COMBAT_ENCOUNTER_SYSTEM.md

music/COMBAT_ENCOUNTER_SYSTEM.md

<!-- Combat / Boss / Encounter System — assembled from the 5-track design workflow (wkvtx1pxa,

2026-07-04): canon audit + boss-system design + encounter ecosystem + local-LLM runtime layer + milestone

pipeline. T99-operational tier (sits under T1_Combat_System_Spec). Invents zero world canon; every claim

traces to ACTIVE canon / a RULED proposal / a batch-1 PROPOSAL. Absolute numbers stay deferred to Phase 5M.

The local-LLM + milestone parts also have standalone drafts: docs/RUNTIME_GENERATIVE_LAYER.md +

docs/MILESTONE_BUILD_PIPELINE.md. -->

Humanity — Combat / Boss / Encounter System

Overview (workflow synthesis):

Delivered. C:\dev\humanity-forgotten\docs\COMBAT_ENCOUNTER_SYSTEM.md — the unified, canon-cited, buildable spec merging all five streams.

What it is: a single T99-operational-tier index-and-systems doc that sits under T1_Combat_System_Spec [ACTIVE v1.0], invents zero world canon, and traces every claim to ACTIVE canon / a RULED proposal / a batch-1 PROPOSAL. Five parts + two appendices:

• Part A — Encounter Ecosystem. Two orthogonal axes (structural vs source); the 9 encounter classes with counts (75 main, ~150-300 sub, 22 threads-as-lens, 120 trade gates, 12 SQ buckets, 72 weapon, 13 school, 33 species/22 familiar, 6-9 environmental); per-chapter layering; the 4 anti-staleness rules; ~350-500 designed-boss total; and a table of which pipeline pass builds each layer.
• Part B — Boss/Combat System. The vril-density thesis; the core loop (prior-art improvement ledger); Pillar-8-as-schema; the evolve-not-swap phase state machine; the ~30-clock enumeration with [DIAL] markers; Composure (the mode-agnostic posture keystone); environmental-boss protocol; the win-AND-death economy; RNG/Sigma balance rule; the full vril-difficulty dial; a worked example; the 7-point novelty ledger.
• Part C — Local-LLM Runtime Layer. The deterministic/generative split, the 4-stage guardrail sandwich, AAA constraints, the continuo F1-F10 failure discipline. (indexes docs/RUNTIME_GENERATIVE_LAYER.md)
• Part D — 5-Chapter Milestone Pipeline. 9 engines in dependency order, Milestone 1 = Ch 2-6, the 8-stage loop + exit gate, cost-staggering. (indexes docs/MILESTONE_BUILD_PIPELINE.md)
• Part E — Build Sequence. E-0 ratify → E-1 give the roster a schema (the single largest structural gap 3.2) → E-2 reconcile taxonomies → E-3 populate Ch 2-6 → E-4 wire runtime packs → E-5 loop to exit gate. Each step names its harness/recall touchpoint and the --emit-baseline discipline so nothing turns the gates red.

Key synthesis contributions (labeled [EXTENSION], each anchored): resolved the audit's Part-3 gaps into bindable proposed shapes — the win-condition enum (3.3), character-engagement = a standalone 8th Composure node decoupled from astral (3.4, RULED D-COMBAT-TD R4), Patterns/Puzzles/Preparation axis definitions (3.5), the phase-trigger enum + tighten/multiply (3.6), the 7↔10↔5 mode-taxonomy reconciliation (3.7), sub-boss differentiation rules that bound the authoring load (3.11), and the knowledge-economy death reward (3.12). A closing traceability table maps all 12 gaps to their section.

Everything Josh-authored (chapters, verbs, names, cardinal numbers) is left to him; all absolute numbers stay deferred to Phase 5M; open contradictions carry their ledger IDs (D-3/D-4/D-5/D-9, anchor #2). The doc correctly sits in the operational tier alongside PIPELINE_ADAPTER/ROADMAP and does not enroll in the authored-canon lint gate.

---

PART A — Encounter Ecosystem

Humanity — Encounter / Boss Ecosystem Taxonomy

Full markdown saved at C:\Users\joshu\AppData\Local\Temp\claude\C--dev-humanity-forgotten\f75a84da-bb17-40db-a02b-876622ee3842\scratchpad\encounter_boss_taxonomy.md. Delivered inline below.

0. Two orthogonal axes

Every encounter sits on two axes at once, and conflating them is the common error:

• Structural axis (where in the chapter): trash / mini-boss / sub-boss / main chapter boss / act-anchor.
• Source axis (what canon system spawns it): Velheim antagonist / thread / trade-gate / side-quest / legendary-weapon / school / fauna / environmental-vril-state / MMO world boss.

Ruling 7's "everything woven" means a chapter draws its 2-4 sub-bosses from different source systems and different threads — never all from the House-of-Velheim op (CHAPTER_BUILD_TARGET.md §1). Classes §1–§9 are points on the *source* axis; §10 shows how they stack onto the *structural* axis.

1. Main chapter boss — 75

One per chapter Ch 2–77, minus the fairy-realm bookends. No rows for Prologue, Ch 1, Epilogue (L1 = traversal/cutscene); Ch 76-77 = one reveal fight. [JOSH-AUTHORED] count + the 5 Tier-A anchors Ch 8/23/38/55/65 (T1_Combat §7 L209-215); the other 70 are the Claude-created, adversarially-reviewed proposal in boss_roster_rows.json (Ruling 2), names still [DECISION NEEDED - JOSH]. Every row obeys Pillar-8 doctrine: blended/never pure-DPS, primary+secondary verb, three-axis Patterns/Puzzles/Preparation (never 100/0/0), moveset-evolving phase escalation, ≥2 counters/signature, Hades reward on win AND death. Structural template = Ch 17 Entu Priestess three-phase. Non-standard mains (Ch 69 argument, Ch 42 Debt Engine) are real playable fights, not cutscenes.

2. Per-chapter sub / mini-bosses — ~150–300 (mid ~225)

2-4 per chapter → one main boss (CHAPTER_BUILD_TARGET.md §1). [DERIVABLE] from 2-4×75; per-chapter count [JOSH-AUTHORED via cascade]. Anti-monoculture rule: sub-bosses surface across different threads and different source systems (exemplars Ch 3 / 42 / 43 each spread Naga-fauna + oppressor-combat + folklore-boss → main). This is where trade-gate, side-quest, thread, and fauna bosses physically appear inside a chapter, each carrying one of the 12 woven facets.

3. Thread bosses — 22 threads, as antagonist expressions (not a separate roster)

[DERIVABLE — no new count.] Thread bosses are a *lens* on §1–§2, not a ninth roster; "sub-bosses across different threads" is exactly how the 22 threads get playable expression. Thread 3 (Antagonist Family) alone anchors distinct forms at Ch 2,3,4,6,8,14,16,17,18,22,23…76 (T1_Threads_22 §Thread-3). The one place a thread IS the boss: Cassius Velheim (Ch 76) + Father/Elven King (Ch 75). Thread 1 resolves via *Witnessing*, not a fight; Thread 6 has no antagonist. Absorbed into the 75 + ~225 totals.

4. Trade-questline gate bosses — 10 trades × 12-tier = up to 120 gates

[MIXED.] "10 trades / 12-tier / every tier quest-gated" is hard canon (CVD §12.1, §14.2); the "120" phrase is NOT in CVD — it's CLAUDE.md's 12×10 arithmetic, flagged not invented. Not every gate is a boss: Outer-Order T1–T5 gates = playable TESTS (mini-bosses), T5/T6 threshold ceremonies = NARRATED, Inner-Order T6–T8 = full tests, T9/T10 = prestige space. Boss-class fraction is [JOSH-AUTHORED per trade via cascade]. Ten trades each have 1 ancient + 1 modern home base.

5. Side-quest bosses — 12 buckets, many bosses each

[DERIVABLE bucket count = 12 canon; per-bucket JOSH-AUTHORED.] SQ1 Folklore…SQ12 Secret Societies (CVD §11.1), each 1M+ words → many bosses across the arc. SQ9 Cryptids and SQ1 Folklore overlap the fauna/folklore classes rather than being disjoint. Appear as chapter sub-bosses and free-standing open-world encounters.

6. Legendary-weapon bosses — 72

36 Brute + 36 Trickery, each with a Legendary quest whose climax is a boss-class encounter granting the Legendary form (T1_Combat §5/§6). [DERIVABLE count = 72 canon; identity JOSH-AUTHORED at Phase 6.] 68 documented + 4 to-add slots (Ch 9-10, 47-48, 29 Sámi, 57 Australia). Completionist/Layer-3 content — questline capstones in the weapon's home region, not on the main ladder.

7. Ability / school bosses — 13

13 CVD Pillar-4 schools, each a teach→grant→showcase questline whose climax boss GRANTS the verb (school_roster_rows.json). [roster count = 13 canon; the 13-row proposal is Claude-created, Ruling 2.] The 185 abilities are canon; schools teach *verbs*, not counts. Maatherion (Ch 69) is where the school and the boss are the same object. Showcase engineered into the NEXT chapter (metroidvania retroactive-unlock). Guardrails: no school teaches Poison; no gate grants the 8 pure numeric buffs; the four threshold schools form a strict relay.

8. Fauna / creature encounters — tameable + hostile, folklore substrate

[JOSH-AUTHORED cardinal numbers — verbatim canon (CVD §17.12):] 33 legendary species = 21 mortal (7 Ground/7 Flight/7 Water) + 12 elven; 22 familiar slots = 21 base + 1 path-variable (childhood familiar, reunited Ch 58 Antarctica); binding-stone mechanic. Non-legendary hostile/tameable fauna = open-ended per region (T0_Creature_Roster, schema live). Three modes: person-to-animal combat (hostile sub-bosses), bonding-quest encounters (21+1 legendary + Trade 1/8), folklore-as-real fauna (CVD's L1 promise "mythological creature as real fauna"; Creature_Roster carries folklore_behavior_substrate + attribution columns). Ch 57 fauna combat IS permitted (Ruling 5). Per-region rosters [JOSH-AUTHORED via 69 region pages]; substrate [deep-research at build time].

9. Environmental / atmospheric bosses — the protected/relief playable layer

6 canon-documented instances (Ch 10, 20, 29, 38, 45-46, 57 — T1_Combat §9.2) + a Claude-proposed 9-landmark ledger (adds Ch 51, 52, 70, staircase beats 36/37). Cannot be defeated by damage — canon 4-step protocol: Perception → Vril-management → Restraint → Purgation; baseline Intuition+Constitution+Integrity; graduated failure (partial purge strengthens the Ch 39+ revisit = Hades loop). This is the ruled answer to "no combat" chapters (Ruling 5/7): L1 stays action-packed everywhere; defending protected people from oppressors is a first-class combat layer alongside the protocol. (9b) Tier-C MMO world bosses (100-player, 4 teams) are categorically separate and out of the Solo-Base count (T1_Tier_C_MMO_Spec).

9a. HL_0004 — what "never left combat-absent" is measured on (RULED, WQ_0045, 2026-08-04)

HL_0004 is universal and canonical: "A culturally protected space still delivers a real fight. Combat there comes from antagonist intrusion AND/OR the tradition's own monsters, honored-adversary champions, or environmental/astral boss encounters — a protected site never needs a manufactured Velheim raid to be playable, and is never left combat-absent." The id's TEXT was never in doubt; its GRAIN was, and 48 registry rows were re-deriving the answer one row at a time (29 declaring an omission, 19 a satisfaction). This section is the answer, so a row cites it instead of restating it. It interprets HL_0004 and changes nothing in it; the hard-line row itself is Josh's surface and is untouched.

The subject is the culturally protected PLACE — not the Zone Catalog boolean, and not the chapter.

• Not the boolean. protected is overloaded, and the pages say so in their own words: central_africa_congo.md's zone-grain note reads "five of the eight zones are protected living-tradition or protected-reach ground". A trinational conservation reach is protected-as-ecology; HL_0004's own word is CULTURALLY. Reading the id off the boolean also inverts the authority hierarchy, because a page-authored region-ambient block — declared in the same catalog as "page authority, not spine canon" — could then put a CVD-derived universal hard line into violation.
• Not the chapter. Ch 11 stages five encounters and four of them sit on unprotected colonial ground, so a chapter-grain reading would let a chapter satisfy a protected-space rule with fights nowhere near a protected space. That drains the id of the thing it exists to force, and the objection stands.
• Not a living community's homes either, and this half was earned rather than assumed. HL_0004's own nouns are "space" and "site", and CVD §17.1's protected-space clause is about sacred, ceremonial and heritage places — temples, rock art, ceremony grounds. A residential camp is covered by HL_0003, which is collective protection of the community. Ch 10 states the same thing in canon's own words: juhoansi_camp is typed "protected community, never a battleground", and the v3 conversion deliberately MOVED the crisis-dance off that ground to san_rock_art_panels, where it is fought. The check that this is a routing and not a convenient carve-out is that it leaves the id's teeth in: Ch 10's protected SACRED site does host its fight, and Ch 11's real debt does not disappear — it narrows from three zones to one, lope_bwiti_ceremony_ground, which is a ceremony ground staging nothing, exactly HL_0004's subject.
• So: a zone is under HL_0004 when it is protected AND it is a cultural SITE — sacred, ceremonial, or heritage ground. Several catalogue class-rows may decompose one site; the site is what the id measures.

A boss PHASE staged on that ground discharges the id, and traverse is a real exercise of it. Ruling a traverse-primary phase "not a real fight" would import a strike-chauvinism Pillar 8 explicitly rejects — no axis is ever 100-0-0, BE_0011's own primary verb is traverse, and the v3 conversion re-grounded BE_0011_S3 from strike to traverse for exactly this reason. The discharge is not free, though: the row must NAME the phase, the boss verb exercised on that ground, and what the ride does NOT claim. EG_0154 is the worked exemplar — it names BE_0011.P1, the primary verb traverse, the boss's own composure node, and then names the two counters that are exercised at the station in P2 and not here. A corridor with no exercised boss behaviour is not a discharge.

Four verdict values, and every declaring row carries exactly one.

• OUT_OF_SCOPE — not culturally protected ground: protected false, or protected-as-conservation. There is no debt and there never was one.
• COMMUNITY_GROUND (HL_0003 governs) — a living community's dwelling and daily ground. HL_0004's subject is a protected site, not people's homes; whether an intrusion is ever brought onto a camp is a spine care call under HL_0003, never a registry-side debt. This is a routing, not an exemption: the ground is not permanently barred from a fight, it is simply not HL_0004's to demand one.
• DISCHARGED — a named boss phase is staged on this site, with its verb and its non-claims named.
• OPEN — a culturally protected SITE that stages no encounter. This is a real obligation on the spine owner, not a settled omission, and the row names the mechanism HL_0004 itself supplies for that ground so the owner is not handed a blank.

A registry lane records the verdict; it never mints the encounter. Staging a beat is the spine owner's act. An OPEN verdict is discharged by a spine beat, never by a registry-side invention — the same law that refused a Ch-9 boss row without a slated beat (CR_0184) and refused an encounter on the Fang shelf at Ch 11 (WQ_0044).

10. Per-chapter layering

open-world trash + hostile fauna (§8)          — dozens, emergent
SUB-BOSS 1 ← thread expression (§3)            ┐
SUB-BOSS 2 ← trade-gate OR side-quest (§4/§5)  ├ 2-4 across DIFFERENT threads/systems
SUB-BOSS 3 ← fauna/folklore OR school (§7/§8)  ┘
MAIN CHAPTER BOSS (§1) ← the peak
   (= school climax at grant chapters §7; = env protocol at protected/relief §9)

The sub-boss ladder is the delivery vehicle for the 12-facet weave (combat/exploration/discovery/puzzles/mystery/lore/trades/gathering/side-quests/engineering/folklore/superstition). Three-layer surfacing always: L1 complete alone; L2 investigatable substrate; L3 metaphysical-through-mechanics, no pre-Ch-76 Grand Sage hint.

11. Anti-staleness (four executable rules)

1. No two consecutive chapters share dominant combat type (T1_Combat §9.3; CVD §Pillar-1.12); only sanctioned same-type pair = Ch 40-41. The 3 flagged same-mode main pairs (4-5, 37-38, 43-44) are fixed in cascade respec where sub-bosses supply the real per-chapter variety.

2. Every 5-chapter window ≥3 distinct combat types.

3. Ten encounter categories rotate (person-to-person, person-to-animal, group/formation, boss, social boss, environmental, stealth, naval, mounted, vril-projection).

4. Win-condition + verb rotation (HP/part-break/pacify/perceive/restraint/declare/bypass) so resolution differs even when opponent type recurs.

Enforcement point = the mode_variety_check registry, run per node by the cascade. Agility-course biome-rotation adds a parallel non-combat variety dial.

12. Rough total

Designed boss-class encounters, Solo-Base: ~350–500 = 75 main + ~225 sub + ~72 weapon + ~13 school + ~22 species + ~6-9 environmental + the boss-class fraction of the 120 trade gates and 12 SQ buckets. Thread bosses (§3) are NOT additive — already counted inside the 75 + ~225. On top of that, canon promises "dozens of designed encounters plus emergent open-world plus infinite side-quest combat" per chapter, so total encounters are effectively unbounded; ~350-500 is the *designed-boss* figure. Distribution: peaks (75 main) with 2-4 foothills each; act-anchors thicken at 8/23/38/55/65 + Realm War 71-73 + endgame 74-77; environmental bosses in protected/relief valleys (10/20/29/38/45-46/52/57/70); weapon/school/species capstones distributed as optional content.

13. Pipeline-pass mapping

Main bosses → T0_Boss_Encounter_Registry (0.E drafted → sign-off → cascade). Sub-bosses/thread tags/env bosses → same registry, Phase-6 /spine-chapter cascade. Trade gates → T0_Quest_Definition_Registry+T1_Trades_Master. Side-quest → T0_Questline_Index+T5. Weapons → T0_Weapon_Registry+Tier-6 (Phase-6 region build). Schools → T0_School_Registry (0 rows → proposal). Fauna → T0_Creature_Roster+T0_Familiar_Registry (DRAFT, schemas live; Phase-6 + deep-research). MMO → Tier-C rows (Session 5J). The 75-row main layer is already drafted; everything else populates during the Phase-6 cascade.

14. Josh-authored vs derivable ledger

• Cardinal canon (never auto-derive): 75-count + 5 Tier-A anchors; 22 threads / 10 trades / 12 tiers / 12 SQ / 72 weapons / 13 schools / 185 abilities / 69 regions; 33 species = 21(7/7/7)+12 elven; 22 slots = 21+1; binding stone; Ch 58 reunion; 6 env instances + 4-step protocol; Rulings 5-7.
• Derivable (flag, don't invent): sub-boss ~150-300 (2-4×75); 120 gates (12×10, phrase is CLAUDE.md not CVD); ~350-500 total; the 9-landmark env ledger.
• Claude-proposals pending sign-off (Ruling 2): 70 non-anchor main rows; 13-school roster; per-chapter sub-boss identities + NPC names ([DECISION NEEDED - JOSH]).
• Open flags: D-9 (Ch 58 non-anchor boss?); thread-ID encoding mismatch (T01..T22 vs THREAD_01..THREAD_22) must canonicalize before FK population; Ch 71-73 battle-myth confirmability + Ch 70-77 fairy-realm serpent = reserved Witness slots, no authored content.

Key source files (all absolute): C:\dev\humanity-forgotten\_source\01_Tier_1_Foundation\T1_Combat_System_Spec [ACTIVE v1.0].md (§7/§9), ...\T1_CVD_Creative_Vision_Document [ACTIVE v1.4].md (§11-14, §17.12), ...\T1_Threads_22_Master_References [ACTIVE v2.2].md (Thread 3), C:\dev\humanity-forgotten\docs\proposals\CHAPTER_BUILD_TARGET.md, ...\ROSTER_PROPOSALS.md, ...\boss_roster_rows.json, ...\school_roster_rows.json.

---

PART B — Boss / Combat System (phases · the ~30 clocks · RNG-vs-scripted · vril-difficulty)

Humanity — Boss & Combat System Design (real systems, canon-true, AAA-first)

Status: DESIGN PROPOSAL. Anchors only to ACTIVE canon; invents no canon (chapter/boss/verb assignments stay Josh's). This is the *systems layer* under T1_Combat_System_Spec — it specifies the interacting-variable machine that the per-boss roster rows (docs/proposals/ROSTER_PROPOSALS.md + boss_roster_rows.json) plug into. It contradicts no hard line I could find; where it extends canon it says so and cites the anchor.

Full file at: C:\Users\joshu\AppData\Local\Temp\claude\C--dev-humanity-forgotten\f75a84da-bb17-40db-a02b-876622ee3842\scratchpad\combat_system_design.md

---

0. The one-sentence thesis (what makes Humanity's combat novel)

**Difficulty is not a stat knob on the enemy — it is a *density knob on the world*, and that same knob is the game's cosmology.** Every action game scales bosses by inflating HP/damage/resistance. Humanity refuses that as the *source* of difficulty (DIFFICULTY_SYSTEM.md, amended 2026-07-30 by the twenty-eighth sitting item J5: some health and damage scaling exists, most of the difficulty comes from complexity, and difficulty never comes solely or mainly from inflated numbers — §9b). Instead the single dial is ambient vril density — literally how far the withdrawn fairy realm still reaches you (T1_Vril §1, §10). Turn the dial down and the *world gives you less*: telegraph slack shrinks, your recharge wells answer slower, and the boss's answer-tree gets pruned toward its one true solution. The boss never gets a bigger number. You have to do more because home reaches you less. That is the improvement: Souls-grade demand *without* the number-inflation lie, delivered as diegetic physics rather than a menu abstraction.

Everything below is built so a designer can turn exactly one global float (vril_density ∈ [0,1]) and have a Story-mode boss and a Sundered-mode boss be *the same authored encounter* at different world-tensions — no separate difficulty-tuned enemy tables, ever.

---

1. The core loop we're improving on (principles extracted, then bettered)

The loop, stated once: *Read the world's tell → answer with the mode-correct primary verb → the boss defends, forcing your secondary verb → land the mode-break to open a vulnerability window → spend it before the environmental cycle or Withdrawal clock closes it → the phase evolves and adds a layer → repeat at higher world-tension.* Win or die, the run pays out.

---

2. Pillar-8 compliance is the schema, not a checkbox

Every boss row already carries a primary verb / secondary verb pair and a blend axis (e.g. strike / declare, 40/35/25) in boss_roster_rows.json. The system reads those fields literally:

• Primary verb — the dominant answer the L1 player finds immediately (action-fun layer).
• Secondary verb — the answer the boss *forces* you into for at least one phase; a boss beatable with primary alone is a Pillar-8 violation (CVD §5 P8: "single-mode in audit = pillar violation"). The engine enforces ≥2 distinct valid counters and a character-engagement field or it fails the gate.
• Blend axis (three weights summing to 100) — the runtime mixer for how much Composure must be broken by combat vs. puzzle/environment vs. character-engagement. "Never single-mode" made numeric.

The seven canonical modes (direct / stealth / puzzle_routed / environmental / vril_management / familiar / astral — CVD §5 + Combat_Spec §9) are the palette. A boss's primary and secondary verbs must sit in different modes for ≥1 phase — one rule that makes every fight blended by construction, not by review.

---

3. The phase model — "evolve, not swap" as a state machine

Model each boss as a layered accretion machine: a stack of Layers; a phase transition pushes a new layer without popping the old one.

Phase 1:  [ L1 primary-verb loop ]                      ← action-fun spine (L1 player)
Phase 2:  [ L1 loop ] + [ L2 forced secondary verb ]    ← the blend arrives; old verb STILL valid
Phase 3:  [ L1 ] + [ L2 ] + [ L3 recontextualize ]      ← character/thesis pressure; both prior layers live

• Every prior layer stays load-bearing. In P3 you're still dodging P1 tells; the fight got *deeper*, not *different*. This is "evolve, not swap" — and the anti-Souls move: Souls P2 often deletes P1's moveset; we forbid deletion.
• Transitions are Composure-gated, not HP-gated where canon allows. Default trigger = a Composure break (§5), so *skill* advances the fight. Rows may declare HP sub-phases where canon wants felt escalation (Ch 76 Cassius: "HP sub-phases → pacify/purge by path"), but even those resolve on a verb condition (pacify/purge), never pure DPS — the Pillar-8 line.
• The template is canon. Ch 17 Entu Priestess three-phase (surface combat → moral descent → philosophical confrontation) is the reference (CVD §5 P8). Phase-3 confrontations resolve on a character verb (declare/weigh/pacify); the boss whose P3 *is* a school of argument (Ch 69 Maatherion, "the school IS the boss") is the purest case — the 42 Confessions *are* the phase, the win is recitation not damage (ROSTER_PROPOSALS.md; CVD §17.3).
• Endgame escalation is authored, not scaled. Ch 73 "game's largest single combat encounter" runs its canonical three sub-phases with a TGP-exclusive familiar paired-operation at Sub-Phase C (RCM L6720); Ch 76 implements Plato's Charioteer as a literal three-phase tripartite-soul boss (RCM L6752). Sub-phases are just additional pushed layers.

---

4. The "dozens of factors on timers" — the full clock enumeration

Every one is a real runtime clock/table; the density dial modulates the ones marked [DIAL], and nothing here inflates an enemy stat.

4.1 Tell timers (windup) — tell_windup_ms per attack. [DIAL]: shown windup = base × (0.6 + 0.4 × vril_density). Wellspring reads generously; Sundered = "slack cut from every telegraph" (same animation, less lead). tell_channel (visual / audio / vril-polarity shimmer) must be readable without color alone (accessibility hard req).

4.2 Cooldowns & rhythm — attack_cooldown_ms; combo_gate (anti-loop); aggression_meter (0–1) that biases the RNG move-table toward pressure the longer you turtle, decays when you press. Anti-passivity governor; shifts *selection weight*, never damage.

4.3 Vulnerability / punish windows — punish_window_ms opened by a counter or Composure break. [DIAL]: base × (0.7 + 0.3 × density). vuln_node = the anatomy/behavior exposed (part-break analog); hitting it advances a Composure sub-clock (§5).

4.4 Status / DoT clocks (bounded, mode-flavored) — bleed / stagger / chill-root / burn / poison / soak (vril-wet → conducts next Air/Lightning hit). Each is a countdown with a stack cap; capping triggers a break event (chill→shatter). boss_status_resist_regen sheds a stack every N s *unless the env cycle is in the matching state* — status couples to the arena, not to a resist stat.

4.5 Environmental cycles (the arena is a clock) — env_cycle_period_s + env_phase: every arena runs a periodic state machine grounded in real regional substrate (tide in a Swahili dhow fight, geyser/altitude pulses at Uyuni, sandstorm intervals, waterfall/obelisk vril-tide — T1_Vril §5; Combat_Spec §9.2). L2/L3 players time offense to the cycle; L1 still wins on reflex, slower. env_trap_nodes = hazards you trigger *against* the boss (often the secondary-verb answer for environmental bosses). The cycle is read-only to difficulty — the dial changes react *slack* (4.1), not cycle speed (that would be fake difficulty).

4.6 Add-wave timers — add_wave_period_s / size / table / cap (weighted, Sigma-bounded). Adds are Composure pressure, not HP padding — they force the secondary verb while you're mid-primary. [DIAL]: density changes *answer availability* (low density → the ambient-vril crowd-clear costs more, refunds less), not wave rate.

4.7 The Withdrawal clock (enrage replacement) — withdrawal_timer_s. On expiry, instead of a damage spike, ambient density drops one notch (vril_density -= step): tells tighten, wells slow, answers prune — the enrage *pressure* with zero inflation, exactly as canon says the realm is *already receding* ("depletion is an ongoing process," T1_Vril §10). Resolving a phase resets/rewinds it — competent play literally holds the realm open.

4.8 Vril-ambient modifiers (global density field) — vril_density (the dial) sets telegraph slack (4.1), punish-window width (4.3), recharge-well answer latency (T1_Vril §5), and answer-pruning depth (§9): one float, four coupled effects, plus the arena's baseline VFX weather so the tension is *visible*. integrity_band_field = the mirrored 5-band from IPW §13 (L1 +10% healing … L5 +10% dmg +10% vuln) — a player-state modifier from the morality system, orthogonal to the dial, and the one stat delta canon allows: on the *player*, chosen by their own alignment, not spawned by difficulty.

4.9 Player-tell adaptation (the boss reads YOU) — player_habit_vector: a rolling histogram of *your* tells (dodge direction, favorite opener, turtle ratio, spam). The move-table (§8) is weighted against your habits — roll-spammers baited into a delayed sweep, turtles guard-broken. Sigma-bounded so it *nudges*, never frame-perfect-punishes. adaptation_gain [DIAL]: higher difficulty → faster learning → a *smarter* opponent, not a stronger one.

4.10 Composure clocks — see §5. 4.11 Phase-transition triggers — see §3. 4.12 Vril-Pulse window — pulse_window_ms, a tight refund input after a spend; [DIAL] narrows/shrinks as density falls. Rewards rhythm, punishes greed via the economy, never via damage.

Count check: ~30 interacting variables (windup, channel, cooldown, combo gate, aggression, punish window, vuln nodes, 6 status clocks, resist regen, env period/phase/traps, add period/size/table/cap, withdrawal clock, density, recharge latency, pruning depth, integrity field, habit vector, adaptation gain, composure + sub-nodes, phase triggers, pulse window) — ~half dial-modulated, none an enemy HP/damage number.

---

5. Composure — the mode-agnostic "posture" bar (the keystone)

Souls-likes give you poise/posture that only combat fills. Humanity's second bar is Composure, and its novelty is that every mode can deplete it — that is how "blended, never single-mode" becomes a *felt mechanic*.

Composure is a boss-side meter with named sub-nodes (part-break analog), each depletable by a specific mode's verb: direct node on parry/heavy-punish; stealth node by break-LOS-and-reengage (backstab-grade); puzzle_routed node by solving the arena interlock (align obelisks / complete inscription); environmental node by triggering a trap on-cycle; vril_management node by matching the boss's polarity and starving its manifestation (§6.2); familiar node via bonded paired-operation (canon at Ch 73 Sub-Phase C, TGP-exclusive); astral node only in the astral layer (Hard-Line chapters, Pillar 11).

• Full break → phase transition / massive punish window. Because nodes are spread across modes, a single-verb player physically cannot break Composure before the Withdrawal clock bites — the system *pulls* you into the secondary verb. Pillar-8 enforced by pacing math, not a reviewer.
• [DIAL]: Composure regen latency scales with density — at Sundered, nodes you don't finish decisively re-arm faster. No node's *value* changes; only the pressure to commit.

Biggest single novelty: posture you break with puzzles, stealth, environment, familiars, and vril — not just your sword.

The routing this section states in prose has a data home: registries/T0_Composure_Routing [DRAFT v0.1]/Sheet1.csv, one row per live ability_class value in T0_Ability_Tree_Registry, carrying that class's primary and secondary Composure node, whether it breaks a node or only enables the break, its engine mode tag, and which counter-class demands a member of the class can answer. Every row cites the sentence above (or §3, §4.1, §4.3, §4.5, §6.1, §6.2, Appendix §1.2, or T99_Translation_Combat §5.4 for the standalone eighth character-engagement node) in its own source_ref column, so the mapping is auditable row by row rather than reconstructed from a hand-coded switch. The table is registered in docs/registry_extensions.json (_deferred.composure-routing) and docs/fk_spec.json. It is INTERNAL routing only — mode and node tokens sit alongside damage types as engine plumbing and never reach a player surface, per the no-mode-taxonomy rule; players meet abilities by name and school. Two facts the table records rather than smooths over: no ability class satisfies a parry_only demand today, because parry is a weapon primitive gated to parry-capable weapons rather than an ability class; and 48 of the 185 ability rows carry an empty ability_class (the whole Layer-B hybrid-element band), so the routing covers 137 of 185 rows until that column is populated.

---

6. Arenas, environmental bosses, "room modifiers"

6.1 Arena as a Hades modifier space — each arena has a vril-site layout: recharge wells (T1_Vril §5), trap nodes, stealth cover, high ground for astral/familiar, the env-cycle emitter. The *same boss* in a different chapter's arena plays differently because the modifier space differs — our aspect/room-modifier variety, doubling as Pillar 10 regional combat identity (jungle/desert/steppe/ice movement, stance systems, weapon-of-origin bonuses).

6.2 Environmental / atmospheric bosses (canon's non-DPS win) — canon defines a boss class whose victory is *not* damage (Combat_Spec §9.2; CVD §12.3), 4-step protocol Perception → Vril-management → Restraint → Purgation. In systems terms: the "boss" is a vril-state field with hidden genuine_echo + negative_manifestation; the player runs a polarity-matching minigame (read the field, gated on Intuition/Constitution/Integrity → tune own vril to resonate with the echo → withhold combat-scale expenditure: a restraint_budget clock that, if you dump burst, destroys both → fail = environmental degradation + faction loss, canon). [DIAL]: low density → tighter restraint budget, fainter echo. Still L1-playable per Josh's ruling (CANON_AUDIT.md, 2026-07-03): L1 must be action-packed in every chapter; the protocol *is* the playable layer for protected/relief chapters, and WGM "no combat" loses to the Pillar-3 hard line — so even a no-antagonist chapter (Ch 57) delivers a real kinetic encounter, not a menu.

---

7. The death/win economy (Hades, formalized and canon-mandated)

Canon mandates reward on win AND death (CVD §5 P8).

• On win: loot (Sigma-bounded), Composure-node mastery credit (permanent read on that node's tell), integrity-routed drop table (IPW-banded), 12-tier mastery XP, and for legendaries a step on the weapon Legendary-quest.
• On death: a Withdrawal-echo payout — *the realm remembers your effort.* You bank the tell-reads earned this attempt (newly-seen boss tells stay revealed next attempt — a knowledge economy, not an XP handout), a substrate-lore fragment (L2/L3 Grimoire, so death *teaches*), and on Wellspring/Veiled a soft catch ("no trial can strand you"). On Sundered/Forgotten One the payout is knowledge only — the door stays hard.
• Strictly better than Hades' currency drip: the death-reward is **diegetic knowledge that lowers the *next* attempt's skill floor without lowering the boss's demand** — you improve because you *learned the tells*: the Souls fantasy paid out on the Hades cadence.

---

8. RNG vs. scripted — the balance rule (canon: Sigma Constraint)

Rule: authored structure is deterministic; texture is bounded-random; nothing is unbounded.

• Deterministic (never rolled): phase order, transition triggers, hard-line beats, env-cycle schedule, the *existence* of each Composure node, integrity routing — fire on state, not dice (RNG_Sigma §3).
• Bounded-random (Sigma-constrained): move selection, add-wave composition, loot, crit variance, status-proc timing — all obey the Sigma Constraint (−1σ floor = no streaks; +1σ ceiling = no droughts, the awaited drop *must* come; reset per event-type). Damage rolls already ±1σ (Combat_Spec §4).
• Move-selection table (per phase): weighted P(move | state) folding in aggression_meter, player_habit_vector, env_phase, withdrawal_timer — then Sigma-bounded so the same punishing move can't roll thrice running and a telegraphed opener you're clearly waiting to punish won't be starved out. Roguelike variety, hand-tuned fairness.
• One line: *If a player could rage-quit over it, it's deterministic or Sigma-floored; if it would get stale being fixed, it's Sigma-bounded random.*

---

9. How difficulty scales — the vril dial, in full, zero stat inflation

AMENDED 2026-07-30 (twenty-eighth sitting item J5): the dial below stands whole and is JOINED by a second family of levers rather than replaced by them — higher difficulty raises the RANGE and COMPLEXITY of the attacks and of the environmental hazards. Bounded health and damage scaling is permitted as the minority contributor (Josh's own words: "ofcourse there will be some health and damage scaling but most of difficulty comes from complexity"), and what stays prohibited is difficulty sourced solely or mainly from inflated enemy numbers. §9b states the amended law in full.

One global float vril_density, five diegetic tiers (DIFFICULTY_SYSTEM.md, RULED & CLOSED): Wellspring · Veiled · Fading · Sundered · → The Forgotten One. It changes four world-demand levers and nothing on the enemy sheet:

1. Telegraph slack (4.1) — higher = longer readable tells; lower = "slack cut."

2. World generosity (4.3, 4.8) — higher = wider windows + faster wells ("vril runs to you as freely as water off a fall"); lower = wells "distant and slow to answer."

3. Answer-pruning (below) — higher = many valid solutions tolerated; lower = "easy answers thin out with the substrate," down to one at the prestige tier.

4. Withdrawal aggression (4.7) — lower density = clock bites sooner, steps harder.

Answer-pruning, precisely: every boss's counter-tree is authored primary + secondary + tertiary+ (Pillar-8 guarantees ≥2). The dial doesn't add power — it prunes leaves off your answer-tree. *Wellspring/Veiled:* full tree, off-mode/cheese-adjacent still work, the world catches you. *Fading (the true tuning — the depleted air the protagonist breathes 75/77 chapters):* primary + forced secondary, no hand-holding, no punishment either. *Sundered:* off-mode/inefficient pruned; execute the intended primary+secondary cleanly; wells sealed slower. *The Forgotten One (post-Ch-77 prestige, NG+ door, not a menu toggle):* pruned to the single true answer; wells sealed mid-trial; the deliberately-standard seventh vril-seat un-pruned via device-integration for *new depth* (not a 13th mastery tier).

Accessibility is orthogonal (Attunements): judgment-free reception layer (colorblind/multi-channel vril-polarity readability, timing windows, remaps) tuning *how demand reaches you*, never *how thick the air is*, gating no content, available in every tier including the Forgotten One. Vril-polarity multi-channel readability is a non-negotiable L1-playability requirement — the environmental-boss perception verb must be legible without color.

Headline improvement: every AAA game lies a little — "Hard" = 3× HP, double damage. Humanity's Hard = *your home reached the edge of its ability to help you*, and that sentence is simultaneously the difficulty setting, the plot, and the physics. Same authored boss, five world-tensions, one float. No enemy has a difficulty-scaled stat block.

---

9b. Nature-derived counter classes and the response vocabulary (RULED, Josh 2026-07-30)

Authority: docs/spine/DECISIONS_PENDING_JOSH.md, the twenty-eighth sitting item J5, and its authority home in _source/01_Tier_1_Foundation/T1_Combat_System_Spec [ACTIVE v1.0] under Nature-Derived Counter Classes and the Complexity-First Difficulty Law. This section is the systems-layer reading of that ruling; the ruling itself is the tier doc's.

The counter class is not a flag a designer picks. It is a reading of the attack's own nature — geometry, reach, duration, persistence, and the body and weapon producing it — and the systems layer treats it as derived data that authoring records and the gates check.

• Too sweeping, too indirect, or too large to dodge, so a parry answers it naturally. Nothing forces the class; the attack's shape produces it.
• Too powerful or too piercing to parry, so a dodge answers it, by the same logic in the other direction.
• Arena-wide, so neither answers it and the vocabulary expands: take cover behind rock or pillar or whatever the arena affords, strike a weak point, or trigger a reaction that stops the boss or the hazard.
• Parry is not only physical. Some parries require the opposite element, so the element matrix is a parry surface as well as a damage surface.
• Every move fits the boss's actual body, weapon, art, size, and scale, including what it throws and what it brings down — falling rock, lava, landslide, tidal wave, tornado, and the rest of the arena's hazards.

9b.1 The seven response classes and where each one is answered

The four earlier classes keep their meaning exactly. The three new ones are the arena-wide answers the ruling names, and each lands on a column that already exists rather than on a promise.

• any — no capability gate.
• dodge_only — the universal dodge, granted from the first playable chapter.
• parry_only — a parry-capable weapon, read from T0_Weapon_Registry.defensive_class.
• neutralize — an elemental counter read off T0_Element_Effectiveness.
• cover — arena geometry, read from T0_Environment_Grammar_Registry.tactical_affordance_class, whose declared enum already carries COVER (docs/proposals/systems/WS7_WIRING_BUILD_PLAN.md L95).
• weak_point — a named part, read from T0_Boss_Encounter_Registry.hitzone_parts and the rail's own part_ref.
• trigger_interrupt — an arena trigger, read from T0_Environment_Grammar_Registry.attack_verbs / loaded_potential / enemy_usable, which is the affordance set FR-020 already requires of every arena.

Two of the three new classes are answered per ARENA rather than per chapter, and the schema has no arena-to-env-feature join today, so gate 28 declares their resolvers not-armed rather than passing them. That is the named missing input, not a silent green: the join is the work order. The row-count half of that note is superseded by measurement — T0_Environment_Grammar_Registry carried 2 seed rows when it was written and carries 178 today, and the affordance those rows offer is now ruled rather than left to per-row taste (§9b.1a).

9b.1a The arming law — attack_verbs is derived from the row's own physics (RULED 2026-08-04, WQ_0038)

A leg-9 critic measured the column and found it near-dead: 13 of 153 rows armed any verb at all, and the thirteen were geothermal, hydro, gaseous, structural, or one volcanic-scree mineral row. The discriminating pair made it a class question rather than a chapter one — EG_0004 (Ash-and-Scree Volcanic Slope) and EG_0133 (the Ch-10 stratigraphic descent) are the same typed profile cell for cell, and only one of them carried a verb. EG_0128 (Ch 9, Afromontane scree) is the same profile and also carried none, so the divergence began upstream of both legs.

The ruling is that this is DRIFT, not intentional sparseness, and the argument is the column's own spec rather than a preference. docs/proposals/systems/WS7_WIRING_BUILD_PLAN.md L105 pins the fixed eleven-member D-verb set, and docs/translation/T99_Translation_Combat.md Step 5 states the keying law verbatim — the set is "keyed to each feature's physical_properties_array and loaded_potential", a lava vent IGNITEs and VENTs, a snowpack slope AVALANCHEs. That is a derivation rule, and under a derivation rule two rows with the same deciding keys must carry the same verbs. Divergence is therefore drift by construction. The alternative reading — that sparseness is the design — was steelmanned and fails on the same evidence: it cannot explain why one of two identical profiles is armed, and it starves three declared consumers (the trigger_interrupt response class above, FACTORY_CONTRACT FR-007's arena mutation vocabulary, and the Chaos D-verb binding at Translation Step 5). Maximal arming fails in the other direction, because it would arm rows whose own columns declare no potential and no gift, which is invention.

The law, and it is one sentence: every affordance a row DECLARES gets the verb that consumes it, every loaded potential a row DECLARES gets the verb that releases it, a row that declares neither arms nothing, and a firewalled row arms nothing regardless. The firewall is protected_care_ref naming any hard line or vril site, or alteration_permanence_default reading hard_blocked. It is deliberately NOT env_boss_verb_disabled: the founding WS7 seed row EG_0001 carries that flag TRUE and arms D2_IGNITE;D10_VENT in the same row, so the column governs what a boss may weaponise back, never whether the surface has verbs at all.

Applied as a corpus-wide sweep on 2026-08-04, additive only — no authored verb was removed, and the sweep is idempotent. 33 rows changed; 32 previously unarmed rows took the verbs their own columns already required, and EG_0019 (Felling Line and Cut-Timber Stack) gained the D2_IGNITE its fuel:flammable gift had always demanded, which is the same defect the boarded finding named on EG_0137 sitting undetected inside the armed set. Zero care-protected row was armed, and no care, permanence, or enemy_usable column was touched. The two rows the finding named are answered: EG_0133 now arms D4_AVALANCHE;D9_HURL on the EG_0004 precedent, so Ch 10's one strike encounter at olduvai_gorge has a real environmental verb and not geometry alone, and EG_0137 arms D2_IGNITE on ground the Zone Catalog already functions for combat.

The sweep overreached on fourteen rows, and the fresh-context critic caught it (finding F2, repaired the same day). Those rows' own notes had already refused the verb, in writing, for a canon reason the derivation rule cannot read: EG_0010, where the encounter's ruled outcome is rescue and not kill, so no ring-out affordance is authored; EG_0120 and EG_0130, where no Ch-9 ability arms an environmental verb at all; EG_0128, where an avalanche loosed on that slope would hit the endemic wolves and ibex the encounter exists to protect; EG_0058 and EG_0067, where every belt of that scrub runs against a village's crop and grazing. Arming them made fourteen rows say one thing in prose and the opposite in the cell, which is exactly the living-source defect that teaches a later agent to hallucinate. All fourteen are restored to none, and the refusal is no longer prose only — each row now carries a machine-readable ARMING DECLINED: <D-VERB>[;<D-VERB>] declaration in its notes. That is the exception mechanism the derivation rule needed and did not have: the law still binds every row that declares nothing, and a row that has genuinely ruled otherwise says so where a script can read it. Measured after the repair: 32 of 178 rows armed, from 13 of 153 before the sweep (14 of 178 in the live tree, EG_0171 having landed in between) — the "14 → 46" line in the sweep's own commit message is off by one against that arithmetic and is corrected here.

The exception cannot be used to escape the law, because the gate holds it in both directions. A declining row is exempt from the gift-consumed, potential-released and profile-consistency teeth for the verbs it names, and three new checks hold the declaration honest: every declined token must be a real D-verb; a row must never ARM a verb it declines, which is the tooth that stops a future sweep re-arming a refusal in silence; and a declaration must answer a live derivation pressure, so a row cannot decorate itself out of the ruling with an inert refusal. One idiom is deliberately not read as a refusal — the registry's own "typed forward and unarmed" (EG_0033, EG_0102, EG_0008) means the verb IS in the cell while no chapter ability triggers it, which is a different statement entirely.

What the derivation actually READS, declared rather than left implied (critic finding F3). The keying is over DECLARED TOKENS — ambient_state_gift and loaded_potential — plus three property readings: thermal_state:hot, a massless gaseous medium, and a high-conductivity hydro body. It does NOT read physical_properties_array generally, so a row carrying flammability:high without a fuel gift arms no D2_IGNITE, and a non-hydro row carrying conductivity:high without a conductive gift arms no D6_CONDUCT. EG_0019 and EG_0174 are the pair that shows it: same category, same flammability, and the only deciding difference is the gift's spelling (fuel:flammable versus material:raw_latex). Whether the property array should arm directly is a real question and it was NOT ruled at WQ_0038; it is boarded for its own pass rather than answered by a sweep. Until then the deferral is kept honest by measurement: the gate counts the blind spot every run — 14 rows on the flammability axis, 5 on the non-hydro conductivity axis at this writing — and FAILS when that census grows, so a new row can never widen an unruled gap unremarked.

The control is now armed on the right axis, which was the third half of the work order. harness/check_env_arming.py is a registered gate with seven teeth — firewall integrity, verb vocabulary, gift-consumed, potential-released, profile consistency, declared-exception integrity, and the property-axis blind-spot ratchet — each with a mutation fixture that must fire it and a control that must not. The old alterable-ground control measured alteration_permanence_default only, so a row read as satisfied while its usable-in-combat affordance was zero: a control armed against the wrong axis, which is the noisy-gate-conceals class. Proof the new one is armed rather than merely green: run against the pre-sweep registry it exits 1 with 33 findings, naming EG_0133, EG_0128, EG_0151, EG_0137 and EG_0019 by id; run against the repaired registry it passes. It is also the single durable deriver — any future arming pass imports required() and declines() from it rather than forking the rule into another script.

Declared unexercised, so a later reader knows it was measured and not overlooked: D3_FLOOD, D5_FREEZE and D8_GROUNDWAVE have no arming trigger, because no live row declares an impounded potential, a frozen-water gift, or a groundwave-bearing property. thermal_state:cold is specifically NOT mapped to D5_FREEZE, because EG_0014 declares that gift and deliberately arms D11_OBSCURE alone.

9b.2 Elemental parry, derived from the recovered matrix

T0_Element_Effectiveness holds the Josh-authored grid as 124 rows (docs/original_predclaude/ELEMENTS_MATRIX_RECOVERED.md). The rail carries three columns for the elemental half: attack_element (the attack's own element), parry_element_rule (physical / element_effective / element_named), and parry_element_ref (the named element when the rule narrows it).

• The answering set for an attack element is every attacker element the grid rates Effective (value 1) against it as defender. Water answers Fire because the grid says so.
• A named element must itself be Effective against the attack element, and gate 28 fails a named element the grid does not rate Effective.
• The grid never says "opposite", it is asymmetric, and Earth, Fire and Metal each rate themselves Effective, so no unique pair exists to read. "Opposite element" therefore reads as effective-against — RATIFIED by Josh at the twenty-ninth sitting (2026-07-30), whose own wording is "the opposite or 'strong' element to the attack" — the strong element against an attack is the one the grid rates Effective against it. The interim interpretation is closed (the T1 section carries the same ratification), and the Phase-5H effectiveness calibration now tunes magnitudes rather than deciding the reading.
• The grid as landed carries 13 attacker elements and 13 defender elements, and all 13 defenders have a non-empty answer set (measured off the registry, not assumed). The axes are not the same 13 — Life only attacks, Vegetation only defends — and that asymmetry is the authored source's, carried not smoothed. The live vocabulary is larger still (31 values by the matrix doc's own count), so every element outside the grid has no derivable elemental answer, and an elemental parry authored on one is reported unanswerable rather than passed.

9b.2b The mitigation scale — a successful parry is not a promised zero (RULED, Josh 2026-07-30)

Authority: docs/spine/DECISIONS_PENDING_JOSH.md, the twenty-ninth sitting, and its authority home in _source/01_Tier_1_Foundation/T1_Combat_System_Spec [ACTIVE v1.0] under The Mitigation Scale. Josh's words: "There's need to be a scale there. Not all 'parry's' whether physical or magic will negate 100% damage, if it's a much stronger attack then their parry or opposite magic then they'll still receive some damage."

The elemental parry is performed rather than selected: at the right parry-window moment the player puts up an elemental blast of the opposite or strong element forward of the incoming attack, and survives untouched or takes reduced damage. §9b.2 above says which element answers; this section says what the answer is worth.

• The outcome of a correct in-window response is a SCALE from full negation down to partial mitigation, positioned by the attack's strength against the answer's strength (the parry itself, or the opposing element's blast).
• Both halves are covered, the physical parry and the elemental one. An attack much stronger than the answer leaves residual damage through either.
• The counter class states the response the attack's nature demands, never the outcome it guarantees. neutralize, parry_only and any are demands to be met, not clean-zero promises, and no rail column implies a zero-damage result.
• The scale is ORDINAL here by rule. No coefficient, threshold, band count or percentage is stated in this doc: magnitudes are Phase-5M's under the tuning firewall, and this section owns the shape only.
• It composes with §9b.3 rather than adding to it. Higher difficulty means stronger and more complex attacks, so the same parry mitigates less against them — a consequence of the two rulings meeting, not a new difficulty mechanism.
• The damage formula already carries the shape this ruling names. T1_Combat_System_Spec under Defensive Layer Integration states that defensive successes "reduce or negate" as "a damage multiplier at zero or fractional value"; the ruling names what selects between zero and fractional, and makes the fractional branch doctrine rather than an implementation detail.
• No schema column is minted for the scale here. The ordinal mitigation column belongs on T0_Boss_Attack_Rail and mints with that rail's first authored moveset (window-plan rank 16), because a column with no rows behind it is a promise rather than a shape.

9b.3 Difficulty is complexity, and the anti-inflation thesis survives

Higher difficulty raises the range and the complexity of attacks and of environmental hazards. Some health and damage scaling exists; most of the difficulty comes from complexity. Both halves are the ruling, and the ordering is the binding part.

• The complexity levers join the four density levers of §9 rather than replacing them: more of the moveset live, wider attack range, more hazards active at once, longer interleaved cycles, more arena state demanding a read.
• Bounded health and damage scaling is permitted as the minority contributor. What stays prohibited is difficulty that comes solely, or mainly, from inflated enemy numbers — which is what §0's thesis was always arguing against.
• New Game Plus is ruled enhanced difficulty and complexity and reads this same law.

9b.4 What is untouched

The chain meter stands whole (visible pips, five to seven or more correct reads in a row, wipe on any mistake). The Ch-2 reduced scope stands (docs/FUN_REBUILD_PLAN.md R2.2). The environmental-boss protocol stands (perception, vril-management, restraint, purgation). A parry-only attack defeats a dodge by its nature — geometry, duration, persistence — rather than by a flag that switches off i-frames, so the implementation expresses the physics and the class stays readable.

---

10. Worked example (reactions + counters + environment integrated)

Mid-game Tier-A boss, strike / declare (direct→character), blend 40/35/25, obelisk-lit desert arena, at Fading:

1. P1 (direct spine): readable sweeps; punish on the window, chip the direct Composure node; obelisk env_cycle pulses vril every ~18s — time your Vril-Pulse to it to stay flush.

2. Node 1 breaks → P2 pushes a layer: add-waves arrive; pure-DPS now steps the Withdrawal clock down. Forced secondary = environmental: trigger the sand-collapse trap on-cycle to clear the wave and expose a second node. P1's sweeps are *still live* (evolve-not-swap).

3. Node 2 breaks → P3 recontextualizes (character): the boss fights to be *heard*; win becomes a declare/weigh resolution (Entu template), integrity-band-routed. Damage alone won't end it — the Pillar-8 floor.

4. Adaptation throughout: left-dodge spam in P1 → P2's sweep baits left; turtling → aggression biases toward guard-breaks. Sigma-bounded = smart, not cheating.

5. Dial to Sundered: *same fight* — tells tighter, obelisk answers slower, environmental+character answers now required (pure-DPS leaf pruned), Withdrawal clock quicker. Nothing on the boss got bigger.

---

11. What is novel (the "don't steal, improve" ledger)

1. Difficulty = world density, not enemy stats — the most-copied AAA convention rejected and replaced with a diegetic physics dial that is *also the cosmology*. No competitor ties its slider to its metaphysics.

2. Composure you break with any of 7 modes — posture that stealth/puzzle/environment/familiar/vril/astral all fill, making single-verb play *mechanically* too slow: Pillar-8 enforced by pacing math.

3. The Withdrawal clock replaces the enrage timer — same escalation, zero inflation; failure makes the realm recede (the plot in real time).

4. Death pays diegetic knowledge (revealed tells) — Hades' cadence delivering the Souls fantasy; you win next time because you *learned*, demand unchanged.

5. **The boss learns *this run* (Sigma-bounded adaptation)** — smarter, not stronger; fair by construction (the σ floor forbids frame-perfect punishes).

6. Answer-pruning as hard mode — remove your shortcuts down to the one true answer, instead of adding enemy power: a more respectful "harder."

7. Environmental bosses are real, kinetic, L1-fun encounters with a non-damage win — "you cannot DPS this" as a *system* (polarity-match + restraint budget), not a cutscene.

---

12. Boundaries — what this does NOT decide (cardinal rule)

Mechanics only. It does not assign which boss sits in which chapter, each boss's verbs, blend weights, phase counts, or win conditions — those are the roster rows Josh authors/approves (docs/proposals/ROSTER_PROPOSALS.md + boss_roster_rows.json, marked *proposals*). Absolute numbers (windup ms, window ms, σ values, cooldowns, dial coefficients) are deferred to the Phase 5M balance audit exactly as Combat_Spec §4 defers them; coefficients above (e.g. 0.6 + 0.4×density) are illustrative shape, not locked values. No hard line is touched; each extension I added (Composure bar, Withdrawal clock, habit-adaptation, death-knowledge economy) is labeled, grounded in an existing anchor, and is a [DECISION NEEDED - JOSH]-class *proposal*, not a canon assertion.

---

PART C — Local-LLM Runtime Generative Layer (index; full: docs/RUNTIME_GENERATIVE_LAYER.md)

What it delivers

A buildable design for the runtime local-LLM generative layer that treats the dumb local model as pure upside over a deterministic truth spine — the /dev/continuo lesson (ROADMAP §8, START_HERE.md wiring doctrine) ported from build-time to runtime. Grounded in the repo's real artifacts, not invented ones:

§1 — The split (the load-bearing table). DETERMINISTIC = anything wrong-in-a-way-a-player-notices: damage, phases/timers/enrage (T0_Boss_Encounter_Registry), drops (T0_RNG_Drop_Table), abilities (T0_Ability_Tree_Registry, 185), fauna spawn/bonding (T0_Creature_Roster), worldstate/integrity-band, all hard-line-bound content, cultural depiction (HL_0002 primary-source rule), character canon (118), voice identity (§17.12/§17.5). GENERATIVE = surface only: boss taunts/barks, adaptive tactical hints fed by the sim's *own* next-move intent, fauna behavior flavor, NPC ambient dialogue/reactions, persona-bounded probe responses. The rule: the model performs, it never adjudicates.

§2 — Guardrail loop. A four-stage deterministic sandwich around one model call: (1) precomputed context pack = recall.py's "all-canon-for-a-node" query frozen per entity (persona + hard-line rule vector + fact whitelist + care band); (2) grammar-constrained decode (malformed output impossible by construction); (3) post-gen validator = runtime cousin of lint_canon_graph.py gates (denylist → hard-line rule vector → canon-graph name resolution incl. retired-alias reverse index → persona → anti-slop); (4) authored fallback on any hard-fail or latency miss. The model owns zero facts; the deterministic layer owns all of them.

§3 — AAA constraints. Latency budgets per class (≤120 ms P99 barks, off-render-thread), 3–8B quantized / 2–5 GB VRAM single shared instance, prefix-cached persona preambles, priority batching + speculative ambient, seedable sampler for replay/speedrun determinism, no write path into the sim, fully offline.

§4 — Alive without slop. Bosses that needle your *actual* death-cause because the sim told the model; folklore-true fauna; bounded NPCs whose deflection is itself characterful; anti-slop as a first-class gate. The differentiator from typical "AI NPC" demos: they let the model own facts and hope; here lies are structurally impossible.

§5 — Failure/reward encoding (the continuo core). A concrete F1–F10 table: each dumb-LLM failure mode → its deterministic catcher → its reward-pattern encoding, with F8 (malformed) and F10 (state mutation) *prevented by construction* rather than caught. Every caught failure feeds a local runtime_generation_log that grows the fallback bank + denylists and becomes eval fixtures; a nightly replay scores candidate model/prompt versions with the same §2.3 checks — a truth-recall-style first-try-pass-rate metric analogous to continuo's 0.965.

§6 — Buildability + Josh flags. Extends existing artifacts (canon_graph.json, planned recall.py, T0_Hard_Lines machine columns, T0_Voice_Registry substrate), reuses gate logic, and carries two [DECISION NEEDED - JOSH] flags (launch posture for high-care Ch 29/42/57 NPCs; target model/quant) plus the do-not-invent reminder that the runtime never self-canonizes — honoring the cardinal rule.

One note for your awareness: harness/recall.py does not yet exist on disk — it's the planned Stage-4 component in CANON_INTEGRITY_ARCHITECTURE.md §D/§E. The spec treats it as the shared truth-store interface for both build-time and runtime and adds one new offline exporter (recall packs) on top of it.

---

PART D — 5-Chapter Milestone Pipeline (index; full: docs/MILESTONE_BUILD_PIPELINE.md)

What it contains, grounded in the existing contracts

(1) The nine engines + dependency order. A contract-cited table (E0 book/research → E1 spine → E2 region → E3 boss → E4 quest → E5 lore → E6 art → E7 audio → E8 UE build), each with its owning BPC section, canon inputs, output store (live registries/ CSVs + build-asset store per PIPELINE_ADAPTER), and the recall.py query it uses. Plus an ASCII critical-path diagram and the hard ordering constraints, each traced to a real BPC cross-pipeline linkage (E1 locked state gates all downstream; E2 zones precede E4/E6; E6/E7 are leaves reading what E3/E4/E5 wrote; E8 is the read-only build-time sink).

(2) The 5-chapter milestone model. Recommends Ch 2-6 as Milestone 1 with node-by-node grounding (Flores → Manggarai → Bali → Borobudur → Sri Lanka), and five defenses: it's the head of the canonical build order, one coherent region cluster (shared books/reviewers), exercises the full verb floor plus the first environmental-protocol seed, contains zero §17.1 high-care chapters (debug the pipeline where mistakes are cheap), and nests wholly inside RR-12. Defines the repeatable 8-stage loop and a rigorous exit gate (machine gates scoped to slice nodes + critic/Josh batch sign-off + *actually playable*), tagged milestone/ms-NN.

(3) Cost staggering. Books filtered to ~9 titles for Milestone 1 (the exact BOOK_ACQUISITION Ch 2-6 anchors), online-first, with the expensive §17.1 acquire-ALL clusters isolated to their own later milestones; scanning scoped to the irreducible per-slice remainder (OCR tooling dormant otherwise, no ChromaDB); gating authored per-slice-row (flat per-milestone burden, free whole-file regression coverage).

(4) recall.py + graph per engine. Every engine is a recall.py consumer, inheriting the cited/complete/no-hallucination output contract, with the specific query shape per engine (recall node CH_NN as the load-bearing cascade query; recall band + recall hardlines for boss/art/audio safety; recall aliases as the §17.7 rename guard; recall counts at the exit gate).

(5) Boss/combat + local-LLM per chapter. E3 builds to the CHAPTER_BUILD_TARGET bar (2-4 sub-bosses across different threads + 1 main; Pillar-8 doctrine; the exact Milestone-1 roster rows from the approved proposal). The local-LLM layer is placed precisely in the BPC's runtime inference tier (never authoring canon): runtime NPC barks, within-encounter reactive behavior tuning, and the UE 5.8 MCP build-time authoring assist — all reading compiled data tables only, constrained by the baked-in recall hard-line fields.

Three [DECISION NEEDED - JOSH] items surfaced with recommendations (slice size 5 vs RR-10 window; Ch 2-6 as the first slice; local-LLM runtime model pin). No world canon invented; the doc sits correctly in the T99 operational tier alongside PIPELINE_ADAPTER/ROADMAP and does not turn the harness gates red.

---

APPENDIX — Combat/Boss/Quest Canon Audit (the baseline the design ENHANCES, never contradicts)

Combat, Boss & Quest-Phasing Canon Audit — the baseline the design must ENHANCE, never contradict

Scope: what *Humanity* has already written about combat, bosses, and quest phasing, as of the 2026-07-04 repo state. Every claim carries a cite. Three parts: (1) CANONIZED mechanics, (2) encounter LAYERS canon implies but has not systematized, (3) GAPS a boss/encounter needs mechanically that no doc specifies.

Source abbreviations and authority order (CVD wins all conflicts per CLAUDE.md):

• CVD = _source/01_Tier_1_Foundation/T1_CVD_Creative_Vision_Document [ACTIVE v1.4].md
• COMBAT = _source/01_Tier_1_Foundation/T1_Combat_System_Spec [ACTIVE v1.0].md
• SCHEMA = _source/00_Tier_0_Master_Indices/T0_Schema_Dictionary [ACTIVE v1.3].md
• ROSTER = docs/proposals/ROSTER_PROPOSALS.md — [PROPOSAL, not canon]
• BOSSJSON = docs/proposals/boss_roster_rows.json — [PROPOSAL, not canon]
• DIFF = docs/proposals/DIFFICULTY_SYSTEM.md — [RULED 2026-07-04]
• BUILD = docs/proposals/CHAPTER_BUILD_TARGET.md — [rulings 5-7, work orders]

Status note that governs the whole map: CVD §12 + Pillar 8/11/12 and COMBAT are hard canon. The 75-boss roster (BOSSJSON), the per-chapter sub-boss rule, the difficulty assist layer, and the encounter-layer taxonomy in Part 2 are PROPOSALS/RULINGS awaiting batch-1 sign-off — not yet registry-populated (ROSTER lines 3, 446; BOSSJSON status field). The T0_Boss_Encounter_Registry is a bare stub: one title row, zero data rows (registries/T0_Boss_Encounter_Registry [DRAFT v0.1]/Sheet1.csv), and mode_variety_check is populated on only the two fairy-realm bookend chapters in T0_Chapter_Index, everything else [POPULATE→...] (Chapter_Index CSV rows 2-3).

---

PART 1 — What is CANONIZED (the immutable baseline)

1.1 The blended-mode boss doctrine (Pillar 8) — hard canon

• Every boss deploys combat + weapons + abilities + puzzles + character engagement in blended combination (CVD L214-217, restated COMBAT L207, L372). This is Gate 8, "immutable canon" (CVD L163).
• No boss = a single mode. Explicit prohibitions: "No boss = pure DPS check. No boss = pure dialogue tree. No boss = pure stealth" (CVD L217; COMBAT L207, L372).
• Executable test (pillar violation if failed): "Boss encounter scoring single-mode in audit = pillar violation. Boss without character-engagement field = pillar violation" (CVD L218). So *character engagement* is a mandatory field on every boss, and single-mode dominance is an audit failure.
• The canonical template is the Ch 17 Entu Priestess three-phase structure: surface combat phase → moral descent phase → philosophical confrontation phase (CVD L217, L682; COMBAT L207, L372). This is named as *the* template every boss inherits.
• Auditing is registry-driven: "Audited per boss in T0_Boss_Encounter_Registry" (CVD L682, L219). The registry is the compliance surface.

1.2 The seven combat modes (§12.2) — hard canon

CVD §12.2 (L654-662) enumerates the modes every chapter must combine ("No chapter coasts on one combat mode"):

1. Direct combat (weapon-and-ability engagement)

2. Stealth / restraint engagements

3. Puzzle-routed combat (boss encounters with non-combat resolution components)

4. Environmental hazards

5. Vril-management combat (routes through vril resource discipline)

6. Familiar-leveraged combat

7. Astral / dream-world combat (Hard Line chapter sanctuary preservation per Pillar 3)

Distribution rules (hard canon, CVD L664; mirrored COMBAT L366-367):

• Two consecutive chapters never share dominant combat type.
• Every five-chapter window contains three or more distinct combat-mode dominant chapters.
• "The system does not change per chapter. The chapter's *mode emphasis* changes per chapter" (CVD L666) — one unified system, per-chapter emphasis, not bolted-on subsystems.

COMBAT §"Combat Variety Framework" extends this to ten encounter categories for its own distribution audit (COMBAT L319-330): person-to-person, person-to-animal, group combat, boss encounters, social bosses, environmental/atmospheric bosses, stealth, naval, mounted, and vril-projection combat. (Note this is a ten-category encounter taxonomy distinct from the CVD seven-mode list — see Gap 3.7.)

1.3 The environmental / atmospheric boss protocol — hard canon

A distinct boss type that is not a person, animal, or conventional opponent — an environmental/atmospheric vril-state that cannot be defeated through damage (CVD §12.3 L668-676; COMBAT §9.2 L332-360).

• Victory is not measured in damage: "They are not 'non-combat puzzles.' They are combat encounters whose victory condition is not measured in damage" (CVD L676).
• Canonical 4-step resolution protocol (COMBAT L336-341): Perception (distinguish genuine vril echo from negative-consciousness manifestation) → Vril-management (resonate with the genuine echo without feeding the negative) → Restraint (no combat-scale vril expenditure — it would destroy the genuine echo too) → Purgation (the negative is purged, the genuine echo preserved).
• Capability baseline: success determined by Intuition, Constitution, Integrity. "No combat. No dialogue. No stealth. Perception and vril-state management only" (COMBAT L358).
• Graduated-consequence failure states (COMBAT L343-347) — this is the canon Hades-loop seed:
• Combat-scale expenditure destroys both → environmental degradation + faction loss.
• Insufficient vril management feeds the negative manifestation, strengthening it for the Ch 39-onward revisit.
• Partial resolution → graduated consequence (some purged, some remains).
• Six documented canonical instances (COMBAT L349-356): Ch 10 San Rock Art Spirit (mechanic establishment), Ch 20 Silk Road oasis ruins, Ch 29 Sámi drum site, Ch 38 Azores triple-plate junction (Midpoint), Ch 45-46 Southwest visionary-site echo, Ch 57 Australia Dreaming-site.

1.4 Restraint / protected & relief chapters — hard canon

• Relief chapters (Ch 27, Ch 52, Ch 57) are structural valleys, NOT combat-absent. "Combat content remains canonical at each; intensity drops below adjacent chapters but presence does not" (COMBAT L368; mountain-range pacing CVD §16.1/§16.4/§17.10 per ROSTER L305).
• Hard Line chapters protect thematic sanctuary via Velheim-intrusion / astral combat, NOT combat absence (Pillar 11, CVD L237, L239; Pillar 3 cross-ref). "Hard Line chapter without Velheim or astral combat layer = pillar violation" (CVD L239).
• Ch 57 is the most protected: canonical hard line = no antagonist family presence (COMBAT L371, L495), maximal research depth, AIATSIS-protocol compliance, no songline-imitating content (COMBAT L307). Explicit corrected memory item: "Ch 57 has no combat" is WRONG — the hard line is *no antagonist-family presence*, not no combat (COMBAT L495, the doc's own Recent-Changes cleanup note).
• Ruling 5 (BUILD L57-64): at protected/high-care chapters the §17.1 ban is on combat against the protected people only; their oppressors (colonizers, slavers, raiders, corporate predators) are canonical combat, and rescuing the protected people is a first-class combat layer alongside the environmental/astral protocol. Fauna combat is permitted (BUILD L63) — though this collides with a T4 source at Ch 57 (see Gap 3.8).

1.5 The damage formula & defensive surface — structurally canonized, numerically deferred

• Formula structure locked at Session 5H; numbers deferred to Phase 5M audit (COMBAT L47, L92-94). Final damage = base weapon damage × cast-pattern mod × persistence-class mod × damage-scaling mod × integrity-band mod × RNG sigma factor, ability-layer damage applied additively after weapon multiplication (COMBAT L51).
• Integrity-band modifier (mirrored 5-band, hard canon): L1 +10% healing received; L2 +5% healing received; L3 +5% all damage received; L4 +5% damage dealt; L5 +10% damage dealt +10% vulnerability (COMBAT L69-73). Combat behavior is integrity-state-dependent by canon.
• RNG sigma constraint (hard line, CVD §17.13): damage rolls bound within ±1 sigma; anti-streak floor at −1σ resets on drop; mandatory drop ceiling at +1σ (COMBAT L79). "RNG should reward patience but never punish exhaustively" (ROSTER L304). This binds all difficulty modes (DIFF/ROSTER L347, L393).
• Defensive surface canonized as block / parry / dodge (COMBAT L376-423), each mapped to specific abilities/weapons (Fluid Dodge, Wind Step, Stone Skin; Shield/Great Shield block; Rapier/Madu/Cutlass/Long-Sword parry). Defensive successes apply before integrity-band and sigma factors, as a "damage multiplier at zero or fractional value" (COMBAT L423). Timing windows numerically deferred to Phase 5M (COMBAT L392, L419).
• Vril-seat-element matching bonuses canonized structurally (25% matched-element, 40% on Ability 2 at Tier 3+ vril-seat) (COMBAT L85-88); absolute values deferred.

1.6 The five Tier-A boss anchors — hard canon (corrected)

• The canonical Tier-A anchor set is Ch 8 / 23 (or 25) / 38 / 55 / 65 (COMBAT L209-215). Each aligns with a hard-line spine chapter peak.
• Ch 8 = early-game vril-mechanic capstone; Ch 23 = mid-Tier-2 trade-arc gate (anchor #2 RULED-LOCKED to Ch 23, D-COMBAT-TD R9, 2026-07-19); Ch 38 = Azores convergence / Midpoint; Ch 55 = Atacama; Ch 65 = Evil-bucket collapse.
• The "fifth anchor at Ch 58" was a confirmed canon ERROR, deleted in commit 7e23381 (ROSTER L20; BOSSJSON correction_applied). Ch 58 as a *non-anchor* boss is open ledger decision D-9.
• Two boss scales are categorically separate (COMBAT L203, L219-247): chapter-level Tier-A Solo bosses vs. 100-player Tier-C MMO World Bosses. Tier-A completion never requires a World Boss. World Boss roster scopes at Session 5J (T1_Tier_C_MMO_Spec), and its data — not the solo bosses — is what the T0_Boss_Encounter schema was originally scope-locked for (COMBAT L247; SCHEMA L848).

1.7 Unified twelve-tier mastery & quest-gating — hard canon

• Every developable capability progresses on the same twelve-tier Hermetic ladder (CVD §12.1 L604-648): Neophyte → Zelator → Theoricus → Practicus → Philosophus (Outer, 1-99) → Adeptus Minor → Major → Exemptus (Inner, 100-199) → Magister Templi → Magus (Prestige, 200+). Combat, trades, ability tree, weapon mastery, familiar bond, language, inscription all share it.
• Tier advancement is quest-gated, never numeric-threshold alone: "Every tier requires a quest-unlock, not just numerical threshold hit" (CVD L977, and per-tier gates L614-646). Adeptus Minor (T6) begins substrate perception and requires a canonical initiation questline (CVD L633). Magister Templi (T9) unlocks ONLY after Ch 77 completion (CVD L643).
• Known-issue-2 preserved verbatim: CVD calls it "twelve-tier" while naming ten tiers (Neophyte-Magus) — flag only, never auto-correct (CLAUDE.md Known-issue-2; ROSTER L420).

1.8 The vril-density difficulty dial — RULED canon (NOT HP inflation)

DIFFICULTY_SYSTEM.md is RULED and CLOSED (Josh, 2026-07-04, DIFF L95). This is the strongest post-baseline addition:

• The dial is ambient vril density — the fairy realm's reach — and it is joined by the complexity levers. "The dial is ONLY vril density; NEVER HP / damage / resistance inflation" (DIFF L12-13), AMENDED 2026-07-30 (twenty-eighth sitting item J5): higher difficulty also raises the RANGE and COMPLEXITY of attacks and environmental hazards, bounded health and damage scaling is permitted as the minority contributor, and the prohibition that stands is difficulty sourced solely or mainly from inflated enemy numbers (§9b). Grounded in the vril-withdrawal canon (T1_Vril §1 threshold-of-perception, §10 contraction-not-collapse / depletion-as-process; DIFF L4-9).
• Four tiers on the realm's-fading-reach axis: Wellspring (Story) · Veiled (Easy) · Fading (Medium — "the game as authored," the depleted air the protagonist breathes 75/77 chapters) · Sundered (Hard — "Nothing here hits harder... you must do more because the world gives less") (DIFF L24-57).
• The Forgotten One — prestige New-Game+ easter egg behind the Abyss-crossing (an earned rite per CVD §12 quest-unlock rule); answer-pruning, never number inflation: "No HP, damage, or resistance value differs, in either direction" (DIFF L59-80).
• Attunements = a separate, always-available, judgment-free *accessibility* layer, NOT a difficulty tier; gate no content; vril-polarity multi-channel readability is a non-negotiable L1-playability requirement (the environmental-boss perception verb must be readable without color alone) (DIFF L82-92).

The proposal-stage difficulty document (ROSTER §5) supplies the underlying design contract that the roster rows obey — one canonical difficulty, in-fiction relief valves, defeat model, assist toggles — but its *names* are superseded by DIFF (ROSTER L1 banner). Load-bearing pieces still live:

• Tuning contract (binding on every boss row, ROSTER L308-313): ≥1 counter per signature attack; ≥2 counters across build variety so no build is locked out; every boss tests its stated primary verb (no undertuned non-tests, no tests of untaught verbs); no mandatory external scaling item / meta-currency (D-COMBAT-TD R2, Josh 2026-07-19: every combat boss's Preparation axis must carry ≥1 FREE in-fiction knowledge/skill route — never a purchase-only route, never a mandatory external scaling item); no pure DPS check; phases evolve existing moves per E33 telegraph-escalation.
• The defeat model (proposed new canon, ROSTER L330-334): on defeat, consciousness withdraws to the last attuned recharge source (recharge-as-checkpoint); every boss defeat writes a journal observation recording one telegraph-and-counter; defeat AND victory both advance something — the Hades loop without an XP-on-death economy.

1.9 The per-chapter build target — RULED work orders (rulings 5-7)

CHAPTER_BUILD_TARGET.md (BUILD) canonizes the per-chapter quality bar Josh ruled 2026-07-04:

• Boss structure: 2-4 sub-bosses/mini-bosses culminating in ONE main chapter boss, surfacing across DIFFERENT of the 22 threads — never all from the House-of-Velheim operation (BUILD L10-17).
• Pillar-8 doctrine binds every boss (BUILD L18-22): blended/never-single-mode; primary + secondary verb; three-axis Patterns/Puzzles/Preparation split, never 100/0/0; phase escalation that EVOLVES the moveset (tighten/multiply telegraphs), never swaps it; ≥2 counters per signature attack; a Hades reward on win AND on death.
• Non-standard main bosses are real playable fights, not cutscenes — the Ch 69 materialized-argument precedent generalizes (BUILD L23-25).
• Everything-woven 12-facet checklist (ruling 7): combat, exploration, discovery, puzzles, mystery, lore, trades, gathering, side quests, engineering, folklore, superstition — each a real playable/investigatable presence, none siloed (BUILD L46-54).

1.10 The agility-course system — hard canon (traversal layer)

77 courses, one minimum per chapter; three variants (Standard / Hardcore / Puzzle); timed, hazard-rich, tied to Book of Trades and collectibles; biome-rotation discipline; integrity-band cosmetic title variants on Hardcore (COMBAT L249-311). This is the canonized traversal-challenge layer parallel to the boss layer.

---

PART 2 — Encounter LAYERS canon IMPLIES but has NOT systematized

Canon defines individual encounter *types* and mastery *anchors*, but the only place these are assembled into a coherent per-chapter encounter stack is the PROPOSAL layer (BUILD + BOSSJSON), which is not yet ratified or registry-populated. The layers below exist implicitly in canon; the systematization is the design gap.

Cross-cutting observation: BUILD L76-80 is the only place that stitches these into one pipeline ("expand from the approved 75-row main-chapter-boss layer to add the per-chapter sub-boss and thread-woven rows; School_Registry, Questline_Index, trade/folklore/creature registries supply the woven facets"). That stitching is a ruling, not a built system — the registries it names are empty stubs.

---

PART 3 — GAPS: what a boss/encounter needs mechanically that canon has NOT specified

These are the holes the design must fill. Each is a thing Pillar 8 / §12 / the build target *demand* but no doc defines.

3.1 Numerical combat values are entirely deferred

The whole damage formula, integrity-band percentages in absolute units, sigma deviation in absolute units, vril-seat-match bonuses, cost-class→damage-range mappings, and every defensive timing window are deferred to a Phase 5M balance audit that has not happened (COMBAT L92-94, L392, L419). No boss can be numerically tuned, and "no boss is a pure DPS check" is currently unverifiable because there are no DPS numbers at all.

3.2 The T0_Boss_Encounter_Registry is an empty stub, and its schema was built for MMO World Bosses, not solo chapter bosses

The registry has zero data rows (registries/T0_Boss_Encounter_Registry [DRAFT v0.1]/Sheet1.csv). Its schema (SCHEMA §4.1 L844-873) is production-pipeline-shaped — boss_id, phase_breakdown_object, engagement_modes_object, ue5_boss_class_ref, mesh/voice/cue/cinematic FKs — and is scope-locked "at MMO build phase," authority split between COMBAT (Gate 8) and T4 (Cassius) (SCHEMA L846-848). It has no columns for the things the roster actually needs: primary/secondary verb, three-axis Patterns/Puzzles/Preparation split, win-condition enum (HP/part-break/pacify/perceive/restraint), per-signature counter enumeration, escalation vector, or the Hades win-AND-death reward. The proposal's rich row model (BOSSJSON) has no schema to land in. This is the single largest structural gap.

3.3 No win-condition taxonomy is canonized

The proposal uses a de-facto enum — HP / part-break / pacify / perceive / restraint (BOSSJSON rows) — but no canon doc defines these victory conditions or their mechanics. CVD says environmental-boss victory "is not measured in damage" (L676) but never enumerates the alternatives. "Pacify," "part-break," and "perceive" as win states are invented at the proposal layer.

3.4 "Character engagement" is a mandatory Pillar-8 field with no mechanical definition

CVD L218 makes "Boss without character-engagement field = pillar violation" an executable test, but nowhere defines what the character-engagement mechanic IS — dialogue rounds? moral choice? relationship state? The Entu template gives a *shape* (moral descent → philosophical confrontation) but no system. Every boss must have it; no boss knows what it is. ✅ RULED (D-COMBAT-TD R4, Josh 2026-07-19): character-engagement is a STANDALONE 8th Composure node, decoupled from the astral node (astral stays Pillar-11 hard-line-gated) — integrity-band-routed, present on every MAIN boss for ≥1 phase, verbs varied and woven-not-bolted onto pure-strike mains; a talk-boss's break = completing the recitation/refutation. Machine form = the character_engagement_node column.

3.5 The three-axis Patterns/Puzzles/Preparation split has no source authority

The "never 100/0/0" axis split is central to the build target (BUILD L20) and every BOSSJSON row carries an axis_split, but it originates in docs/ROADMAP.md §0.E design principles — not in CVD or COMBAT. It is a proposal-layer construct that needs ratification to become a bindable spec (and needs its axes *defined*: what counts as "Preparation" mechanically?).

3.6 Phase-transition / escalation mechanics are asserted, not specified

"Phases evolve existing moves, never swap the moveset" (E33 telegraph escalation; BUILD L20, ROSTER L313) is a strong doctrine, but no doc defines how a phase transition triggers (HP threshold? part-break? timer? narrative beat?), how telegraph timing scales, or how "tighten/multiply telegraphs" is parameterized. The Entu three-phase template is narrative, not mechanical.

3.7 Two divergent mode/encounter taxonomies coexist unreconciled

CVD §12.2 lists seven combat modes (L656-662); COMBAT §"Combat Variety Framework" lists ten encounter categories (L319-330). The "no two consecutive chapters share dominant combat type" rule is audited against which taxonomy? The mode_variety_check schema (SCHEMA L144) tracks a *third* set — five booleans: combat / puzzle / learning / character / transformation. These three vocabularies need a single canonical mapping before any mode-variety audit is runnable. (And mode_variety_check is populated on only 2 of 79 chapters — Chapter_Index rows 2-3.)

3.8 Live source contradictions at protected chapters

• Ch 57 fauna combat: RULED (D-COMBAT-TD R9, 2026-07-19) — environmental + astral ONLY; no fauna combat, no person/training combat (the saltwater croc stays a staged hazard, never a cull); T1_Combat's Ch-57 clause rewritten accordingly; the care-tier gradient ratified (the tradition's own monster IS fought at Ch 43; nothing living targeted at the max-care hard-line). Former conflict text: COMBAT permitted person-to-animal combat; T4 said none. A boss/encounter designer cannot resolve Ch 57's fauna layer until ruled.
• Ch 43 region: RULED (R9) — BOTH regions as built (Appalachia opioid epicenter PRIMARY + Gullah Geechee Sea Islands SECONDARY, both at witness-depth); neither source doc overruled; ledger D-3 closed.
• Ch 59-63 revisit order: CLOSED-AS-RECONCILED (R9) — the sources already agree (59 Africa / 60 Asia / 61 Europe / 62 Americas / 63 Pacific); the ledger's cited conflict was a stale line-shift artifact; D-5 closed.
• Ch 23-vs-25 anchor #2: RULED (R9) — Ch 23 IS Tier-A anchor #2; the T1_Combat fork deleted; BE_0023 carries the machine tier_a_anchor_2 token.

3.9 Familiar-leveraged combat (mode #6) is named but unspecified

It is one of the seven canonical modes (CVD L661) and a proposed relief valve (ROSTER L321) and assist (ROSTER L366), but no mechanics exist for how a bonded familiar fights, commands, or contributes to a boss encounter. 22 familiar slots and 33 legendary species are numerical canon with no combat behavior spec.

3.10 Vril-projection combat (a mode) is a one-line stub

"At high vril mastery, combat proceeds through vril projection rather than weapons" (COMBAT L330) — an entire high-mastery combat mode compressed to one sentence, with a proposed aim-assist (ROSTER L367) but no mechanics.

3.11 The sub-boss tier has no phase/count/threading rules beyond the ruling

BUILD L12 says "2-4 sub-bosses across different threads → one main boss," but there is no spec for how a sub-boss differs mechanically from a main boss (fewer phases? single-axis-allowed? no character-engagement requirement?). Since Pillar 8's "no single-mode" test is stated for *every* boss (CVD L218), it is unclear whether sub-bosses must also be fully blended — which, if enforced, is 3-5× the Pillar-8 authoring load per chapter that no doc has costed.

3.12 The Hades win-AND-death reward has no reward-economy spec

"Defeat AND victory both advance something" is doctrine (BUILD L21; ROSTER L333; canon-seeded by the graduated-consequence failure states at COMBAT L343-347). But outside the environmental-boss protocol (the only place canon actually implements it), there is no reward-on-death economy for conventional bosses — the proposal offers journal observations and "remembered tells" (BOSSJSON rows) but explicitly disclaims "no XP-on-death economy" (ROSTER L333) without defining what the positive economy IS.

---

One-paragraph synthesis for the design pass

The doctrine is strong and largely immutable: Pillar 8 blended-never-single-mode, the seven modes with the no-consecutive-dominant + 3-per-5-window rules, the 4-step environmental-boss protocol with graduated-consequence failure, the Entu three-phase template, the five Tier-A anchors (Ch 8/23/38/55/65, the Ch-58 "fifth" deleted), the integrity-mirrored damage formula, the ±1σ hard line, the unified twelve-tier quest-gated mastery, and — freshly RULED — the vril-density difficulty dial that is explicitly not HP inflation (Wellspring/Veiled/Fading/Sundered/→The Forgotten One). The per-chapter build target (2-4 thread-woven sub-bosses → one main boss, primary+secondary verb, three-axis split, evolve-not-swap phases, ≥2 counters, Hades reward on win and death) is a RULED work order. What is missing is everything mechanical and structural: all numbers (Phase 5M never ran), a boss registry that has 0 rows and a schema built for MMO World Bosses rather than the 75 solo rows the proposal needs, no canonical win-condition taxonomy, no definition of the mandatory "character engagement" field, no phase-transition mechanics, an unreconciled seven-vs-ten-vs-five mode vocabulary, unspecified familiar and vril-projection modes, and several unresolved protected-chapter source contradictions. The 75-boss roster, 13-school roster, sub-boss tier, and difficulty assist layer are proposals/rulings, not populated canon — they must be ratified and given a schema before they bind. The design should ENHANCE this baseline by filling the mechanical gaps (Part 3) and systematizing the implied layers (Part 2) without contradicting the doctrine and rulings in Part 1.

Key file paths:

• C:\dev\humanity-forgotten\_source\01_Tier_1_Foundation\T1_Combat_System_Spec [ACTIVE v1.0].md
• C:\dev\humanity-forgotten\_source\01_Tier_1_Foundation\T1_CVD_Creative_Vision_Document [ACTIVE v1.4].md (Pillar 8 L214-219; §12 L602-686; §12.2 seven modes L654-662; Pillar 11/12 L235-247)
• C:\dev\humanity-forgotten\docs\proposals\ROSTER_PROPOSALS.md and boss_roster_rows.json (75 rows) and school_roster_rows.json (13 rows)
• C:\dev\humanity-forgotten\docs\proposals\DIFFICULTY_SYSTEM.md (RULED) and CHAPTER_BUILD_TARGET.md (rulings 5-7)
• C:\dev\humanity-forgotten\registries\T0_Boss_Encounter_Registry [DRAFT v0.1]\Sheet1.csv (empty stub) and _source\00_Tier_0_Master_Indices\T0_Schema_Dictionary [ACTIVE v1.3].md §4.1 L844-873 (Boss schema) + L144 (mode_variety_check)

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