CRITIC_ac-traversal-stealth-legibility.md

pipelines/CRITIC_ac-traversal-stealth-legibility.md

CRITIC PASS — ac-traversal-stealth-legibility (fresh-context refutation)

Draft under review: LENS_ac-traversal-stealth-legibility_DRAFT.md (16 findings).

Critic method: every cited file OPENED; every numeric claim RE-COUNTED from the live artifact;

every MISSING re-searched independently with my own positive controls; dedupe re-checked against

docs/DESIGN_GAP_REGISTER.md v1.2 (not v1.1 — see AMEND-A); devil's-advocate-against-timidity

run on all recommendations. Repos read-only; nothing written outside this scratchpad.

Status: IN PROGRESS — extended as verification completes (write-early rule).

---

A. Independent re-verification log (what I opened, what it said)

Canon repo

Game repo (build evidence)

---

A2. Second verification batch (the claims the first batch did not reach)

My independent corpus sweep (562 files under docs/ + _source/01_Tier_1_Foundation/ + registries/, excluding book_extractions/archive/research_notes)

Zeros, each with a live control in the same run and same tool:

climb[ _-]?affordance        0        social stealth               0
climbab                      1        blend(ing)? (in|into)        0
heat[ _-]level|wanted[...]   0        traversal_affordance         0
down-climb|rappel|abseil|                perception_tier              0
controlled fall|drop-hang    0        notoriet*                    2  (both = the gap being named)
CONTROLS: composure 348/64 · agility course 111/58 · identity_hiding 518/114 · disguis* 149/51

The controls reproduce the draft's numbers to the digit. The searches can match; the zeros are real.

The one exception is below (AMEND-F).

---

B. VERDICTS — 16 findings

**Killed outright: 0. Survived: 16 (7 amended). One recommendation limb struck as a re-litigated Josh

ruling.** The audit's evidentiary hit-rate is unusually high: of ~90 discrete citations I re-opened,

every substantive claim held except three, and all three are recorded below.

CONFIRMED — stands as written (9)

AC-2 · climb capability + no vertical exit test · DESIGNED-UNBUILT · CONFIRMED.

The strongest finding in the audit and I could not dent it. Every load-bearing element verified

independently: MantleMaxHeight = 130.0f ("tallest vaultable ledge (uu)"), climb count in

HumanityCharacter.cpp = 0 against 24 mantle hits, W-MOVE's exit line is the canopy-HORIZONTAL

course (MV-3's own canon quote), and TODO_CANON L1205 already flags a 700 m cliff-vertical traverse

as intentional-SEVERE and un-sculpted. The finding's real content is the *plan defect* — the capability is

scheduled, the verification is not — and that distinction is correct and is exactly what the

honest-tier register exists to catch. The fallback in contested item 3 (tag the vertical sites

traversal_capability_pending so the slope teeth stop certifying them) is the right de-risk if W-MOVE

cannot absorb the mode. Land as-is, priority HIGH.

AC-4 · social stealth has no moment-to-moment mechanic · WEAKER · CONFIRMED.

WS_025 verbatim including the NPC_recognition_system consumer that no system answers to; 12

stealth-dominant chapters exact; social stealth/blend in = 0 against identity_hiding = 518. The

dependency on register #5 is stated, not duplicated. The self-refutation (b) — "we are villages, not

cities" — is answered with the correct data. Land as-is.

AC-6 · the stealth spec is one sentence · WEAKER · CONFIRMED.

Verified to the line: L366 is the *entire* stealth spec, one bullet in a ten-item list, against a full

Block/Parry/Dodge section at L417-464; and the built verb has no LOS, concealment, or positional

precondition (I read the constructor — it is four lines). The "greybox tier" self-refutation is correctly

rejected: a missing precondition is a *shape* defect, not a tuning one. One nit: the "no pure stealth

boss" text is CVD L220 (Pillar 8's gameplay-product bullet), not L217 (the header). **Land, and see

the timidity note under AC-12 — the two should be authored as one act.**

AC-7 · notoriety has no ladder, no clear-verb, no player-side state · WEAKER · CONFIRMED.

notoriet* = 2 hits and both are the gap being *named*; heat_level|wanted_level = 0; awareness_tier

64/64 at 35/16/7/6 and it is operative-knows-family, not network-knows-player, exactly as claimed; the

Antagonist Hunter Trigger is ratified (CVD L168-170 via _VOL2 L517-518) and the two EVADE near-miss rows

exist. The extension boundary against #19 (rungs · named persistent pursuer · clear verb) is real and

stated. The devil's-advocate pass is correctly calibrated: today the ruled R-17 sandbox impulse

produces silence, and silence-by-omission is the defanged outcome. Nothing here is care-inflated.

Land as-is.

**AC-9 · markerless navigation has no written doctrine while U11.21 builds against nothing · WEAKER ·

CONFIRMED.** This is *not* a re-litigation of the pre-registered divergence — the draft endorses the

divergence and audits the absence of an enforceable rule, which is a different object. U11.21 is real,

BUILD-NOW, wave TRANS; U11.22 governs only fast-travel; T1_UI_UX_Spec is 35 bytes; the CVD carries no

navigation policy. The recommendation lands inside already-queued U11.31/rank-20 work. Nits: the

divergence lines are register L207 and L227 (draft said L205/L226). Land as-is.

AC-10 · 53 vista sites, no mechanic that pays for stopping · DESIGNED-UNBUILT · CONFIRMED.

My own scan returns 53 vista sites across 40 chapters and 380 climactic_affordance lines —

identical to the draft. U11.25, RB-8's vista half, and the T-C awe-terrace pattern all verified, and the

draft's point is correct that none references the others. Correctly verdicted DESIGNED-UNBUILT rather

than MISSING, and the anti-collectible-bloat guard (L2 payload, zero map pins) is explicit. Land as-is.

AC-14 · rank 7's scope excludes traversal while the EG registry holds two rows · WEAKER · CONFIRMED.

The cheapest high-leverage item in the audit and every number is exact: 2 rows × 18 cols,

environment_grammar_refs 0/11 on a catalog where footfall_surface is 11/11, rank 7 verbatim

scoped to "every slice combat surface". Extra support I found that the draft missed:

T99_Translation_Combat §2.7 cites "WS7 environment-grammar seeds EG_0001-0004" while only

EG_0001-0002 exist — the registry is thinner than its own ratified contract assumes. **Land as-is,

and land it BEFORE rank 7 starts** — this is a one-word plan edit that prevents a second pass over

12 regions.

AC-16 · the QA loop has no affordance-legibility instrument · WEAKER · CONFIRMED.

Best-cited finding in the audit; all six QA-program citations verbatim at the cited lines, including

the Tier-2 driver's own positive control. First-class under the Josh Gate by the runbook's explicit

rule. The fixture set genuinely is free (the 70 traversal + 26 stealth + 53 vista rows I counted

myself). Land as-is, priority HIGH.

AC-3 · descent has no grammar · MISSING · CONFIRMED.

My independent sweep with a stricter verb list (`down-climb|rappel|abseil|controlled fall|controlled

descent|drop-hang|ledge hang`) returns 0 with controls live; MV-9 is verified as a damage model with

no verb; §5D.2 L267 mentions descent only as a failure mode. Correct. *Observation, not an objection:*

AC-3 mints nothing of its own — it is one enum value and one clause inside AC-1/AC-2. It is honest to

land it as a sub-item of those two rather than as an independent work order.

AMENDED — survives with the amendment (7)

AMEND-A · AC-1 · re-class MISSING → WEAKER, and re-point the reuse target.

The climbability gap is real (traversal_affordance = 0 on my own sweep; nothing declares a climbable

surface). But the draft's self-refutation missed the decisive counter-argument, and I found it:

docs/proposals/systems/environment-combat.md §5.1 L132-139 already designs a **six-class affordance

taxonomy whose sixth class is literally TRAVERSAL SURFACES** — "ice (slide/low-friction), mud (slow),

water (swim/soak), scree (slip), updrafts (lift), narrow beams (balance)… they *compose with the MOBILITY

operators*". The draft names "ELEVATION, COVER, CONCEALMENT, and three more" and never notices that one

of the three is its own subject. Two consequences:

1. Class. A designed traversal-affordance class that is *silent on climbability* is WEAKER, not

MISSING. The finding survives; the verdict word changes.

2. Reuse-first defect. Minting a parallel traversal_affordance_class enum beside

tactical_affordance_class splits a taxonomy that is already unified. The reuse-first move is to

extend §5.1's TRAVERSAL SURFACES class with the climb/ledge/descent values and let the registry column

carry the whole six-class vocabulary. The *actual* authoring defect the finding should name is

sharper and better: L138 types "cliffs/pits" under HAZARD VOLUMES only — canon currently says a

cliff is something you fall off, never something you climb. That single typing decision is the

affordance-law gap, stated in the taxonomy's own language.

AMEND-B · AC-11 · the load-bearing registry count is wrong by an order of magnitude.

The draft: "T0_Vril_Site_Registry [ACTIVE v0.1] carries 42 rows with site_class=waterfall."

Actual: the registry has 42 rows TOTAL, of which 3 are site_class=waterfall —

VS_CH02_001 (Liang Bua, recharge_passive), VS_CH03_001 (Mama Klara's ritual site,

recharge_ritual), and VS_CH7677_001 — the Grand Sage Waterfall, Ch 77, source_authority CVD §17.9.

The remaining 39 rows are dimensional_threshold ×9, pyramid ×6, vortex_anchor ×6, megalith ×5, tomb ×3,

and so on. Consequences:

• The proposed reuse set ("the site_class=waterfall rows as the affordance set") is **two ordinary

waterfalls and the reveal site**, not forty.

• The reveal-safety argument in contested item 2 — "a move the player has performed at forty ordinary

waterfalls carries no signal" — is built on the false count and does not currently hold. On

today's data the player's waterfall interactions are Ch 2, Ch 3, and the Ch-77 reveal. That is

reveal-*adjacency*, not reveal-camouflage.

• Separately, T1_Combat_System_Spec L43 does not say what the draft says — L43 is about Tier 6

Weapons_Detailed docs and Tier 2 ChapterDesign ownership. The real build-sanctum/recharge-site line is

L571 ("Quick-change is diegetic and sited: it happens at a waterfall or a vril recharge zone…").

The finding is not dead: waterfall-as-recharge is genuine canon (T1_Vril §5.1 L137-146; CVD L1003 via

§18.2), and T1_Vril L145 explicitly says "multiple chapters across the arc carry waterfall recharge

sites; per-chapter detail in T2" — a tier that was deleted. So the honest re-framing is: **the

signature-move idea has a real canonical hook and a data population that does not exist yet, deferred to

a deleted tier.** Re-verdict the data half WEAKER, drop priority to LOW-MEDIUM, and hand the reveal call

to the director with the true numbers. *If the director wants a stricter landing bar, AC-11 is the one

finding I would cut.*

AMEND-C · AC-12 · the absence-claim is false, and the finding gets STRONGER for it.

The draft: "backstab|assassinat*|takedown|stealth kill returns **no player-verb definition anywhere in

the corpus** — only the two design lines above and per-boss *prohibitions*."

My sweep returns 40 hits across 18 files, and docs/original_predclaude/extracted/Weapons_Trickery.md

L279-332 contains a complete player-verb definition for the Dagger:

• "Ability 2: Backstab — from behind, triple damage" (a positional precondition, already authored)
• "Ability 4: Throat Strike"; "Mastery (Lvl 8): Ghost Step — all dagger attacks from stealth have

zero detection chance for a brief window"

• Legendary "The Venetian Stiletto" (Ch 28 Renaissance-Italian intrigue questline), whose ability

scales with *intelligence gathered on the target through quests and inscriptions*

• Balance line: "Assassination specialist. Underperforms in open combat. Excels above all weapons in

stealth context."

This is not stray prose: T1_Combat_System_Spec L192 names pre-restructure Weapons_Trickery as the

authority for the canonical 36-Trickery roster. And T0_Weapon_Registry carries WPN_039 Dagger

(TRICKERY/BLADES) with damage_type blank, canonical_chapter_acquisition blank — the schema has no

ability columns at all. So the correct verdict is ORPHANED, not absent: the verb, its precondition,

its stealth-detection mastery, and its legendary assassination weapon are all authored in the recovered

source and none of it reached the live registry, the T1 spec, or the runtime.

This makes the recommendation *cheaper and better*: restore the canonical dagger ability set rather

than invent a precondition for UHumanityStealthAbility.

Second amendment, same finding — a better doctrine home. The draft grounds the timidity correction in

CLAUDE.md (a working-agreements file). The reuse home is docs/translation/T99_Translation_Combat

§2.6, a ratified engine-translation section titled *"The R-17 bar — real fights, don't defang; §17.1

care is collective, never individual-unkillability"*, which already states verbatim that §17.1 "does

NOT make individual antagonists/combatants who belong to that culture unkillable", and pairs it with

the HL_0003 deterministic wall protecting communities. The timidity correction is therefore

*applying an existing ruling*, not making a new one — which materially lowers the director-risk on

contested item 1. Devil's-advocate-against-timidity: I endorse this finding without reservation. The

blanket per-boss "no assassination" is care-inflation the repo's own punchlist already flagged, the §17

floor is untouched, and the fix is rename/reframe-by-scoping, never removal.

AMEND-D · AC-13 · one limb struck as a re-litigated Josh ruling; the rest stands.

The draft: "EVADE appears as a win_condition value with no definition anywhere", and recommends

"define EVADE in the win-condition enum documentation the same way BYPASS is defined."

Decisive counter-evidence — docs/translation/T99_Translation_Combat.md L512-515:

"EVADE — the undefeatable near-miss class: escape-by-wit-and-cost IS the win; the encounter is

explicitly not winnable by defeat (the Ch 54/56 Verderer Hunter-Trigger near-misses,

BE_0054_S4/BE_0056_S3). [RULED — D-EVADE-TOKEN, Josh 2026-07-20; minted on the back-fill's

evidence that the 9-verb enum could not honestly express the class; the interim BYPASS retired]"

The definition names the exact two rows the finding is built on, sits in the ruled 10-verb enum

(L499), and is corroborated at BUILD_PLAN_END_TO_END.md L547 ("D-EVADE-TOKEN (2026-07-20, the 10th

verb) … closes … the win-condition taxonomy"). This is a Josh ruling, and the wave rule forbids

re-litigating one. Strike the enum-definition limb from the recommendation.

Everything else in AC-13 survives and I verified it independently: 281 rows, traverse ×8, passage

×10, all eight IDs and win-conditions exact; two of the eight inside the slice; T1_Combat_System_Spec

§Combat Variety's ten categories (L360-369) carry no traversal/pursuit entry; W-SPACE's

approach→threshold→arena→payoff is a standing-fight shape. The pursuit/passage arena grammar is the

real gap and is worth the finding on its own. Re-class the finding MISSING → WEAKER accordingly (the

win-condition taxonomy exists; the *encounter shape* does not).

AMEND-E · AC-5 · canon already defines the perception bands; drop the invented enum.

The gap (no data column) is real and verified: T0_Creature_Roster 150 × 52 with zero

sight/percep/aggro/detect/awareness/patrol columns, TG-12 verbatim, C1.23 verbatim, Composure_Stealth

live. **But the draft treats the perceptive/non-perceptive distinction as undefined and invents a

four-band ladder (none | mundane | trained | vril_perceptive).** Canon already rules it:

T1_Ability_Tree L884, the canonical Cloaking entry —

"…becoming difficult to register for non-perceptive observers. **Effective on guards, common NPCs,

and animals; ineffective against substrate-perception capable NPCs (Throat-bonded antagonist family

vessels, Crown-bonded allies, vril-detection technology).**"

That is a binary plus a named carrier list, already authored. The schema pass is therefore a

transcription, not a design decision — cheaper, faster, and it cannot drift from canon. Replace the

four-band enum with canon's own two bands + the carrier flags. This *strengthens* the finding: it removes

the only part of it that needed a ruling. (Also: C1.23 is at PRE_5090_BUILD_PLAN.md L1759, not L1720.)

AMEND-F · AC-8 · one bonding-requirement slip.

The draft says both §5F.3 passives require "Third-Eye-bonded (Ch 38) plus Throat-bonded (Ch 56-57)".

L365 (Static awareness) is Crown-bonded plus Third-Eye-bonded; only L369 (Hostile substrate

detection) is Throat + Third-Eye. The conclusion is unaffected — Third-Eye is Ch 38, so both remain

post-slice — but the sentence should be corrected. Everything else verified exactly, including the

A_E_017 Ground Sense row (tier 2.0, perception_passive, blank unlock_chapter_ref) and the 153-blank

count. The reveal-discipline self-refutation (c) is sound: Ground Sense reads rock at tier 2 and does not

touch HL_0028's tier-6 gate.

AMEND-G · AC-15 · one misplaced accessibility citation.

The draft points the multi-channel legibility discipline at T99_Translation_Combat §2.6. §2.6 is the

R-17 bar section (see AMEND-C); the color-independent shape/audio-channel rule is the line at L239

("carries an audio + shape channel, not color alone"). Re-point the clause. The finding itself is

verified — §5D.2 L267 verbatim, MV-7 L556 verbatim, footfall_surface 11/11 with only a movement-modifier

consumer, and the NATURAL_VOICE §3 quote (L35-37) is verbatim accurate. *Small strengthening note:*

environment-combat.md L139 already says traversal surfaces "compose with the MOBILITY operators", so

the design tie exists at proposal tier — the read channel is what is absent, and the finding should

say so in those terms.

---

C. Cross-cutting checks

Line-number drift (systemic, minor). Six cites are off by 1-45 lines while the quoted *content*

matched at the true location: W-SPACE COMPLETE L815→L854; C1.23 L1720→L1759; register

L205→L207 and L226→L227; TODO_CANON L1206→L1205; CVD L217→L220. One cite pointed at

unrelated text entirely (T1_Combat_System_Spec L43 — see AMEND-B). Recommend a mechanical line-refresh

before landing; none of these changes a verdict except L43.

Dedupe / pre-registered divergence. Clean. I re-read all eight cited register entries (#1, #4, #5,

#17, #19, #22, #24, #32) plus the v1.2 matrix. Every boundary the draft draws is real and none of the 16

findings re-mints a register entry. Note the draft calls the register "v1.1, 45 entries" — it is now

v1.2 (the 2026-07-28 aspect×lens fold at L381). This lens is the *claiming* lens for matrix cell 24

(L420) and its three NEW aspects are already minted as ac-traversal rows at L435-436, so the audit is

in-scope by the register's own board. The only genuine re-litigation is AC-13's EVADE limb (AMEND-D).

Devil's-advocate against timidity. No finding is care-inflated; none proposes removal over reframe;

the §17 floor is untouched throughout. AC-12 is a *correction* in the bold direction and I endorse it

(strengthened by §2.6). AC-7's pass is correctly reasoned. **One place I would push BOLDER than the

draft:** AC-6's stealth-spec recommendation lists only two answers (Cloaking = active, crowd = free) and

defers the lethal takedown to AC-12's later doctrine line. That sequencing ships a defanged spec that

AC-12 then has to re-open. Author AC-6 and AC-12 as one act — the takedown belongs in the spec

section's outcome list from the first draft, scoped by §2.6's collective-care rule and HL_0003's

deterministic wall.

Reuse-first audit. 14 of 16 recommendations land on an existing column, wave, spec section, or

rubric. The two exceptions are corrected above: AC-1's parallel enum (AMEND-A → extend the existing

six-class taxonomy) and AC-5's invented four-band ladder (AMEND-E → transcribe canon's two bands).

The draft's "new systems proposed: 0" claim holds after those amendments.

---

D. CLOSING VERDICT

Killed: 0 · Survived: 16 · Amended: 7 · Contested: 4

Landing recommendation: LAND, with the seven amendments applied.

This is a high-quality audit. I re-ran every positive control independently and they reproduce the

draft's numbers to the digit (composure 348/64 · agility course 111/58 · identity_hiding 518/114 ·

disguis* 149/51 · 53 vista sites / 40 chapters · 380 climactic_affordance lines · 141 combat / 26 stealth

site rows · awareness_tier 35/16/7/6 · 281 boss rows with traverse ×8 / passage ×10 · 153 blank ability

unlocks · 2 EG rows · environment_grammar_refs 0/11 · MantleMaxHeight 130 · climb count 0). Of ~90

citations I re-opened, three failed substantively — a false registry count (AC-11), a false absence-claim

over a recovered source (AC-12), and a re-litigated Josh ruling (AC-13's EVADE limb) — and two of the

three make their findings *stronger* once corrected.

Land order: AC-14 first (a one-word plan edit that must precede rank 7), then AC-2 + AC-16

(the slice-gating pair — capability and the tooth that proves it), then **AC-6 + AC-12 authored as one

spec act, then the rest. AC-11 is the weakest survivor** and is the finding to cut if the bar tightens.

Contested — for the director

1. AC-11's reveal adjacency, on corrected data. The registry holds 3 waterfall rows, one of which

IS the Ch-77 Grand Sage reveal site. The draft's "making it ordinary is the reveal-safe choice" rests

on 40 ordinary waterfalls that do not exist in data (T1_Vril defers them to the deleted T2 tier). The

call belongs to whoever owns the reveal ladder, and it should be made on these numbers, not the

draft's. Gate #25 grand_sage_silence is unaffected either way.

2. AC-1's class demotion (MISSING → WEAKER). I am ruling it on the discovery of the designed

TRAVERSAL SURFACES class. A director who reads "climbability specifically has no representation

anywhere" as the finding's true object may prefer to keep MISSING. Either way the reuse target must

change to extending §5.1's taxonomy, and the sharper defect to name is L138's cliffs-as-hazard-only

typing.

3. AC-12's boss-prose sweep. Lower-risk than the draft assumed now that T99_Translation_Combat

§2.6 is identified as the governing ruling — the sweep applies an existing rule rather than making a

new one. Still apply at director level per the subagent-classifier lesson, since it edits canon prose.

4. AC-2's vertical exit test vs the W-MOVE window. I endorse the draft's own fallback: if the climb

mode cannot land, ship the tagging half (traversal_capability_pending on vertical zone rows) so the

slope teeth stop silently certifying sites the player cannot traverse. That is the honest-tier

register doing its job, not a schedule fight.

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