CRITIC_hades-death-run-realm-identity.md

supernatural/CRITIC_hades-death-run-realm-identity.md

CRITIC — hades-death-run-realm-identity (fresh-context refutation pass)

Draft under review: LENS_hades-death-run-realm-identity_DRAFT.md (20 findings, HD-1..HD-20)

Critic posture: refutation-first. Every citation opened. Every zero re-run with my own tool.

Every MISSING tested against DESIGNED-UNBUILT. Every finding tested against

docs/DESIGN_GAP_REGISTER.md (pre-registered divergences + already-covered entries).

Plus the DEVIL'S-ADVOCATE-AGAINST-TIMIDITY pass.

Headline: this is an unusually well-evidenced draft. Citation accuracy is near-perfect —

I re-ran every positive control independently and each reproduced to the exact count. The

kills/amendments below are almost all register-collision failures (the author did not check

DESIGN_GAP_REGISTER.md against his own findings) rather than fabricated evidence.

---

VERIFICATION LOG — citations I re-ran myself

No fabricated citation found anywhere in the draft. That is worth stating plainly.

---

THE DECISIVE CONTEXT THE DRAFT DID NOT READ

docs/DESIGN_GAP_REGISTER.md grew from 463 to 1036 lines during this wave — the v1.3 fold

(P1-B + realm doctrine, entries 46-93) landed while this lens was drafting. The author checked the

register's divergence lists but did not check v1.0 entry #9 or the v1.3 realm entries. That is

the source of both kills and most amendments. Three pre-registered items govern this lens:

1. Adjudication 4 / claim-board L945 — the boundary is PINNED IN OUR FAVOUR.

"RecordRetryLatency lands once, on that side; the RUN-ECONOMY half is the Hades lens's and

its P1-C brief carries the boundary note" [L943-946]. HD-8's boundary declaration is not just

legitimate, it is pre-blessed. Cell 13 reads "boundary PINNED — execution-surface half landed

with sm64 (§v1.3 entry 53)" [L408]. No Part-A finding is out of scope.

2. Entry #9 (v1.0) — "Rest / safe-site session-anchor loop", WEAKER, "Lowest priority", whose

recommendation is already "make the safe-site a readable anchor beat (save, restock,

companion talk, the vril-recharge already there)" and which carries a **pre-registered

divergence: "WoW's 'zone heartbeat between excursions' is a deliberate divergence** — WGM

§4/§6.7 walked-geography makes Ch 1-58 a continuous pilgrimage… do not import it" [L173-190].

3. Entry 73 (v1.3, L-R2, ranked HIGHEST) — "The charter's ruleset row is per-WORLD while the

corpus authors its rules per-SITE and per-ENCOUNTER — the machine layer cannot tell a realm from

a place… the first call in the Wave-2 mint" [L672-676]. And entry 77 — "The entire realm

machine layer is specified and unbuilt… **Recorded as the honest baseline so no finding above is

misread as MISSING**" [L700-705].

---

FINDINGS — per-finding verdict

HD-1 — knowledge economy declared, schema-landed, zero writers · SURVIVES, UNAMENDED

Every element re-verified independently (table above). The 8-hits-across-2-files zero is real and

the integrity_score reachability control is the correct kind. Class DESIGNED-UNBUILT is right

and the author resisted the easy MISSING. Recommendation is reuse-first and true: UWorldStateSubsystem

exists, both tags exist, composure_nodes is populated on BE_0002. Strongest finding in the audit.

HD-2 — the death payout repeats one sentence forever · SURVIVES, UNAMENDED

281/281/264 reproduce exactly. ResolveActiveFightConsolation has no attempt index — read the code

myself. FORKB L525 mint item 15 and L384 "most-repeated surface in the game" both verify verbatim.

The sequencing argument (decide shape *before* the natural-voice pass rewrites 281 strings) is the

finding's real value and it is correct.

HD-3 — "the world ADVANCES on death" has no bound · SURVIVES · CONTESTED-CARRIED

P2.14 verifies at L1780 verbatim. I independently verified the self-refutation:

harness/check_no_dead_end.py's docstring scopes it to *path reachability* — "at least one

through-line to N and onward to the Epilogue stays open" — and confirms the layer is unbuilt

("TODAY the opportunity/relationship/persona layer is unbuilt… PASSes as a SCAFFOLD"). Demand

monotonicity is genuinely a different property. The author's (a) rebuttal holds on the source.

Contested only on the re-litigation question the draft itself raises — I agree with the draft that

bounding the *encounter's* demand leaves the *world*-advance doctrine intact. Director call.

HD-4 — the ruled difficulty ladder has no defeat model · SURVIVES, UNAMENDED

All eight term counts reproduce exactly on a 111-line file, soft catch = 2 repo-wide, Hades

control matches. The shapes-vs-magnitudes distinction is sound and the Accessibility precedent

(a non-numeric shape statement already living in that doc) is a genuine reuse argument.

HD-5 — the death payout's only surface is the DEBUG HUD · SURVIVES, AMENDED

Code verifies exactly; T1_UI_UX_Spec is a 35-byte stub (I stat'd it), confirming FORKB L534.

AMENDMENT (register collision, in the finding's favour but it must not open a second lane):

v1.3 entry 50 + adjudication 3 already route content into that same doc — "the gameplay camera

clause set plus the onboarding and HUD-legibility findings land in **T1_UI_UX_Spec's fill at rank

20… (spec-doc; rank 20, one pass not five)" [L524-531]. The defeat surface must join that

rank-20 fill**, not be proposed as "the first authored section" independently. Same doc, same pass,

one owner. Also cite register entry 22 (the four byte-empty T1 stubs) as the parent.

HD-6 — retry anchor is nearest-by-distance · SURVIVES, AMENDED (premise improves)

Code, the 1-hit Checkpoint zero and the 60-hit control all reproduce. **AMENDMENT — the "add one

clause" framing is wrong because the clause is ALREADY CANON:** T99_Translation_Combat §2.5 L161

rules "consciousness withdraws to the last attuned recharge source (recharge-as-checkpoint)."

The draft cites §2.5 for HD-9 and misses that it already answers HD-6. This strengthens the

finding — reframe from "add a clause" to "the shipped code contradicts ruled canon, and the

FORKB tooth is the acceptance criterion." Second amendment: v1.3 entry 46 finds

T0_Vril_Site_Registry.mastery_tier_required "corrupted on 30 of 42 rows by a quoted-list column

shift" [L505-508] — the proposed anchor registry needs that repair first; sequence behind entry 46.

HD-7 — the ruled RESCUER can leave, no successor named · SURVIVES, AMENDED (narrowed)

J-3 (L478-480), J-6 (L490) and D-23 all verify verbatim; D-23 is genuinely open ("resolution

FLAGGED"). AMENDMENT: FORKB L518 shows the enum is `outcome ∈ {checkpoint_return, ally_revive,

ally_revive_unavailable}` — the *unavailable* state is already enumerated. The gap is therefore

the ordering of the fallback chain, not the existence of the failure state. Narrower, and still

real. Cross-ref v1.1 entry 40 (killed-companion branch) as the adjacent arm.

HD-8 — QA loop cannot see a defeat as an ECONOMY event · SURVIVES, UNAMENDED

The Record* family is exactly six (I enumerated them); QS-13..18 verifies at L362; the

"enters as a diagnostic" rule verifies at L114-115; FORKB A7's zero-coverage line verifies at L478.

The boundary declaration is pre-blessed by adjudication 4 (above). The APPEND_LEGAL_IDS

landmine warning is a real reuse hazard correctly carried.

HD-9 — graduated persistence beats the reset · SURVIVES (record-only)

Both quotes verbatim (COMBAT_ENCOUNTER_SYSTEM L221, HumanityCharacter.cpp L349-352,

T99 §2.5). Correctly filed DIVERGENT-BY-DESIGN (favourable). Keeping a "we are ahead" row in an

audit is good practice and matches the register's own posture.

HD-10 — twenty rift dungeons, one Hollow row, no dungeon table · SURVIVES, AMENDED

Twelfth-sitting (L19-42) and eighth-sitting (L233-235) rulings verbatim; T0_Hollow_Codex = 1 row /

14 columns confirmed; the four genuinely-new items verbatim at L137-141; derive-before-authoring at

L151-152. The "do NOT mint a Dungeon registry" instinct is correct and reuse-first.

AMENDMENT (register collision): v1.3 entries 47 and 48 already cover the same twenty

dungeons from two other angles — 47 (no escalation discipline inherited) and 48 (no factory rider;

"the item schedule, the dungeon count and the capacity budget are three artifacts and gate 30 cannot

see any of them"). Neither covers the container/rows question, which is HD-10's real and

surviving claim. Narrow it to that and land it as a sibling of 47/48, not a fresh headline.

HD-11 — no bail-out / re-entry / reroll contract on repeatable content · SURVIVES, STRENGTHENED

The integration-record citations verify exactly (L138 "a new content shape for the solo arena";

"farm the cap" at L403 and L572). Gate 27 check_rarity_firewall.py exists in harness/ — the

reuse target is real. STRENGTHENING EVIDENCE the draft missed: v1.3 entry 53 registers the

*identical defect class* for a different content class — the Agility Course System "defines three

variants, rewards and protections with no failure, reset or re-entry rule of any kind" [L539-541].

Two ruled repeatable-content classes with the same hole is a pattern, not an isolated gap. Cite it.

HD-12 — repeatable content has no difficulty authority · SURVIVES · CONTESTED-CARRIED

T99 §2.2 L126 and COMBAT_ENCOUNTER_SYSTEM §9 L236 verify verbatim, as does the leveling ruling

(DECISIONS L28-34). The compose-the-two argument is sound. The draft correctly refuses to propose

level-scaling. Contested as the draft flags: does "keys off REALM AND SITE" permit a rift to declare

its own band? I read yes (a rift *is* a site), but it is a ruled line and the director should confirm.

HD-13 — nine realm rows own seven env-identity columns, all empty · SURVIVES, AMENDED

Reproduced exactly with my own script: 76 rows; node_kind base 50 / revisit 9 / mythic_realm 9 /

atlantic_waypoint 6 / story_anchor 1 / bookend 1; the nine realm ids incl. hades; **0/10 on all

seven columns**, 4/50 on base; chapters 10/10 and time_period 10/10 as the reader-control;

fairy_realm env_profile_status = stub. Register grep for all seven column names = 0 hits, so

this measurement is genuinely new. §5.4 Wave 3 verifies at L1272.

AMENDMENT: REALM_DESIGN_PROGRAM L1496-1497 already schedules this — "T0_Region_Index —

the ten realm rows filled past 9-11 of 25 fields." The population is not unnoticed. What is genuinely

additive is (a) the measured 0/10 per named column and (b) the divergence tooth (a mythic_realm

row must differ from every base row on ≥N of the seven). Land the tooth; present the fill as

"already scheduled at L1496, here is the column list and the baseline."

HD-14 — realm-ruleset schema minted with no link to the realm rows · KILLED

Two independent failures.

(1) The load-bearing premise is FALSE. The draft states "the ten live realm rows in

T0_Region_Index are simply not in its field of view." I opened the single region_index hit

the draft counted but did not read. It is L1496-1497, in the charter's own downstream-edits list:

"T0_Region_Index — the ten realm rows filled past 9-11 of 25 fields; three retired-coinage keys

resolved." The charter has the realm rows in view and already commits to filling them. The term

count was right; the inference from it was wrong. *(This is the draft's one search-hygiene failure:

a count of 1 was treated as absence without opening the hit.)*

(2) Already-covered register entry. v1.3 **entry 73 (ranked HIGHEST, "the first call in the

Wave-2 mint")** is this exact concern in its more general and more urgent form: "the machine layer

cannot tell a realm from a place."

PRESERVE THE RECOMMENDATION, DROP THE FINDING. The region_id soft-FK plus the ownership split

(T0_Region_Index owns the renderable env profile; T0_Realm_Ruleset_Registry owns the law delta)

is a good concrete answer to entry 73's open question. Route it there as an input to the Wave-2

first call. It must not enter as a new row.

HD-15 — realm identity is physics, never encounter grammar · SURVIVES, AMENDED ×2 · class CONTESTED

The substance reproduces: over the field-family block, `enemy / creature / encounter / boss / hazard /

fauna / loot / drop / bestiary / opposition are all 0, with reward` = 1 as the matching control.

Register grep confirms no entry anywhere covers a per-realm bestiary or reward mix (entry 31 is

per-enemy *element* data; entry 4 is silhouette_class) — so the core claim is genuinely new. The

eighth-sitting Hollow ruling (realms have a declared native opposition class, L229-232) verifies and

is the right supporting argument.

AMENDMENT 1 — the reuse target is a category error. The draft recommends "**enrol the nine realm

rows in the existing mode_variety_check**." mode_variety_check is not a registry — it is a

column on T0_Chapter_Index, keyed on chapter_id, and I measured it at 79/79 populated. Ch 66-74

are therefore *already* in it; realm rows in T0_Region_Index cannot be enrolled in a chapter column.

The correct reuse home is the **check_realm_ruleset.py T-RLM-\* teeth family the charter already

proposes at §2.7** (T-RLM-BASELINE / PARTITION / PROVENANCE / EXPRESSION / TRIGGER / FLOOR…), which is

where a per-realm divergence tooth belongs. *(Side note: CLAUDE.md's "mode_variety_check populated

0/79" is stale — the register's own corrections section flags this at L1035.)*

AMENDMENT 2 — the hazard limb is already assigned. The charter at L1498-1499 says

"T0_Environment_Grammar_Registry — each realm's altered law lands as EG_ rows, **which is what stops

nine gorgeous realms from having ordinary combat in front of them**." The hazard_class column is

therefore already covered in intent. What survives un-covered is **native_hostile_families and

reward_mix_class** — the roster and the reward mix. Narrow to those two; they are still the sharpest

lesson in the title set.

CLASS CONTESTED: entry 77 pre-registers the whole realm machine layer as DESIGNED-UNBUILT

"so no finding above is misread as MISSING." I hold that a *field family absent from the design* is

MISSING while the *registry* is DESIGNED-UNBUILT, but the director should confirm the split.

Evidence-hygiene note: the draft's second control (roster = 1) reproduced as 0 in my window;

its block bounds differ slightly from mine. reward = 1 still carries the control. Immaterial to the

verdict, worth correcting in the record.

HD-16 — the compounding budget beats the four Hades biomes · SURVIVES (record-only)

§2.6 L640-648 verbatim, including "Unlimited novelty per realm produces nine tech demos; one

signature law plus reprises produces a road." T-RLM-COMPOUND is named in §2.7 as claimed.

HD-17 — the nine-band bond ladder grants NOTHING · SURVIVES, STRENGTHENED

RULED 2's nine bands, THE BOND's Family terminus, and the Ch-65 calibration target all verify

verbatim at L1820-1830. T0_Character_Index = 164 rows with exactly six loyalty columns, **all

join/leave/risk, none a yield** — confirmed by reading the full column list. The positive control is

the strongest kind and it verifies: T0_Familiar_Bond_Ability (5 rows) really does carry

bond_level / bond_threshold / grants_class / command_permission / integrity_band_gate —

familiars have a payoff table and companions do not. T0_Bonus_Option_Registry really carries

unlock_gate_ref / condition_expr / expected_pick_band_ref / build_defining, so the reuse

target is real. The "dead content silently" warning is in the ruling itself, as quoted.

Adjacents to cite (not collisions): v1.1 entries 37 (loyalty data ~6 rows of 163), 39

(no computational surface for loyalty_belief_axis), 43 (BOND tops at Family, no partner track).

None of the three is "a band yields nothing," so HD-17 stands on its own. **v1.3 contains zero

companion/bond/gift content** (I grepped) — this territory is unclaimed by the concurrent waves.

HD-18 — no player-initiated way to INVEST in a relationship · SURVIVES · CONTESTED · AMENDED

RULED 3 verifies verbatim, and the draft's reading is correct on the text: "never from observed

deeds" is written about the departure belief-distance scalar. Festival rows verify (16 rows, the

offering-economy gameplay_hooks); the amulet-as-gift beat verifies at ROADMAP_POST_5090 L579-581,

running player→NPC as the draft says.

MANDATORY AMENDMENT the draft missed. The same RULED 3 paragraph contains a ratified invariant

that directly bounds this ask: "opposition ceilings at 100 and the trigger caps at 90, so complete

opposition always forces departure and investment never buys immunity." Any give/offering verb

must be declared incapable of purchasing departure immunity. That is a bounding constraint, not a

fork — exactly the register's "Ruling: No (bounded)" shape. Without it this finding would quietly

reopen a ratified invariant.

CONTESTED as the draft honestly flags. I add: if Josh's intent extends to the positive direction,

this becomes DIVERGENT-BY-DESIGN and belongs in the divergence list.

HD-19 — our keepsake exists, is load-bearing, has no mechanical form · SURVIVES, UNAMENDED

I dumped all populated cells of the single T0_Equipment_Registry row: EQ_0001, `string_status =

placeholder, equip_slot = unassigned_stat_slot, wearable, signature, fairy_realm`,

CH_PROLOGUE, with the provenance chain in notes and J-6 in source_ref. Exact. T0_Imprint_Affix

exists as a registry, so routing effects there instead of adding effect columns is genuinely

reuse-first. The "ours runs outward" reframe is a legitimate design observation, not flattery.

HD-20 — the between-run hub has no return-cadence contract · KILLED

Three independent failures.

(1) The load-bearing premise is uncited and contradicted. HD-20 asserts "the player will visit

these sites more than any other class of place in the game" and "a place the player re-enters

dozens of times." This is the one claim in the entire draft with no citation, and the ruled

world structure contradicts it: register #9's own constraint cites **WGM §4/§6.7 walked-geography —

"Ch 1-58 a continuous pilgrimage (fast-travel return-only pre-Vimana)."** On a one-pass forward-walked

road, 42 sites spread across 79 nodes are passed, not re-entered dozens of times.

(2) Already-covered register entry. #9 "Rest / safe-site session-anchor loop" is the same gap

("the reward-for-returning 'zone heartbeat' is absent") with the same reuse home and an explicit

recommendation already covering "companion talk" at the safe site. It is ranked **"Lowest

priority."** HD-20 would enter as a MEDIUM duplicate.

(3) Its distinctive limb is a PRE-REGISTERED DIVERGENCE. The visit-count-keyed refresh *is* the

"zone heartbeat between excursions," which #9 declares "a deliberate divergence… do not

import it." Independently, v1.3 F-R10 finds "the **Hades-the-game warmth mechanism is

structurally unavailable on a one-pass road**" [L907-909].

PRESERVE ONE CRUMB: the T0_Scene_Spec_Registry soft-FK is a mechanism #9 never named. Attach it

as an implementation note on #9 at #9's priority. Do not land HD-20.

---

DEVIL'S-ADVOCATE-AGAINST-TIMIDITY PASS

Run over all 20 findings. No care-inflation or defanging defect found. Specifically:

• Nothing proposes removal, softening, a content warning, a protective exclusion, or routing an

antagonist away from a target. No monster is defanged. No culture is protected from being playable.

• HD-3 is the one finding that could read as making the game easier, and it pre-empts the charge

correctly: the ask is "make the cost AUTHORED instead of accumulating," and it explicitly holds

"Sundered and The Forgotten One stay as hard as ruled." HD-4 reinforces the same floor —

soft catch on Wellspring/Veiled only, "knowledge only — the door stays hard" on the top two tiers.

That is the R-17+ calibration, not a retreat from it. Both pass.

• HD-17 and HD-18 both run the pass in the correct direction — HD-17 argues a nine-band ladder

yielding nothing is "under-designed… not tastefully restrained"; HD-18 explicitly names the

gift-mechanic objection "the timid read and I am ruling against it," and grounds reciprocity in

authored festival rows rather than importing a menu of hearts. Correct posture on both.

• One over-boldness check: HD-18 is the only finding adding a player-facing verb, and it is the

one I have bounded (the "investment never buys immunity" invariant) and left contested. Appropriate.

The residual risk is the opposite of timidity and worth stating: this lens's Part A pushes toward

a *generous* defeat economy. The floor that keeps it honest is HD-4's tier split, and it must land

with HD-3 rather than after it, or a bounded-demand invariant could be read arc-wide instead of

tier-scoped.

---

VERDICT

Killed: 2 · Survived: 18 · Amended: 8 · Contested: 5

*(Killed: HD-14, HD-20 — both for already-covered register entries plus, in each case, a refuted

load-bearing premise. Amended: HD-5, HD-6, HD-7, HD-10, HD-11, HD-13, HD-15, HD-18. Contested-carried

to the director: HD-3, HD-10 scope, HD-12, HD-15 class, HD-18.)*

LANDING RECOMMENDATION — LAND WITH AMENDMENTS

This draft clears the bar comfortably. I opened every citation and re-ran every positive

control with my own tooling; the counts reproduced to the digit in more than twenty independent

checks, including six code sites, eight registries and three term-count sweeps. **No fabricated

evidence anywhere.** The class discipline is honest — DESIGNED-UNBUILT is used seven times where a

weaker auditor would have minted MISSING, and two findings are filed as favourable divergences.

The systematic weakness is one thing only, and it is a process gap rather than a competence gap:

the author did not re-read DESIGN_GAP_REGISTER.md against his own findings, and the register grew

573 lines mid-wave. Both kills and five of eight amendments trace to that. The single recurring

technique error is treating a term count of 1 as absence without opening the hit (HD-14) — the

mirror image of the repo's own "a search that cannot match reports zero" lesson, and worth naming in

the lens program's standing rules as its sibling: *a search that DID match must be read.*

Land in this order: HD-1 → HD-2 → HD-4 (the Part-A spine; HD-4 must land with HD-3, not after) →

HD-6 (reframed as code-contradicts-canon, sequenced behind entry 46) → HD-5 (into the rank-20

T1_UI_UX_Spec fill, not a separate pass) → HD-8 → HD-17/HD-19 → the rest.

Route, do not land: HD-14's ownership split → entry 73's Wave-2 first call. HD-20's Scene_Spec FK

→ an implementation note on entry #9.

Director calls needed before landing: the five contested items, of which **HD-18 is the only one

that touches a ratified invariant** and is therefore the only genuine fork.

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