CRITIC_presentation-craft-anim-audio.md

pipelines/CRITIC_presentation-craft-anim-audio.md

CRITIC PASS (fresh-context refutation) — lens presentation-craft-anim-audio

Draft under review: scratchpad/lens_wave/LENS_presentation-craft-anim-audio_DRAFT.md

(a copy also sits in-repo at docs/pipeline_review/lens_wave_2026-07-28/LENS_presentation-craft-anim-audio_DRAFT.md).

Both repos read-only. Every verdict below was re-derived from the artifact, not from the draft's

summary of it: I re-ran each positive control myself, in Python (os.listdir + csv.DictReader,

never glob — the registry directories carry [ACTIVE v0.1] brackets, which glob silently eats as a

character class and returns an empty match, the exact "search that cannot match" class).

Headline: this is an unusually clean draft. Zero findings die. Sixteen of sixteen survive on

substance. Fourteen carry amendments; two of those amendments are load-bearing (PC-6's scope is

wrong-doc, PC-8's recommendation inverts the lesson it names). One care-inflation was caught by the

devil's-advocate pass (PC-12). Five items are contested and go to the director.

---

METHOD NOTE — what I re-verified independently

Every numeric claim in the draft was re-counted. Where my count differs from the draft's, the

difference is recorded as an amendment even when the conclusion is unaffected — a lens that will be

quoted at a contract freeze cannot carry counts that do not reproduce.

Code citations spot-opened and confirmed line-exact: HumanityAbilityFoundry.cpp:38-44 (the

"variant-selection key is SYNTHESIZED" comment), :42 (the 1-12 clamp), :44 (the %s_T%d printf),

:63-69, :98-107 (+5% / +8%); HumanityAttackLibrary.h (no montage/notify field; "A ref, not baked

art."; "the owner drives the windup->active transition through its OWN telegraph->strike timers");

HumanityCharacter.cpp:138-150 (the single ABP_Unarmed bind); HumanityGameplayCueNotifies.cpp:9-45

(five families, "art rides the 5090"); HumanityEncounterVolume.cpp:338 ("the color-independent SHAPE

channel (accessibility floor)"); factory_contract.json:154 (FR-006's tooth: NONE(...)).

Doc citations confirmed: COMBAT_PROGRAM_ADDENDUM.md:77-83 + :109-113; ASSET_DROPIN_CONTRACT.md:69,

:73-77, :110-113, :139-142, :187-191; COMPARATOR_LENS_PROGRAM.md:142-145 (the amendment,

verified first-hand, not via the realm program's quote of it); REALM_DESIGN_PROGRAM.md:1216-1219

and :1635-1640; DESIGN_GAP_REGISTER.md:102-108, :213, :216-217, :338-340, and cells 11/12

CLAIMED-QUEUED at :406-407; COMBAT_ENCOUNTER_SYSTEM.md:267 + :415; BATCH1_RULINGS.md:55 + :75;

QA_WATCHING_PROGRAM.md §4.3 ("only ordinals, bands and relations gate") + §4.4 (band_source,

tier_applicability, "blind per-region distinctness") + QS-20's motion_class enum;

PRE_5090_BUILD_PLAN.md:845-846 and :1979; T99_Translation_Audio.md §2.1/§2.2/§2.4 and Music Step 1's

seven-value theme_type enum at :292-293; flores_island.md Section 10 at :844-867.

---

PER-FINDING VERDICTS

PC-1 — DR-2 has no tier axis · SURVIVES (WEAKER, HIGHEST) · amended (counts)

Refutation attempted three ways and failed each time.

• *Could the tier axis be hiding under another name?* No. I scanned the whole 191-line contract

case-insensitively: presentation is a literal zero, and both hits for variant are the substring

inside "invariant". Both tier hits are "PROPOSAL-TIER" and "care-tier column". The eleven-row

table at §2 carries six columns and none of them is a tier, a variant, or a multiplicity.

• *Is this DESIGNED-UNBUILT?* No, and the draft's own rebuttal is the right one: the ASSETS are

5090-gated, but the CONTRACT CLAUSE is desk work that freezes pre-hardware. FR-006's own tooth

field records the hole verbatim (factory_contract.json:154) — a rider that records a hole is

evidence of the hole, not its closure.

• *Is the VFX N/A deliberate and therefore fine?* Verified real (§2 note at :73-77, open item 5 at

:187-191) and the draft's read is right: the swap guarantee is written for a single

placeholder→final swap and cannot express *n* variants per id.

Reuse-first: genuine. Every element of the fix already exists — the Batch-0 mechanism

(registry_extensions.json + fidelity_baseline.json, "proven across ten registries already",

:139-142), the degrade-never-null discipline in ResolveSound, and the five-state enum.

AMENDMENT (binding, cosmetic-severity): the positive-control counts are wrong — mesh_id_ref is

15 (not 11) and generation_status is 10 (not 7). Fix before landing; a lens quoted at a freeze must

reproduce.

PC-2 — no fidelity/ceiling tier anywhere · SURVIVES · CLASS AMENDED (HIGHEST)

Substance is airtight: realm appears zero times in 1920 lines of the doc that owns per-asset

budgets, against tier×162 / budget×126 / Nanite×76 in the same read. §5.6 F8's check column

literally ends "vs the class budget" — and there is no realm class to gate against. The scalability

rebuttal is correctly killed (§7.3 scales cost across hardware, not ambition across content classes).

AMENDMENT (binding): the verdict class contradicts the draft's own stated convention. Its method

note says "where a thing is RULED but has no home in the architecture about to freeze, it is reported

WEAKER" — and that is exactly PC-1's basis. PC-2 is structurally identical: the charter amendment

IS a landed ruling (COMPARATOR_LENS_PROGRAM.md:142-145, verified first-hand) and DR-2 carries no

field for it. Split the verdict explicitly:

• the requirement (reserve a realm tier) = WEAKER — ruled, homeless;
• the mechanism (any fidelity/budget class at all) = MISSING — genuinely undesigned.

This is not pedantry: a MISSING on a ruled item reads to a later agent as "nobody decided this," and

the whole point of the amendment is that somebody did.

PC-3 — two orthogonal tier axes, one word · SURVIVES (WEAKER, HIGH) · amended (+strengthened)

All three collisions verified: PresentationTier in code (mastery), the charter's realm tier

(fidelity), tier_applicability in QA (footage maturity). The naming ratification costs one line and

is unrecoverable after the freeze — the priority is right.

AMENDMENTS: (a) the citation for the realm-tier phrasing is REALM_DESIGN_PROGRAM.md:1179-1181,

not :1178-1180 (:1178 is the knowledge-progression lens row). (b) The finding is **stronger than

written** — there are at least two more live "tier" senses it did not name: the twelve-tier MASTERY

levels themselves (where Josh ruled 2026-07-27 that LEGENDARY/MYTHIC are **rarity grades, never tier

names** — CLAUDE.md known-issue 2), and readability_tier, which DESIGN_GAP_REGISTER.md:107 already

flags as "a false friend (twelve-tier inscription mastery)". Five senses, one token. Add them; the

recommendation does not change.

PC-4 — synthesized variant key, uncosted request size · SURVIVES (WEAKER, HIGH) · amended · one half CONTESTED

The core is verified line-exact: the key is synthesized (:44), the code says so in its own comment

(:38-41), and the registry carries 185 rows / 35 columns / 31 elements with no

presentation_tier_ref.

AMENDMENT 1 (sizing). The draft's "31 × 12 = 372" is the CLAMP CEILING, not the live key space.

tier_numerical spans 1.0-8.0 in the live registry, so today's synthesizable space is 31 × 8 = 248

— which is, tellingly, the "248+" the draft itself used in PC-1. State both: 248 live, 372 at the

clamp ceiling. The argument survives either number; the inconsistency does not.

AMENDMENT 2 (tooth scope). "The ladder is outside the audit that is supposed to prove the contract

holds" overstates it. W11's tooth (:110-113) asserts the stable-id UPROPERTY equals its registry

RowName; the ability's ability_id IS a RowName and IS audited. What is unaudited is the DERIVED

variant key. Say that precisely — the weaker true claim is still decisive.

CONTESTED (director/Josh). The "presentation BANDS, not 12 tiers" recommendation is the single

most consequential proposal on this lens, and it narrows a Josh-nodded hard requirement ("the 5090

asset passes generate TIERED variants as a hard requirement, not an option";

COMBAT_PROGRAM_ADDENDUM.md:111-113) in two directions at once: 12 rungs → 4 bands, and 31 elements →

10 base elements with hybrids inheriting. The feasibility case is strong and honest (hand-authoring

248 tiered Niagara systems solo is not a plan), the band boundaries are derived from real canon (the

5→6 and 11→12 mastery gates), and D10 is satisfied (tier-12 stays distinguishable from tier-1). But

this is a decision brief for the director, not a finding to apply silently — it touches a ruling

and it costs the 21 hybrid elements their own presentation identity. Route it through the standing

decision protocol with the status-quo option (12 rungs, generated) steelmanned.

PC-5 — frame data is timer-authored, no animation binding · SURVIVES (WEAKER, HIGH) · amended

The zero is real and I reproduced it exactly: 6 hits in 2 files, all of them the one stock

ABP_Unarmed bind plus an include. My positive control (213 hits / 32 files) is larger than the

draft's but proves the same thing — combat-presentation code is densely reachable and the animation

zero is a true zero. The header carries no montage, notify, or animation field; durations are floats

driven by the owner's own timers. The ruling ask (canon ROW authoritative, montage retimed to it) is

the right way round — it protects the ratified counter matrix and the QA-measurable telegraph lead.

AMENDMENT: animation is not wholly undeclared. T1_Build_Pipeline_Contracts [ACTIVE v1.0]:60

already names the authoring surface — "Animation Montages for combat hit reactions and

ability-execution animation," IK Rigs for retargeting, Motion Matching for locomotion. Cite it and

narrow the claim to what is actually absent: the authority rule between a purchased montage's

notify windows and the canon row's seconds. That is genuinely unruled, and PC-5's tooth (a QT-5

check that fires when they diverge beyond tolerance, shipping as a must-not-fire fixture with zero

montages bound) is exactly right and costs nothing today.

PC-6 — craft has no doctrine home · SURVIVES, but SCOPE AMENDED — the framing is wrong-doc

This is the one finding I came close to killing. The scan is real (T1_Combat_System_Spec:

animation×0, VFX×0 across 667 lines, with parry×16 / audio×5 / music×3 proving the read is

live — all six reproduce exactly). But the conclusion "animation and VFX craft has no doctrine home"

did not search the other T1 doc the wave runbook itself names. It does have a home:

_source/01_Tier_1_Foundation/T1_Ability_Tree [ACTIVE v1.4].md carries the presentation ladder in

five places — :16, :93, :527, :583, :1323 — and :583 enumerates the craft vocabulary

outright: *"The mastery presentation ladder per docs/COMBAT_PROGRAM_ADDENDUM.md §5 renders each

ability's animation refinement, effect density, audio confidence, and follow-through from the

capability's current tier… The ladder is presentation only."* T1_Combat_System_Spec:504 cross-refers

to the same rulings. So the ladder is anchored, owned, and firewalled against power double-dip at T1.

What survives, and it is still worth landing: the craft LAW is absent. A four-noun list

("animation refinement, effect density, audio confidence, follow-through") is an anchor, not a

doctrine — nothing rules commit-vs-cancel policy, root-motion policy, a hit-reaction taxonomy, or the

per-tier fluidity vocabulary the DR-2 prompt composer must quote as its input. Register entry 29's

adjacent camera/cut hole is real and correctly routed rather than duplicated.

I positive-controlled the surviving half corpus-wide (whole canon repo, all .md/.json/.csv,

case-insensitive): root motion = 1 hit — the draft itself; cancel window = 1 — the draft;

anim notify = 0; hit reaction = 3, all of them T1_Build_Pipeline_Contracts' tool declaration

plus the draft; montage = 20 across 8 files, 9 of them the draft. Control: follow-through returns

61 hits across 43 files in the same sweep, so the search plainly can match. The craft law is

absent from the entire corpus, not merely from the combat spec — that half of PC-6 is airtight and

should be re-stated with this evidence rather than with the combat-spec-only scan.

AMENDMENTS (binding): (1) rewrite the current-state paragraph so it does not claim homelessness —

claim *contentlessness*, and cite T1_Ability_Tree:583 as the existing anchor. (2) The reuse-first

recommendation must reconsider its owner: landing the craft law in T1_Combat_System_Spec while the

ladder's anchor lives in T1_Ability_Tree re-splits an ownership the corpus already resolved. Either

land it under the existing T1_Ability_Tree anchor (§7 / §5H.4), or state affirmatively why the

defensive-layer taxonomy makes the combat spec the better seat and cross-link both. CONTESTED —

director's call.

PC-7 — five cue families, a much larger state vocabulary · SURVIVES (DESIGNED-UNBUILT / WEAKER, MED-HIGH)

Every content number reproduces exactly (309 uassets, 14 SFX mostly _placeholder, 1 Combat

StateTree, 0 Niagara). The five families and the "art rides the 5090" comment are verbatim. The

recommendation — a closed FGameplayCueParameters set declared in the same DR-2 clause as PC-1, so one

Niagara system per family branches instead of forty existing — is the best economy argument on the

lens and costs one struct declaration now.

AMENDMENT (cosmetic): "each a subclass whose entire body is UE_LOG(...)" is imprecise — the

UE_LOG lives in UHumanityGameplayCueNotify_Base::HandleGameplayCue; the five subclasses each set

only FamilyLabel. The point (a proven channel with no art) is unaffected.

PC-8 — nothing rewards expressive variety · SURVIVES as a gap (MISSING, MEDIUM) · RECOMMENDATION INVERTED — must be amended

The gap is real, the divergence check is correct in substance (a presentation-only channel grants no

stat and shows no rank, so no-scoreboard is not re-litigated), and the refusal of the timid framing

("a style rank is tonally wrong") is exactly the calibration this repo asks for — that half passes the

devil's-advocate pass cleanly.

But the recommendation does not implement the lesson it names. DMC5's instrument produces variety

because it decays on repetition. The proposed input — HumanityChainMeter's unbroken-chain state —

rewards *the correct answer repeated*, which the draft itself identifies two paragraphs earlier as

"the exact opposite incentive to variety." Promoting the rendered band on an unbroken chain makes

repetition look better, not worse.

Second problem, verified in the header: HumanityChainMeter is a puzzle-boss instrument

(RatifiedPipFloor = 5, RatifiedPipStep = 2, RatifiedPipCeiling = 9, SnapToRatifiedRung,

ScalePipsForDifficulty). A channel keyed on it is invisible in ordinary combat — i.e. in almost all

of the slice.

AMENDMENT (binding): keep the finding and the "no new meter, no UI, no stat, no storage"

discipline, but change the transient input to a variety signal — e.g. a rolling count of distinct

abilities/verbs used in a short window, decaying when the same one repeats — and give it a non-boss

home so it reads in normal encounters. The chain state can remain a *second* input for boss fights.

CONTESTED — the input choice is a design ruling, not an editorial fix, and it should be settled

before this becomes a rubric row.

PC-9 — T1_Audio_Spec is a vacant authority seat · SURVIVES (WEAKER, MED-HIGH) · CLEAN

Verified: the file is 34 bytes (title line + blank), T99_Translation_Audio.md is 820 lines

with §2.1 licensing, §2.2 MetaSounds+Quartz, §2.4 the Kokoro/ElevenLabs split and §5's per-region

checklist all present as described, and register entry 22 names exactly these stubs. The verdict

"misfiled, not missing" is the honest one, and the authority argument is real (T1 wins conflicts; a

T99 operational doc and a region page's Section 10 cannot arbitrate each other).

Dedupe checked and cleared: entry 22 owns the *stub fact*; this finding contributes the **delegation

design** plus the three laws the stub should actually own. Not a duplicate. The DOC_MAP obligation is

correctly carried. No amendment.

PC-10 — adaptive score has a seam and no dramatic law · SURVIVES (WEAKER, MED-HIGH) · amended (three citations)

The decisive zeros reproduce: theme_type declares seven values including place and path

(T99_Translation_Audio.md:292-293), and place = 0 rows, path = 0 rows across all 12 registry

rows. 76 regions, 79 nodes and 3 integrity paths therefore have no theme row of any kind. The

reveal-discipline angle is real and well spotted — Music Step 1 does bind a CVD §17 check on

"Grand Sage pre-reveal thematic leak" (:308), so leitmotif economy is a §17 surface, not just craft.

AMENDMENTS (binding):

1. The theme_type breakdown is wrong: 5 character, 1 family, 2 event, 3 ending, 1 concept = 12.

The draft's "6 character" makes its own list sum to 13.

2. The positive control says character ×4; it is ×5. Control holds, count wrong.

3. AU-5 is at PRE_5090_BUILD_PLAN_VOL2.md:1279, not :1275 (:1275 is AU-3, the foley row). The

quoted AU-5 text is otherwise verbatim-correct.

4. The "workable for simple adaptive music… becomes unwieldy for complex interactive scores" quote

lives at T99_Translation_Audio.md:190 (§2.2), not inside Music Step 6; the SPIKE is rowed as

open spike #5, which *gates* Music Step 6. Re-cite as "§2.2, spike #5 against Music Step 6."

PC-11 — per-region ambient identity, join empty 0/76 · SURVIVES (DESIGNED-UNBUILT, MED-HIGH) · amended (count) · highest-feasibility item on the lens

Reproduced exactly: all three arrays — region_footstep_sfx_id_ref_array,

region_environmental_ambient_sfx_id_ref_array, region_ritual_context_sfx_id_ref_array — are

populated 0 of 76, flores_island's own row included, while the SFX side points at it. The

positive control is real (page_status / node_kind parse for all 76). Section 10 on flores_island.md

is as rich as described (:844-867). AU-4's and AU-13's absence lines are verbatim-correct

("ambient array declared unpopulated, island silent"; "arrays on 10 registries declared unpopulated").

The recommendation is the best on the lens: zero-token, zero-model, deterministic, through the

sanctioned harness/apply_*_columns.py writer pattern (apply_zone_catalog_columns.py,

apply_region_page_sections.py, apply_region_node_kind.py all exist), with the

docs/fidelity_baseline.json refresh in the same commit per the standing rule. Nothing here needs

authoring judgment — it is a pure inversion of data that already exists.

AMENDMENT (binding, changes the work order): flores_island carries 18 SFX rows, not 8. The

8 named ambient beds (SFX_00013-SFX_00020) are joined by SFX_00029 / SFX_00030 (Ebu Gogo

deep-gallery presence rustle; distant broken-speech murmur) — so the ambient-bed set the script must

invert for Flores is 10, and a script written to the draft's "8" would silently drop the two

creature-presence beds, which are precisely the Ch-2 stalk atmosphere. Note also that SFX_00019

(Manggarai Village Edge Dusk) is sfx_class=cultural_traditional with layering_role=ambient_bed, so

the inversion must key on layering_role, not sfx_class, or it drops that row too. Fix the number

and state the key — this is the one amendment that would otherwise ship a defect into a deterministic

writer.

PC-12 — R-A16: realms have zero audio identity; the selector drops the crown jewel · SURVIVES (MISSING, HIGH) · CARE AMENDMENT

Every registry claim reproduces exactly, and the selector trap is the sharpest decision-grade finding

on the lens: node_kind counts are `base 50 · revisit 9 · mythic_realm 9 · atlantic_waypoint 6 ·

story_anchor 1 · bookend 1, and fairy_realm` — the declared apex of the ceiling — is the single

story_anchor row. Any filter keyed on node_kind='mythic_realm' returns 9 and **silently excludes

it**. All ten realm rows carry empty ambient arrays, empty ambient_elemental_affinity, and blank

env_profile_status (fairy_realm: stub). 0 of 30 SFX rows and 0 of 12 theme rows anchor to a realm;

the 12 SFX region values are all earthly, exactly as listed.

I also confirmed the draft's own caveat about the safer selector: page_status='dimensional' returns

11 because it includes epilogue (node_kind=bookend). So *neither* existing column is a clean realm

selector — which strengthens reservation clause 4 (generation_tier='realm' written onto ten rows).

DEVIL'S-ADVOCATE / CARE AMENDMENT (binding). Recommendation (3)'s "reference-first sonic sourcing"

for living-cosmology realms imports the Panel-A image rule onto audio wholesale, and that is

care-inflation measured against the corpus's own ruled line. The repo already has the correctly

calibrated version, authored on the region page: *"No reproduction of protected sacred music — the

gong-waning register is reference-composed, never sampled from consecrated performance"*

(flores_island.md:856-857). That line bans sampling consecrated performance while explicitly

permitting reference-composition — which is the delegated-bolder posture. A blanket "reference-only"

read would functionally freeze generated audio for four of ten realms (Duat, Naraka/Meru, Xibalba,

Annwn per REALM_ANALYSIS_ART_PIPELINE §2.5) for no §17 reason: composing an original score in a

tradition's register is not depicting a living tradition's sacred object. Restate recommendation (3)

in the flores wording, cite it, and let the ruled care_tier column (which §2.5 already routes off)

carry any tighter case. The §17 floor is untouched by this amendment — no sacred-performance sampling

either way.

PC-13 — the slice's systems-teaching audio vocabulary · SURVIVES (COVERED) · CLEAN

All eight gameplay_cue rows verified verbatim, ids and names both: SFX_00021 Gather-Yield Confirm,

SFX_00022 Mineral Vein Strike, SFX_00023 Vril Recharge Fill Loop, SFX_00024 Vril Recharge

Fill-Complete, SFX_00025 Pip Fill Tick, SFX_00026 Pip-Wipe Damp Thud, SFX_00027 Chain-Complete

Resolution Swell, SFX_00028 Pips Reveal Subtle Appearance. Recording a COVERED so the lens is not

read as "audio is a hole" is exactly the honesty the register asks for. No amendment.

PC-14 — the ladder's pass/fail is a two-point check · SURVIVES (WEAKER, MED-HIGH) · amended (citations + reuse home)

D10's text is verbatim-correct at ENGINE_OPTIMIZATION_DOCTRINE.md:1411, and the binary criterion is

real: *"is the tier-12 variant identifiable as tier-12 at this class?"* — a tier-1-vs-tier-12

question over a 12-rung ladder implemented as a linear ramp. A ladder flat from rung 2 to 11 passes

it. The gap is real and the QA machinery genuinely does not cover it.

AMENDMENTS: (1) the criterion sentence lives in §6.5 (:1623-1628), not §5.4 — §5.4 holds D10's

rule text (:1411) and D6 (:1407), whose Check column already says "per-tier capture." Say it

precisely: the per-tier captures are already required; what is missing is an **ordinal criterion

across them**. That is a smaller, truer, and much cheaper ask — and it is exactly what the

recommendation proposes. (2) Name the reuse home explicitly: QA_WATCHING_PROGRAM.md:387 row RB-9

("Presentation — HUD BUILD-NOW · cinematic PARKED · audio SKELETON-NOW") is the existing presentation

board this row lands on. The must-fire fixture (a synthetic ladder whose rung 3 copies rung 2 must

FAIL) obeys §4.4's self-testing-gate law by construction.

PC-15 — no positive standard for a signature look · SURVIVES (WEAKER, MED-HIGH) · amended (dedupe + nuance)

Both positive controls reproduce exactly. The observation is the most transferable one on the lens

for a solo dev whose art is generated, and §6.4's pre-hardware board is verified explicitly shippable

with no GPU ("it should be shipped rather than deferred", :537), so the recommendation is cheap and

real.

AMENDMENTS: (1) "every one of those grades *against slop* — a negative standard" overstates it.

§5.4 grades "against reference-title bars" and runs an unlabeled-plate assignment test — a real

positive-ish instrument, but a per-realm distinctness test, not a cross-title signature. Narrow the

claim to that and it is unassailable. (2) Dedupe risk, must be routed: aspect R-A15 ("Per-realm

art direction and world identity") is CLAIMED by the realm doctrine wave that is RUNNING right now

(DESIGN_GAP_REGISTER.md:447; REALM_ANALYSIS_LENS_GAPS_2026-07-27.md:96-99, which says in terms that

the presentation lens "covers animation and VFX craft, not world-identity art direction"). The

cross-realm SIGNATURE is genuinely outside R-A15's per-realm scope — but it must land as the

cross-realm head of the same art bible that wave is producing (§4.2), not as a parallel artifact,

or two in-flight waves author competing art bibles. CONTESTED — boundary call for the director.

PC-16 — 31 elements, no visual law · SURVIVES (MISSING, MED-HIGH) · amended (quote fidelity)

The zeros are the strongest positive-controlled result on the lens: ELEMENTS_MATRIX_RECOVERED.md

returns 0 for all five of color / colour / silhouette / VFX / visual / palette across 105 lines

while element returns 17 in the same read. T0_Element_Effectiveness is 124 rows of

attacker/defender/effectiveness/axis/source with nothing presentational; the ability registry has 31

element values and no visual column. The accessibility floor is verified ruled and built

(COMBAT_ENCOUNTER_SYSTEM.md:267, :415; BATCH1_RULINGS.md:55, :75;

HumanityEncounterVolume.cpp:338).

Dedupe checked and cleared: DESIGN_GAP_REGISTER.md:102-108 covers **silhouette-at-range for creature

families and wayfinding landmarks; this finding covers per-element presentation identity**. Different

surfaces; the draft's "this is the other half" is correct.

AMENDMENT (quote fidelity): "readable without color alone" is presented in quotation marks with a

file:line citation, but it is a paraphrase — the sources say "colorblind/multi-channel vril-polarity

readability" and "vril-polarity multi-channel readability is a **non-negotiable L1-playability

requirement**". Either quote the real strings or drop the quotation marks. On a §17-adjacent

accessibility ruling, a paraphrase-in-quotes is the kind of thing that gets re-quoted as canon.

---

THE REALM-TIER RESERVATION (the seven clauses) — verified

Each clause was checked against the machinery it claims to ride, and all seven ride real machinery:

• Clause 1 — the Batch-0 mechanism is real and is described in the contract as "proven across ten

registries already" (:139-142). The "enum value must exist at freeze" argument is **inference, not

citation** — sound inference (the standing rule is that any registry edit refreshes

fidelity_baseline.json in the same commit), but label it as reasoning so it is not later re-quoted

as a cited fact.

• Clause 2 — tier as request parameter, never a path segment: correctly preserves §1's drop-in

invariant. Verified against §1 and the §2 convention paths.

• Clause 3 — generation_prompt_hash and regeneration_trigger both exist in the declared

assetgen block (:81-83), so the self-enforcement claim is real.

• Clause 4 — ten rows verified individually (nine mythic_realm + fairy_realm as story_anchor).

Strengthened by my finding that page_status='dimensional' also returns epilogue, so no existing

column is a clean selector.

• Clause 5 — the naming ratification; see PC-3's amendment (five senses, not three).
• Clause 6 — realm×0 across 1920 lines is confirmed, and §5.6 F8's "vs the class budget" is

verbatim. The clause's "either a realm column or an explicit 'realm inherits region budgets, ruled'

line — but the silence is not fine" is the right shape for a doctrine doc.

• Clause 7 — §4.4's tier_applicability and the "blind per-region distinctness" audio measure are

both verified, as is the audio rubric's SKELETON-NOW / ARMED-AT-HANDSHAKE status. The realm band

must be declared at freeze or a ceiling surface passes against a region floor.

Cost honesty ("clauses 1-5 are schema and desk work, GPU-free") is accurate on everything I checked.

Divergences and the refuted-and-dropped list — verified

The five self-refutations in "Refuted and dropped" are all TRUE refutations, and this is the section

that most raises my confidence in the draft's method: the greybox ladder seam is live and recorded at

PRE_5090_BUILD_PLAN.md:845-846; audio design exists at 820 lines plus AU-1..AU-13 + QT-AU +

DR-AU1/DR-AU3; the QA program's Pillar 1 sequence capture, two strips (§1.2) and burst clock (§1.3)

are real, as is QS-20's motion_class enum; DR-AU1 names the music lanes; and

HumanityCombatFeelSubsystem.cpp exists with SetGlobalTimeDilation and CF-1/CF-6/CF-8 all present in

the file. An audit that kills its own best headlines is an audit worth landing.

Pre-registered divergences: all four respected. One citation amendment — no-scoreboard is not in a

"§Covered/Divergent" section; it lives as entry 8's bounded constraint at DESIGN_GAP_REGISTER.md:169

and :171 ("a visible skill-up scoreboard *would* contradict CVD §12.1's practitioner framing"), plus

the wave runbook's own guard. Re-cite; the substance of PC-8's divergence check is unaffected.

Devil's-advocate-against-timidity — full pass

Run over all sixteen findings and the seven clauses. One care-inflation found (PC-12's

"reference-first sonic sourcing" — amended above to the ruled flores line). Everything else passes:

PC-8 correctly refuses the "a style rank is tonally wrong for a grounded rated-R world" instinct and

keeps the mechanism while dropping DMC5's presentation; PC-16 binds hue to an already-ruled

accessibility floor rather than inventing a new constraint; PC-12's §17.1 framing is collective and

does not defang anything; nothing anywhere proposes removal where rename/reframe would serve. No

finding softens the R-17+ bar, and no recommendation treats a monster, a deity, or a hostile faction

with kid gloves. The §17 hard-line floor is untouched throughout.

---

CLOSING VERDICT

Killed: 0 · Survived: 16 · Amended: 14 · Contested: 5

Contested items for the director:

1. PC-4's band collapse (4 bands × 10 base elements vs 12 tiers × 31 elements) — narrows a

Josh-nodded hard requirement and costs the 21 hybrids their presentation identity; feasibility case

is strong. Decision brief, not a silent apply.

2. PC-2's verdict class — MISSING vs the draft's own ruled-but-homeless = WEAKER convention.

Recommend the split verdict (requirement WEAKER / mechanism MISSING).

3. PC-6's owner — T1_Ability_Tree already anchors the ladder at five sites; landing the craft law

in T1_Combat_System_Spec re-splits a resolved ownership.

4. PC-15 vs R-A15 — the realm doctrine wave is RUNNING and owns per-realm art direction; the

cross-realm signature must land as the head of that wave's art bible, not beside it.

5. PC-8's input — chain-state promotes repetition, the inverse of the lesson; the variety signal is

a design ruling.

LANDING RECOMMENDATION: GO-WITH-FIXES. Land the audit after applying the fourteen amendments and

routing the five contested items. Three amendments are must-fix before anything downstream consumes

this: PC-11's Flores count (8 → 10 ambient beds, keyed on layering_role not sfx_class — a

deterministic writer built to the draft's number ships a real defect and drops the Ch-2 Ebu Gogo

presence beds), PC-6's wrong-doc scope (as written it tells a later agent the ladder is homeless

when T1_Ability_Tree owns it), and PC-12's care line (as written it would freeze generated audio

for four realms for no §17 reason).

Priority ordering endorsed as written: **PC-1, PC-2 and the seven-clause reservation are genuinely

contract-critical and must precede the DR-2 freeze** — the reservation is the correct concrete

discharge of the binding charter amendment, it rides only machinery that already exists, and every

clause in it becomes materially more expensive the day after the freeze. **PC-11 should land first

regardless** — it is zero-token, zero-model, needs no authoring judgment, and turns QT-AU's non-silent

floor from vacuously green into a real measurement on the slice's own region.

Generated by harness/site/structure_site.py — the URL path is the repo path. review root