

• (a) Invented birds, with real practice and history in the reveals (the recommendation).
• (b) A small bank of real common birds (perhaps twenty), each mark hand-verified against a cited guide, on some levels only.
• (c) Invented birds on the boards, plus a real species named as a "looks like" note in the reveal.

(a) nothing. (b) a curated bank with sources and per-region choices (which garden?), plus ongoing verification. (c) one verified line per board.

• Wave 52 (2026-10-01, information only): the three dawn-chorus boards shipped with invented birds under option (a): RUSTCAP, DUNWING, PIPKIN, SLATEBACK, MOSSNECK and CHALKBILL. Every bird, plate and song is declared the boards' own. Still yours to decide.
• Wave 53 (2026-10-01, information only): the hide's three boards shipped with invented birds under option (a): PIPKIN, BRACKLE, SEDGET, FENNOCK, MOTTLEBY, QUILLET, WHINNOCK and TARRIBEL, with invented plates, phrases, keys and calls. Still yours to decide.
• Wave 54 (2026-10-01, information only): the field guide's three boards shipped with invented birds under option (a): MEWHAWK and CRAGMEW (raptors), MOCKLE and PIEJACK (mimics), and the hide's eight names, with invented entries, marks and logs. The real Eurasian jay and common buzzard appear only in jay_buzzard's catalogue title and its reveal, which quotes Wikipedia ("Eurasian jay"); the reveals of usually_winter and mark_splits quote Wikipedia on Roger Tory Peterson. Still yours to decide.
• Wave 55 (2026-10-01, information only): the marsh's three boards shipped with invented birds under option (a): REEDLARK, MUDWING and SEDGEPIP, with invented maps, sector roses, windows, tie rules and calls. The real robin and cuckoo appear only in the catalogue titles and in higher_branch's reveal, which quotes Wikipedia ("Common cuckoo"); the other reveals quote Wikipedia ("Mark and recapture") and Ralph, Droege & Sauer (1995, USDA Forest Service Treesearch 31755) on point counts. Still yours to decide.

People all over the world hear words in bird songs. In English the barred owl says who-cooks-for-you. In Japan the practice is called kikinashi: the sankōchō's call is heard as tsuki hi hoshi, hoi hoi hoi ("moon, sun, star", which gave the bird its name), and the lesser cuckoo's as tokkyo kyoka kyoku ("patent office"). These are the reveal lines for C-1305 Who Cooks for You and C-1307 A Phrase for Every Bird. The best-known Japanese example, the bush warbler's hokke-kyō, is the name of a Buddhist sutra, so it is left out under the register's rule on sacred content. The recommendation is (a).

• (a) Quote the secular examples in romanised Japanese with an English gloss, alongside the English ones; leave out the sutra reading (the recommendation).
• (b) English mnemonics only.
• (c) Include the bush warbler too, explained as a name.

Zero. Reveal copy only.

• Wave 53 (2026-10-01, information only): phrase_bird (C-1307) shipped on option (b) as the conservative default: its reveal quotes one English mnemonic, the barred owl's "Who cooks for you" (Wikipedia, "Barred owl"), and no Japanese example. Adding the kikinashi examples later is a reveal-copy change. C-1305 is unbuilt. Still yours to decide.

The most natural way to "sing it back" is to tap the rhythm. Today C-1303 and C-1308 are rebuilt from chips: tap-only, untimed and commute-safe. A tap-along form would be timing, which IDEA-050 holds (precision timing as its own family). The proposal Sing Along was folded into C-1308 as a rider held on IDEA-050 instead of becoming its own entry. This item asks a different question. If IDEA-050 opens timing, should the tap-along become the primary form of these boards, with chips as the fallback, or stay an optional extra? The recommendation is (a).

• (a) Chips stay primary; tap-along is an optional rider once IDEA-050 allows it (the recommendation).
• (b) Tap-along becomes primary on the two boards once IDEA-050 allows it; chips stay as the accessibility and reduce-motion twin.
• (c) No tap-along at all.

(a) or (b): a timing rider per board after IDEA-050. (b) also changes how the boards score. (c) nothing.

• Wave 52 (2026-10-01, information only): twice_over (C-1303) shipped tap-only, untimed: the strip is tapped note by note, and a tap starts a phrase there. Untimed taps are the primary form; no tap-along was built, since IDEA-050 holds timing. Still yours to decide.
• Wave 53 (2026-10-01, information only): the hide's three boards are tap-only and untimed as well; no tap-along was built. Still yours to decide.
• Wave 54 (2026-10-01, information only): the field guide's three boards are tap-only and untimed too; no tap-along was built. Still yours to decide.
• Wave 55 (2026-10-01, information only): the marsh's three boards are tap-only and untimed as well (other_branch advances one tick a tap); no tap-along was built. Still yours to decide.

Two real counting schemes supply the rules behind two boards. The garden count asks for the highest number of each species seen at any one time, never the total, and only birds that land (C-1310 The Most at Once). The winter count began in December 1900 as counting birds instead of shooting them, in a 15-mile circle (the reveal for C-1324 The Big Day). Naming them credits the practice exactly. One carries a religious holiday's name, and both are run by organisations. The recommendation is (a).

• (a) Name both in reveals, with their dates, as the practices the rules come from (the recommendation).
• (b) Describe them generically ("a garden count", "a winter count begun in 1900").

Zero. Reveal copy only.

• Wave 52 (2026-10-01, information only): most_at_once (C-1310) shipped with the garden count described but not named, option (b) in the build sheet's own wording. The rules come from the RSPB's garden-count instructions (discoverwildlife.com and rspb.org.uk, fetched 2026-10-01). Naming it later is a reveal-copy change. Still yours to decide.

On a 412 px phone the stage header (COGNITIVE:HUB │ STAGE 2/4 │ STN 4E-02 │ 0:04) is about one clock too wide, so the clock wraps to a second line on every stage. Polish wave P1 stopped the line ending on a dangling │: the separator and the clock now wrap together, so line two reads │ 0:04. The line is still two lines. Making it one is a choice about the brand and the instrument's voice, so it is filed rather than made. The recommendation is (b): the brand already heads Home, and the stage screen is the instrument in use.

• (a) Keep it as it is now: two lines on narrow phones, with the clock and its separator kept together.
• (b) Drop COGNITIVE:HUB from the stage header only: STAGE 2/4 │ STN 4E-02 │ 0:04 fits one line (the recommendation).
• (c) Keep the brand and shorten the rest, for example 2/4 │ 4E-02 │ 0:04.

(a) nothing; (b) or (c) a one-line change in PuzzleScreen.tsx's header and a screenshot check.

Two of the oldest boards draw their pieces from generator data that breaks the glyph law. sequence_rebuild shows emoji fruit (🍓 🍑 🥝 🫐 🍊 🍎), and one of them (🫐) is missing from the phone font, so it draws as an empty box. card_memory's cards use ★ ♥ ♣ ♠ ◐ ◑ ◈ ▣, outside the sanctioned set. The symbols live in the generators (MemoryGenerators.ts), so replacing them moves those types' golden entries and needs a CACHE_VERSION bump. P1 could not do that, because its rule was that generator output must not change. What should replace them is a visual-language choice. The recommendation is (b) for both boards, done in one batch with a declared golden move.

• (a) Leave both as they are.
• (b) Replace the items with drawn tokens from the sanctioned glyphs and the category palette (shape × colour), keeping the item counts. The golden moves once, declared, with one cache bump (the recommendation).
• (c) Replace them with short words (PLUM, FIG, PEAR) in mono, so the boards read as text.

(a) nothing; (b) or (c) a generator change for two types, a declared golden move, a CACHE_VERSION bump and a renderer tweak.

Futoshiki's vertical inequalities are drawn with the logic glyphs ∧ and ∨, which are outside the sanctioned set. The legend says "arrows point to the larger number", but the marks follow < and >, whose point faces the smaller number. The board is correct; the legend misdescribes it. P1 left both alone, because changing either changes how the player reads the board. The renderer holds both, so no golden moves. The recommendation is (b).

• (a) Leave the glyphs and the legend.
• (b) Keep the marks and fix the legend: "the open side faces the larger number" (the recommendation).
• (c) Redraw the marks as the sanctioned ▴ ▾ ◂ ▸ pointing at the larger number, and keep the legend's wording.

(a) nothing; (b) one string; (c) a renderer change plus a screenshot check. No change to generators in any option.

On several boards a disabled primary key keeps its dark onAccent label on the dark dead face, so it reads as a blank key until the board arms it. P2 fixed band_number (BUILD THE STAIR) per board. The P2 survey at 360 px also showed blank-looking dead primaries on pool_rule, which_way_up, certain_likely, ghost_box and answer_already, which were left alone. Most older boards already switch the label to textFaint when dead, so the fault is a missing per-board style, not a global rule. It is the same family as IDEA-119 (Key's face height).

• (a) Keep fixing board by board, as each board is polished.
• (b) Sweep the remaining boards in one polish wave (a style per board, no Key change).
• (c) Let Key dim its children's text when disabled (a global change; every dead key is restyled at once and every board needs a screenshot check).
• Polish wave P3 (2026-10-01, information only): the five boards named above (pool_rule, which_way_up, certain_likely, ghost_box, answer_already) were fixed board by board, with the same per-board style P2 gave band_number: the dead primary's label is textFaint, and its face takes the key height minus 4. That is the existing per-board practice, not a choice among (a), (b) and (c). Key.tsx is unchanged. The question above is unchanged and remains yours.
• Wave 53 (2026-10-01, information only): the three new boards (from_hide, phrase_bird, singing_silent) were built with every dead primary keeping a faint visible label, per board, as the existing practice. Key.tsx is unchanged, and the question above remains yours.
• Wave 54 (2026-10-01, information only): the three new boards (usually_winter, mark_splits, jay_buzzard) follow the same per-board practice: mark_splits' empty new-bird keys keep a faint "1 ○" label. Key.tsx is unchanged, and the question above remains yours.

(a) nothing now; (b) about a dozen renderers and their screenshots; (c) one Key.tsx change and a full screenshot sweep. No generator changes in any option.

At level 3 and above the carcase is up to five panel columns wide (about 390 px), wider than a 360 px phone's board, and both edges were cut off. P2 scales the whole drawing down to the measured width, never up: at 360 px the joint keys are about 37 px square instead of 50, below the 44 px tap target the ring lattices keep. At 412 px they are about 44 px. The alternative keeps 50 px keys and lets the board scroll sideways, which hides part of the carcase. The recommendation is (a).

• (a) Keep the fit-to-width scale (shipped in P2).
• (b) Keep 50 px keys and scroll the carcase horizontally.
• (c) Re-lay the carcase in fewer columns (moves the drawn layout; the generator's panel coordinates stay, so no golden move).

(a) nothing; (b) one ScrollView and a screenshot check; (c) a layout rewrite in the renderer.

On a railed board, Board.tsx centres the left rail's contents in the full height of the play surface (rail has justifyContent: 'center'). When the surface holds more than the grid the rail labels (a ledger, keys, a law line above the rows), the labels slide away from their rows: on crystal_common ALDER sat two rows below its row, sock_ran's tints sat beside the wrong loads, and tape_empty's TAPE A sat beside the third drive. Polish wave P3 fixed those three board by board (the rail column fills the rail and starts at the top, or below a measured law line). A change in Board.tsx would align every railed board at once, but it moves the rail on boards whose rails are meant to centre (a short tag rail against a tall figure). This is an engineering-taste call with a visible effect on many boards, the same family as IDEA-119 and IDEA-136, so it is filed rather than made.

• (a) Keep aligning board by board as each board is polished (P3's practice).
• (b) Sweep the railed boards in one polish wave (a style per board, no Board.tsx change).
• (c) Start the left rail at the top in Board.tsx and let boards that want a centred rail ask for it (a global change; every railed board needs a screenshot check).

(a) nothing now; (b) a survey of the railed boards and their screenshots; (c) one Board.tsx change plus a full screenshot sweep. No generator changes in any option.

Key gives every key a 48 × 48 minimum, and a board can override it through style. Polish waves have done so where 48 px keys cannot fit a 360 px phone: P3 drew morse_fence's posts 22 × 40, and P4 drew touch_frame's film chips one lane high (40 px, and as narrow as one 24 px film column, the width the object occupies), and the steppers of every_wheel and straight_line at 44 × 44. The alternative was Key's minimum, which made touch_frame's chips spill over the next lane's label and pushed every_wheel's last ▸ past the board's edge. Narrow keys are harder to hit; the 44 px steppers meet the common 44 px guideline, but a 24 × 40 chip does not. Whether a floor applies to these overrides (and what it is) is an accessibility and taste call, so it is filed rather than made.

• (a) Keep the per-board overrides as shipped (fit first; boards decide).
• (b) Set a floor of 44 × 44 for overrides: narrow chips such as touch_frame's grow to 44 and the lane layout is re-spaced (a wider plan, or chips that overhang their film columns).
• (c) Keep the drawn size but widen the hit area with hitSlop where neighbours allow (no visual change; overlapping slops resolve to the later key).

(a) nothing; (b) a re-layout of touch_frame's plan and a check of morse_fence; (c) one prop per board and a click check. No generator changes in any option.

• Polish wave P5 (2026-10-01, information only): sheet_drags declared 32 px detent tags, but Key's minimum drew them 48 wide, so they overlapped and the row ran past the board. P5 did not override the minimum: the rail now lays the tags out at 48 px, so they stagger onto two rows. No key in P5 is under 44 px. The question above remains yours.
• Polish wave P6 (2026-10-01, information only): grade_burning's six grade keys broke 5 + 1 at 360 px; they are now 44 × 48 so the row fits, which meets the 44 px floor named in (b). No key in P6 is under 44 px. The question above remains yours.
• Polish wave P7 (2026-10-01, information only): behind_under declared 30 px rail tags, but Key's minimum drew them 48 wide, so neighbouring tags overlapped (F, A, C and E at d3). As in P5's sheet_drags, P7 did not override the minimum: the rail now spaces the tags at 48 px, and they stagger onto a second row when five do not fit. Rows that wrapped one key alone (sheet_shades, off_spiral, smaller_hole, thirty_two, spread_thin) were split into even rows rather than narrowed; reel_eight's six dancers fit one row at 48 px. No key in P7 is under 44 px. The question above remains yours.
• Polish wave P8 (2026-10-01, information only): on the twenty P8 boards, keys declared 36–44 px tall that Key's minimum already drew at 48 are now declared 48, so the drawn size is unchanged; the keys that override the minimum to 44 (rank_cups' side keys, air_cell's grade chips) stay 44 with a 40 px face. between_numbers' scrub steppers wrap as pairs instead of leaving a lone ▸, and concede_deny's three verbs fit one row at an 88 px minimum width (their height stays 48). No key in P8 is under 44 px. The question above remains yours.
• Polish wave P9 (2026-10-01, information only): no key in P9 is under 44 px. Two cases were left for this question rather than overridden: lawn_stripes' lane keys, drawn 38 × 26, and any_ten's d4–d5 slot ring (filed separately as IDEA-141). Rows that wrapped one key alone were split into even rows instead of narrowed: any_ten's chip bank, curb_cable's d5 slots (4 + 3 / 4 + 4), question_wears' bands (3 + 2 / 3 + 3) and rank_cups' cups (3 + 3). Under-44 keys on pick_tray, key_couple, long_pole and pearl_strand were not touched. The question above remains yours.
• Polish wave P10 (2026-10-01, information only): keys declared 38–44 px that Key already drew at 48 are now declared 48, so their size is unchanged. pick_tray's 44 px side keys keep a 40 px face. one_hundred's verdicts narrowed from 104 to 84 px minimum width; their height stays 48. No key in P10 is under 44 px, and the under-44 keys of pick_tray, key_couple, long_pole and pearl_strand were left as they are. The question remains yours.

Board lines join their facts with " · " and hold each dot to the fact before it, so no line opens on a dot (nobreak.ts's line). On a narrow card that cannot hold both facts the break therefore falls after the dot, and the first line ends on it: corners_meet's ball card reads BALL ⌀15 · / PATH 24, and its cloth cards read SIDE 32 · / REACH 22. P6 noted the ball card for P7. A clean change exists (a bare line break between the two facts on these cards, no dot), but it changes how every two-fact card on the board reads, and the house rule is applied the same way on many other boards, so it is a taste call and is filed rather than made.

• (a) Keep the house rule: the dot stays with the fact before it, even at a line's end.
• (b) On cards too narrow for both facts, set the two facts on two lines with no dot (corners_meet's parcel and cloth cards first, then others as polish finds them).
• (c) Widen the cards (two to a row instead of three) so the facts share a line, at the cost of a longer board.

(a) nothing; (b) a line per card renderer and a screenshot check; (c) a re-layout of corners_meet's card rows and a click check. No generator changes in any option.

At d4–d5 any_ten sets 18–22 slots on a ring 288 px across. The slots are declared 30 px, but Key's 48 px minimum draws them at 48, so neighbouring hit areas overlap and the faces stretch into pills. P9 split the chip bank evenly and centred the ring for 48 px keys, which mends d1–d3; at d4–d5 there is simply not enough circumference for that many 48 px keys. Any fix changes how the board looks and how easy each slot is to hit, a layout and accessibility taste call, so it is filed rather than made.

• (a) An elliptical or taller ring, so the larger counts get more circumference (a longer board).
• (b) Two concentric rings, inner and outer, splitting the slots between them.
• (c) Keep the ring and accept the overlap, with a hitSlop or draw-order rule that settles which slot an overlapping tap lands on.
• (d) Keep it as it is.

(a) or (b) a re-layout of the ring in the renderer and a click check at d4–d5; (c) a prop or ordering rule and a click check; (d) nothing. No generator changes in any option.

The six gates on Home's causeway carry world names (the Gearworks, the Tally Works, the Lattice, the Archive, the Signal Press, the Switchyard), and so does the stage line on the solve screen: "STAGE 2 · THE LATTICE". The names make the worlds places to travel to, which is DEC-069's direction. They also add a second vocabulary beside the Chart's areas (LOGIC, NUMBER, SPATIAL, MEMORY, LANGUAGE, JUDGEMENT), so a player has to learn that the Lattice is Spatial. The accessibility labels already say both ("The Archive. 26 of 37 memory stations sighted."). Built today: the world names (WORLDS O2).

• (a) Keep the world names on gates and the stage line (built).
• (b) The plain area names everywhere: LOGIC, MEMORY and so on.
• (c) Both: the world name with the area under it on the gate nameplate, and "STAGE 2 · THE LATTICE · SPATIAL" on the stage line, at the cost of a longer line.
• (d) Names you pick.

(a) nothing; (b) or (d) a table change in app/world/worlds.ts and its copy tests; (c) a second label row on each nameplate (the plates already alternate rows) and a wider stage line, with a 320 px check.

When a board is solved, the wavefront lights every key outward from the tapped one. The keys light amber, the game's own colour, because the lit keys are the player's ("yours", WORLDS §0), and the world law keeps amber out of every world's palette, so the light reads against any world. The alternative is to light them in the world's signal hue, so a solve looks like the world answering. That ties the solve to the place, but it spends the one colour that today always means "you did this", and it can lower contrast on the darker world hues. Built today: amber.

• (a) Amber in every world (built).
• (b) The world's signal hue.
• (c) The world's hue for the wave's ring, amber for the keys.

(a) nothing; (b) or (c) one ink lookup in WorldFrame and a contrast check on all six world hues at 360 px.

On the Chart, a world band shows Sepia in her glass tank. To keep the Chart inside its frame budget her rig steps at 7.5 Hz in the band, while the world's own mechanism steps at 15 Hz. Measured at 4× CPU on the web: with her at 15 Hz the band cost 4.7–5.3 ms a frame, with 5 long frames in 16 windows, all garbage collections; at 7.5 Hz it cost 3.7–4.6 ms with 1, and 0 in the final runs. She is about 60 px wide there, and her breath, fins and cloud read the same, but a blink can fall between two steps and not show. Built today: 7.5 Hz.

• (a) Keep 7.5 Hz and accept the occasional missed blink (built).
• (b) Show her still in the band, at rest, and keep only the world's ambient moving.
• (c) Hold each blink for at least one step at 7.5 Hz, so none is missed.
• (d) 15 Hz, at the measured cost above.

(a) nothing; (b) a prop and a test; (c) a small change to the rig's blink timing and its tests; (d) a constant and a re-measure. The device run, still owed, may move these numbers.

A world band appears on the Chart only when the player arrives through a gate on Home. Tapping a region inside the Chart opens its card as before, with no band, so the plain Chart keeps its layout and nothing moves under the player's finger. The alternative shows the band on any region tap: the worlds become visible to a player who never uses the gates, at the cost of the chart opening a gap and shifting regions on an ordinary tap. Built today: through a gate only.

• (a) Through a gate only (built).
• (b) On any region tap inside the Chart as well.
• (c) A small "▸ ENTER" key on the region card that opens the band.

(a) nothing; (b) a condition in ChartScreen.tsx and a check that the shift reads as intended; (c) one key on the card and its copy test.

The band sits inside the Chart's scroll, above the selected region, so it scrolls away as the player moves down the chart. A first build held it fixed above the scroll view, always in view, but at 360 × 740 that left about 30 px of chart above the region card. Built today: it scrolls.

• (a) It scrolls with the chart (built).
• (b) Fixed above the chart, always in view, with less chart visible.
• (c) Fixed but shorter: a strip of the world rather than the whole band.

(a) nothing; (b) the first build's placement back, with a layout test; (c) a second, cropped band layout and a screenshot check at 360 and 412 px.

Your 2026-10-04 direction (DEC-069) named spatial, kinetic and memory as fields to adventure across. Spatial and memory are Chart areas. Kinetic is not: no ability in the Chart is motor or kinetic. So kinetic is built as the look of every world (gears, pendulums, marble runs, escapements), most of all in Logic's Gearworks. A real kinetic world would need new abilities and new puzzle types first, and new stations appended to the Chart, never renamed (DEC-066). Built today: kinetic as the shared look (WORLDS O1).

• (a) Kinetic is the look of all six worlds (built).
• (b) A seventh world later, with new abilities and types researched for it first.
• (c) One of the six worlds carries the word in its name (the Gearworks as "the Kinetic Works", say), with its area unchanged.

(a) nothing; (b) a research territory, new types, appended stations and a seventh gate on a causeway that is already tight below 348 px (IDEA-148); (c) a table change, as in IDEA-142.

The six gates share Home's causeway. Below about 348 px wide each gate's tap box falls under 44 px, the floor this project holds keys to (see IDEA-139). Phones that narrow are rare, but 320 px layouts exist on some small Android devices and under large display-size settings. Built today: the single row, unchanged at every width.

• (a) Keep the single row and accept under-44 targets below about 348 px (built).
• (b) Keep the row and give each gate a hitSlop, so the tap area clears 44 px while the drawing stays the same.
• (c) Two causeway rows of three gates below 348 px.

(a) nothing; (b) a prop and a 320 px tap check; (c) a second hub layout, its scene art at that width and screenshot checks.

The Switchyard is the Judgement world. Its signal mast carries four lamp housings: one dark, and one decoy, Sepia's, lit in the muted signalDim as a lure rather than an alarm; crimson stays on small marks and nothing reads as an error. A dim lamp still draws the eye a little, and a player may read it as a station or a hint. An unlit decoy keeps the joke for those who look closely and asks nothing of anyone else. Built today: dim.

• (a) Dim, in signalDim (built).
• (b) Unlit: the same housing with a dark lens, like the mast's dark lamp.
• (c) No decoy; the mast keeps three housings.

(a) nothing; (b) one fill change in app/world/scenes/switchyard.ts and its WorldScenes test; (c) the housing removed, the mast redrawn and the test updated.

Every world draws one fixture per station (a weight, a crystal, a flap, a lamp) in the same field along the bottom of the band, laid out the same way whatever the station count. A world with few stations (the Signal Press has 23) then shows its fixtures small in a field sized for many, and the art gets less of the band than it could. The alternative scales the fixtures up and the field down for those worlds, so each reads as full; it costs a second layout rule and makes the six bands differ in size. Built today: the same layout for all.

• (a) The same layout for every world (built).
• (b) Bigger fixtures in a smaller field below a station-count threshold.
• (c) The same field, with fixtures spaced wider rather than larger.

(a) nothing; (b) a scale rule in app/world/scenes/fixtures.ts, a band-height check on the Chart and screenshots at 360 and 412 px; (c) a spacing rule in the same file and a screenshot check.

A tap that goes into a world (Today's key, a gate, Practice here, a History row) opens the screen under a veil in that world's field colour with an edge rim. A plain screen (History, Journey, Settings from Home) has no world to borrow, so it arrives under a carbon veil over 180 ms. Carbon keeps the world colours meaning "you are going somewhere"; a world colour on plain screens (the day's first area, say, or the last world visited) would make every move feel like travel, at the cost of that meaning. Built today: carbon.

• (a) Carbon for plain screens (built).
• (b) The colour of the day's first world.
• (c) The colour of the last world visited, carbon on a fresh launch.

(a) nothing; (b) or (c) a colour rule in app/ui/arrival.ts, a case in tests/world/Phase5.test.ts and a timing check; (c) also keeps the last world in memory.

A selected non-primary key now latches 1 px down and rises when deselected, so a pick reads as held. Primary keys (SUBMIT and its kin) do not latch: they act rather than hold, and a latched primary could read as already pressed. Latching them too would make every key behave alike. Built today: no.

• (a) Non-primary keys only (built).
• (b) Primary keys latch too while their action is pending.

(a) nothing; (b) one condition in app/ui/Key.tsx, its tests and a click-solve pass at 360 and 412 px.

History, Journey and Settings now share one header with a room plate, so each screen reads as a room at Outstation IX: History is "The ledger", Journey "The logbook", Settings "The workbench". The names are in the house voice and say what each room keeps, but they replace the plain screen names a player may look for, and "ledger" and "logbook" are close. Built today: "The ledger" / "The logbook" / "The workbench".

• (a) "The ledger" / "The logbook" / "The workbench" (built).
• (b) Plain names: "History" / "Journey" / "Settings".
• (c) Other room names of the owner's choosing.

three strings in app/screens/PuzzleHistoryScreen.tsx, ReflectionScreen.tsx and SettingsScreen.tsx, and their tests.

Home idles at about 6 ms per frame at 4× on the web, right at the gate. A trace puts most of it on Sepia: each tick morphs her paths and repaints her surface, and the stepped ambient re-layers with it. Ticking her more slowly, or in coarser steps, would bring idle under 6 ms, at the cost of a stiffer Sepia on the screen a player sees first. Built today: the current tick.

• (a) The current tick (built).
• (b) A slower tick on Home only (for example half the rate), the band and the board unchanged.
• (c) Coarser steps: fewer pose frames per breath and blink, the same rate.

(a) nothing; (b) or (c) a cadence constant or a pose table in app/world/, a timing run at 4× and a look on a phone.

Batch 9, "number, letter and lamp grids (1 of 2)", is the only batch not on main: 19 boards (operator_fill, every_tribe, closed_both, stamp_out, tear_dots, fragments, every_crumb, tub_mangle, ladder_scan, two_letters, word_ladder, rhyme_wheel, rhyme_sort, morse_fence, carry_final, voiced_once, transdeletion, letter_bank, alternade_split). It has been built in an unmerged agent worktree, but it stops at a cleanup step that needs your approval before it runs, and it has never been independently graded. Until it lands, these 19 boards are on their pre-upgrade drawing (audit: 8 F, 8 C, 3 B) and are the only boards in the game below the DEC-072 standard. Built today: nothing merged.

• (a) Approve the cleanup; the batch then merges, goes to the independent grader and to second passes like every other batch, and ships as the next APK.
• (b) Discard the worktree and rebuild batch 9 fresh from main, to the brief as it stands now (BATCH_BRIEF S1–S4 and the grade rulings since).
• (c) Hold batch 9 for now.

(a) the cleanup, a merge, one grade and probably one or two second passes; (b) a full batch build from scratch, plus the same grading; (c) nothing, but 19 boards stay below par. Either (a) or (b) bears on the 40 MB soft budget (IDEA-158).

batch 12 redrew breakers_pink's tomato with its turned facets red and its untouched facets green, following DEC-072's "real icons". The old board, and IDEA-053 and IDEA-097, held colour to a lightness-only difference: no hue carries information. Red against green reads as a ripening tomato at a glance, but it is the one pair that colour-blind players (about 1 in 12 men) mix up, and it carries the information the puzzle asks for. Second pass 8 left the hues unchanged on instruction. blend_grade's honey (batch 21) is a related case: an amber ramp in place of grey, keeping its lightness order. Built today: red and green.

• (a) Keep red and green as built.
• (b) Back to lightness only: a pale and a deep blush of one hue, as IDEA-053 and IDEA-097 hold.
• (c) Red and green plus a second cue that does not need hue (a lightness step, or a fine hatch on turned facets), so the facet reads either way.

(a) nothing; (b) or (c) a palette change in gm2dParts.tsx' TomatoView, a reshoot and a grade. Ruling IDEA-097 would settle this class for every board.

batch 12 lays pick_tray's sample of beans out to the measured screen width rather than in the generator's 12 columns, so that every bean is drawn at least 33 px across at 360 px. A bean is still named by its index, so nothing about the answer changed; what changed is that the tray no longer has the generator's fixed grid, and a bean's row and column differ between phones. Built today: width-flowed.

• (a) Keep the width-flowed tray as built.
• (b) Restore 12 columns with smaller beans; targets fall under 32 px at 360 px (IDEA-139).
• (c) 12 columns in two half-trays of 6, stacked, at the cost of more height (d5 is about 1,135 px already).

(a) nothing; (b) or (c) a layout change in gm2dParts.tsx' SortTray, a reshoot and a grade.

DEC-068 replaced the 34 MB gate with a soft 40 MB budget. The board upgrade took the APK from 35,481,581 bytes (v5.0.1) to 39,565,869 bytes (v5.11.0), 4,084,288 bytes in ten releases, mostly new icon families and drawn board parts. v5.11.0 is 434,131 bytes under the budget. A batch of about 19 boards has added between about 270 KB (batch 12) and 310 KB (batch 21) on its own, and a release carrying three batches up to about 620 KB; batch 9 may cross the line, and any puzzle wave after it will. The R8 test build (IDEA-116) measured about 4.4 MB smaller in September but has never been tried on a phone. Built today: the 40 MB soft budget.

• (a) Raise the soft budget, to 45 MB or 50 MB.
• (b) Keep 40 MB and turn on R8 and resource shrinking first (IDEA-116), after a phone test.
• (c) Keep 40 MB and spend a size pass on the icon families (shared parts, fewer paths).
• (d) Drop the budget and keep only the size report at every release.

(a) or (d) a DEC entry; (b) a test APK, a run on your phone and a build-config change; (c) about a wave of work, with a risk to the drawings the grader passed.

"Yes you can raise the 40 MB budget." The same message asked for more polished, more fun visuals and a costing of generated art (Higgsfield, Nano Banana), which a raster-art track would spend the room on.

DEC-072 made every board's tiles, keys and cards hover. Reduce-motion stills them, but it also stills the solve power-up, the fly-through and every other motion. A player who likes the motion but finds a hovering board busy while thinking has no middle setting. A "still boards" row in Settings would stop the board hover only, and keep the screens' motion. Built today: nothing; the hover follows reduce-motion only.

• (a) Leave it: the hover follows reduce-motion only.
• (b) A Settings row, off by default, that stills the board hover only.
• (c) Still the hover automatically after the first tap on a board, so it greets the player and then keeps still.

(a) nothing; (b) one setting in SettingsStore, one check in useHover, a copy line and a test; (c) the same check driven by board state, plus a look on a phone. Either is small, but it adds one more thing to explain.

the owner asked for a game that feels 3D, "like a world", an "interdimensional puzzle land", and wondered whether voxel is the way. Round 2 (Nano Banana, tmp-art/nano/round2_contact.jpg) drew the same world map three ways: the flat-with-depth illustration of WORLDS §5 (crisp lines, cel shading), a voxel diorama (chunky cubes, tilt-shift), and a soft 3D toy render (rounded vinyl and clay). A mock-up of a voxel spatial puzzle screen is in the same sheet. Voxel has one practical edge: the rotatable puzzle pieces must be drawn by code (the turntable test showed the image model cannot hold a shape across angles), and code draws cubes natively, so a voxel world and voxel pieces would match. Built today: the illustrated direction (WORLDS §5).

• (a) Voxel everywhere: worlds, characters and pieces.
• (b) Soft 3D toy for worlds and characters, code-drawn voxel pieces with the same materials.
• (c) Illustrated worlds and characters (today's direction), code-drawn voxel pieces.
• (d) Mixed by world: each world in its own style, as separate dimensions.

the art is cheap in every case (inside the Google AI Pro plan, about a day of generation per round). The choice sets the post-process, the prompts for every later asset and how pieces are shaded.

the owner wants "a wide range of different characters". Round 2 drew a cast of eight, in two styles: Sepia in her cart, a clockwork owl (Logic), a pangolin shopkeeper with an abacus (Number), a paper-fold fox (Spatial), a jellyfish librarian (Memory), a hermit crab with a gramophone shell (Language), a heron signalwoman (Judgement), and a small traveller robot. Each would host its world, greet the player at its gate, and react to solves there; Sepia stays the guide who travels between them. Built today: Sepia only, as a code-drawn rig.

• (a) This cast, as drawn.
• (b) This cast with changes the owner names.
• (c) More than one character per world (residents and visitors), for variety day to day.
• (d) Sepia only.

art for each character (an expression sheet, about 6–8 images) and the hooks to show them; the APK grows by roughly 0.3–0.6 MB per character as WebP.

two standing rules stand in the way of the owner's direction. SPIKE_VISUAL L1 and WORLDS §5.2 allow generated raster only for skies and back walls, at most 150 KB a plate and 2 MB in all. WORLDS §5.2 also keeps Sepia and every creature as code-drawn vector rigs, because FLUX could not draw a cuttlefish; round 1 showed Nano Banana can (tmp-art/nano/r1/05_sepia_sheet.jpeg). The owner said "go ahead" on round 1, which agents read as intent, not as these law changes. Built today: both laws as written.

• (a) Lift both: raster for worlds, characters and textures, with a new total budget (for example 15 MB of art), and Sepia drawn from the generated sheet.
• (b) Lift the raster limit only; Sepia stays a code rig, restyled to match the new sheet.
• (c) Keep both laws.

(a) or (b) a DEC entry and a new size figure for the APK budget (IDEA-158 raised it with no figure); (c) nothing, but the worlds stay mostly vector.

the voxel spike (docs/cycle7/SPIKE_VOXEL.md, app/ui/voxel/) draws a piece the player can drag to turn, at 60 fps on the web. On boards where the skill is turning the shape in your head (void_fit, handedness and similar), free turning would let the player look instead of reason. Boards where a fixed view is the puzzle (hidden_count, sightlines) keep the voxel look but never turn. Built today: no board turns; the component is dev-only.

• (a) Free turning only on boards where seeing all sides is not the skill (the pilot: paint_ledger or colour_count).
• (b) A turn budget on mental-rotation boards (for example two quarter turns).
• (c) Turning unlocks after the answer is committed, as a reveal.
• (d) Free turning everywhere.

(a) the pilot board; (b) or (c) a small rule in the component and a line of copy; (d) nothing, at the cost of the puzzles' difficulty.

the owner asked for music, music puzzles, tapping on the beat and playing instruments together. Until now the product was sound-free by house rule (docs/cycle3/PUZZLE_CATALOG.md §1092: "no headphones, no quiet carriage"), and many players will play muted on a train. Built today: no sound at all; no audio library is installed.

• (a) Music and sound effects only; every puzzle still plays in silence.
• (b) As (a), plus sound puzzles that also work silently: the beat shown as a pulse and felt as haptics, the notes shown as notation.
• (c) As (b), plus some puzzles that need sound, flagged so a muted player is never served one on a muted day.

(a) one audio library, a Settings row, about 3–6 MB of music; (b) a rhythm engine with a latency calibration step; (c) as (b), plus a way for the session to know the player is muted.

Suno's free plan gives about 10 songs a day on the older v4.5 model, MP3 only, and its songs may not be used commercially, which matters if the game ever goes on the Play Store. Pro ($10 a month, about 2,500 credits) and Premier ($30) give commercial rights for songs made while subscribed, plus v5, WAV and stems; upgrading later does not cover songs made on the free plan. One source reports a September 2026 change capping commercial downloads at 20 songs a month on Pro; unconfirmed. Built today: nothing; the owner is not yet signed in to Suno in the agents' browser.

• (a) Free plan: prototypes only, regenerated on a paid plan before any release.
• (b) Pro for a month or two while the soundtrack is made.
• (c) Premier, for more songs and Suno Studio.

(a) nothing now, rework later; (b) $10 a month; (c) $30 a month. - Found 2026-10-20: the owner's Suno account is already on Pro (1.6k credits). Each download now spends one "Unlock & Download" from a monthly allotment: 27 left, refreshing 14 Nov 2026. The confirmed limit is that allotment rather than credits, so agents download only takes the owner picks.