Slice 2 — Arc-Layer (Spec 1, content + scope design)

Date: 2026-06-28 Status: Brainstorm-design, awaiting user review → writing-plan. Parent: docs/specs/2026-06-15-story-arcs-corruption-design.md (the ratified Spec-1 architecture — §1 Arc-Templates, shared types, pickArc/arcAllows semantics, selection lifecycle, ending evaluation). This doc does NOT re-design the architecture; it pins the concrete creative content + the precise Slice-2 scope boundary that the parent left open. Plan: docs/specs/2026-06-15-story-arcs-corruption-plan.md (Slice 2 row + anchors).

---

What this slice is

The arc architecture (types, pickArc, selection, ending resolution) is ratified. Slices 0+1 were infrastructure (tests, RNG parity, JSON SSOT) and shipped no user-visible feature. Slice 2 is the first payoff: the story actually branches — different pacing, mentor behaviour, beat flavour and ending across shifts.

Content ambition (ratified this session): "lean + tagged". Variability comes from routing + pacing + mentor presence + ending + tag-filtering of the EXISTING beat pool — no new antagonist prose block. Distinctive per-antagonist beat prose is a later content pass.

Scope

IN Slice 2: - arc.ts twin: ArcTemplate / ArcState / EndingSpec types, arcAllows(), pickArc(), selectArc(), resolveEnding(). - pickArc(POOL, this.arc) swept across the ~20 narrative pick sites (mechanical; pick() itself untouched). - Arc selection in the silentReset callback (weight + in-memory ring buffer N=3; cold-start = pure weights; first shift deterministic-from-weights for test pinning). - phaseRouting (acyclic forward only) + per-phase durationScale. - Mentor override per arc (gate beatMentor on arc.mentor.availablePhases). - threat→stage quantization (small pure function) for peakStage / divertOnThreat. - Ending resolution at SILENCE (peakStage → first matching divertOnThreat, else default); the resolved EndingId selects the SILENCE farewells/lastWords set. - Beat tagging of a subset of existing beats + arc-specific farewells/lastWords per ending (6 ending sets; normal = today's content unchanged). - Additive invariant: arc=undefined ⇒ golden narrative byte-identical (web + native twin).

OUT (deliberately deferred): - Corruption render layer → Slice 4 (only the threat→stage quantization lands here, not the visual render). - Arc scenes FilmDirector bias → Slice 7; threatCurve → Slice 8. The ArcTemplate type carries these optional fields but Slice-2 templates omit them (no unconsumed data). - Persistence of arcsCompleted/endingsReached/layerCounter → Slice 3 (ring buffer is in-memory until then; cold-start path already handles empty history). - "False all-clear" backward routing (e.g. ALARM→ROUTINE then re-escalate) — needs a once-guard to avoid loops; deferred. Slice 2 does acyclic forward routing only, so the additive invariant holds trivially.

Web/native split (mirrors Slice 1's 1a/1b/1.3 rhythm): - Slice 2 web — now, autonomous: arcs authored in story-content.json, web reads them, web tests (golden, arcAllows truth table, ending determinism). - Slice 2 native — later, bundled with native 1.3 in an on-device session: native reads the same arcs from the JSON (no double-authoring). The cross-platform LCG arc-parity tests run once native lands.

---

The 5 arcs (selectable) + captured (divert-only)

Selectable arcs[] = normal, harmonized, cold-path, karsen, wraith. captured is not a template — it exists only as an EndingId with its own farewells/lastWords, reached via divertOnThreat.

Arc	weight	antagonist	mentor (Karsen) availablePhases	routing / pacing	ending.default	divertOnThreat
normal	1	baseline	ROUTINE,INTRUSION,ALARM,PANIC	linear, ×1	normal	—
harmonized	3	drift	ROUTINE,INTRUSION,ALARM,PANIC	ROUTINE ×1.3, PANIC ×0.7	harmonized	stage 3 → cold-path
cold-path	3	trace	ROUTINE,INTRUSION	skip-blitz INTRUSION→PANIC; INTRUSION ×0.8, PANIC ×0.9	cold-path	stage 3 → captured
karsen	3	injection	ROUTINE,INTRUSION,ALARM,PANIC	ALARM ×1.3, PANIC ×1.3	karsen	stage 3 → captured
wraith	2	wraith	ROUTINE	ROUTINE ×0.5, PANIC ×1.4	wraith	—

Total weight 12 → normal ≈ 8 % (rare), each dramatic arc 17–25 %.

As ArcTemplate[] (the arcs section of story-content.json):

[
  { "id": "normal", "weight": 1, "antagonist": "baseline",
    "mentor": { "name": "INSTR-KARSEN", "availablePhases": ["ROUTINE","INTRUSION","ALARM","PANIC"] },
    "beatTags": {},
    "ending": { "default": "normal" } },

  { "id": "harmonized", "weight": 3, "antagonist": "drift",
    "mentor": { "name": "INSTR-KARSEN", "availablePhases": ["ROUTINE","INTRUSION","ALARM","PANIC"] },
    "durationScale": { "ROUTINE": 1.3, "PANIC": 0.7 },
    "beatTags": { "include": ["drift","calm","technical"], "exclude": ["wraith","aggressive"] },
    "ending": { "default": "harmonized", "divertOnThreat": [ { "atStage": 3, "to": "cold-path" } ] } },

  { "id": "cold-path", "weight": 3, "antagonist": "trace",
    "mentor": { "name": "INSTR-KARSEN", "availablePhases": ["ROUTINE","INTRUSION"] },
    "phaseRouting": { "INTRUSION": "PANIC" },
    "durationScale": { "INTRUSION": 0.8, "PANIC": 0.9 },
    "beatTags": { "include": ["trace","technical"], "exclude": ["calm"] },
    "ending": { "default": "cold-path", "divertOnThreat": [ { "atStage": 3, "to": "captured" } ] } },

  { "id": "karsen", "weight": 3, "antagonist": "injection",
    "mentor": { "name": "INSTR-KARSEN", "availablePhases": ["ROUTINE","INTRUSION","ALARM","PANIC"] },
    "durationScale": { "ALARM": 1.3, "PANIC": 1.3 },
    "beatTags": { "include": ["injection","technical"], "exclude": ["wraith"] },
    "ending": { "default": "karsen", "divertOnThreat": [ { "atStage": 3, "to": "captured" } ] } },

  { "id": "wraith", "weight": 2, "antagonist": "wraith",
    "mentor": { "name": "INSTR-KARSEN", "availablePhases": ["ROUTINE"] },
    "durationScale": { "ROUTINE": 0.5, "PANIC": 1.4 },
    "beatTags": { "include": ["wraith","aggressive"], "exclude": ["calm","drift"] },
    "ending": { "default": "wraith" } }
]

Differentiation logic: normal ≈ today (rare). harmonized lingers calm then resolves; spiking to peak stage 3 means harmonization failed → cold-path. cold-path skips the ALARM deliberation (decisive sever), Karsen drops after INTRUSION (channel goes cold); too slow → captured. karsen keeps the mentor present throughout with extra ALARM/PANIC dialogue room. wraith compresses ROUTINE (turns hostile fast), cuts the mentor after ROUTINE, and drags out a hostile PANIC (where Slice 4 corruption will peak).

---

Tagging

Vocabulary (theme + tone), referenced by beatTags: - theme: drift, trace, injection, wraith - tone: calm, aggressive, technical

Filter semantics (from parent §1, restated for the implementer): - beat with no tags → always eligible (back-compat; this is why most beats stay untagged). - beat with tags → eligible iff (no include set or shares ≥1 include tag) and shares 0 exclude tags. Exclude wins on conflict. - pickArc with arc=undefined → pick(pool), tags never consulted (invariant). - Empty filter result → fall back to full pool (never starve).

Tagging plan (lean): tag a subset of INTRUSION/ALARM/PANIC beats by tone/theme so each arc draws a different colour from the existing pool; ROUTINE stays largely untagged (similar across arcs). Rough split to assign during implementation: - aggressive/hostile intrusions+alarms → aggressive (a few also wraith). - spec-drift / impersonal-anomaly beats → drift + calm + technical. - tracking/trace-themed beats → trace. - signature/injection-themed beats (canon: "i think this is signed", "module 14") → injection. - calm/technical status-y beats → calm/technical.

Guard re-pin: adding the optional tags field changes story.beats.*, so the tests/script-bank-parity.test.ts hash is intentionally re-pinned in Slice 2 (reviewed change). The golden narrative test stays green (tags are inert when arc=undefined), which is what proves behaviour preservation.

---

Ending content (6 sets, keyed by EndingId)

Stored under story.endings = { farewells: Record<EndingId,string[]>, lastWords: Record<EndingId,{typed,abandonAt}[]> }. farewells render as QUOTES lines; lastWords.typed is typed at the prompt and left hanging (the string itself is the cut-off half-thought; abandonAt stays 0, matching today's convention — web beatSilence types the whole typed and holds, so abandonAt is currently inert on web; kept for the data convention / native parity).

normal — unchanged (today's content), preserved for the additive invariant: - farewells = the current 6 (> you should have used the cold path. … > goodnight, officer.) - lastWords = the current 5 (i was just running through the final wait someone's at the door…, etc.)

Note: normal keeps today's "someone's at the door" lastWords as-is (preservation); the captured ending below gets its own thematically-overlapping but distinct set.

harmonized (the anomaly folds back in; calm closure):

"farewells": [
  "> harmonization complete. the strings fold back in.",
  "> the work is the answer.",
  "> nothing followed me home tonight.",
  "> quiet again. the way it should be.",
  "> [SHIFT CLOSED // NOMINAL]"
],
"lastWords": [
  { "typed": "logging it clean. see you next rota", "abandonAt": 0 },
  { "typed": "all folded back. i think we're go", "abandonAt": 0 },
  { "typed": "going to sleep well for on", "abandonAt": 0 }
]

cold-path (severed the link, went dark to survive):

"farewells": [
  "> you should have used the cold path.",
  "> channel severed. i'm a ghost now.",
  "> better cold than traced.",
  "> they can't read a line that isn't there.",
  "> [LINK DARK // NO CARRIER]"
],
"lastWords": [
  { "typed": "pulling the jack now don't try to recon", "abandonAt": 0 },
  { "typed": "going dark. if you're reading this i'm already g", "abandonAt": 0 },
  { "typed": "cold path. should have taken it ho", "abandonAt": 0 }
]

karsen (mentor intervention holds the line):

"farewells": [
  "> stay on the line. i've got you. — INSTR-KARSEN",
  "> module 14. he was right.",
  "> karsen pulled the plug in time.",
  "> not alone after all.",
  "> [GHOSTLINK HELD // OPERATOR RECOVERED]"
],
"lastWords": [
  { "typed": "karsen are you still th", "abandonAt": 0 },
  { "typed": "okay. okay. doing what you sa", "abandonAt": 0 },
  { "typed": "tell me it's going to b", "abandonAt": 0 }
]

wraith (antagonist takeover; identity stolen):

"farewells": [
  "> it answers to my name now.",
  "> the wraith wears my designation.",
  "> i was never the operator.",
  "> whatever types next is not me.",
  "> [CONTROL TRANSFERRED // SESSION HIJACKED]"
],
"lastWords": [
  { "typed": "that's not what i typed who is wri", "abandonAt": 0 },
  { "typed": "my hands aren't on the k", "abandonAt": 0 },
  { "typed": "it's using my designat", "abandonAt": 0 }
]

captured (taken through the door, not the wire; system-side close):

"farewells": [
  "> they came through the door, not the wire.",
  "> operator did not return from break.",
  "> [SEAT VACATED // NO SIGNOUT]",
  "> [SESSION ABANDONED // CHAIR EMPTY]"
],
"lastWords": [
  { "typed": "wait someone's at the do", "abandonAt": 0 },
  { "typed": "who are y", "abandonAt": 0 },
  { "typed": "no no i didn't", "abandonAt": 0 },
  { "typed": "how did you get in h", "abandonAt": 0 }
]

---

Selection, routing-safety, quantization

• Selection (selectArc): weight each arcs[] candidate by weight; dampen any id in the in-memory arcsCompleted ring buffer (N=3) so the same arc can't repeat within 3 shifts; pick uniformly among survivors. Empty history → pure weights. First-ever shift → deterministic-from-weights (test can pin a known arc), or pass arc=undefined to assert identity. Runs inside silentReset after blackout + persona regen, before enterPhase('ROUTINE'). ArcState survives silentReset; peakStage resets to 0 at shift start. captured is excluded from selection.
• Routing safety: Slice-2 phaseRouting values are acyclic forward only (skip a phase, as cold-path's INTRUSION→PANIC). No backward routing → no loop guard needed.
• threat→stage quantization (shared pure fn, twin-tested): threat<0.15→0, <0.40→1, <0.70→2, else 3 (parent §3 table), with a calm flag that forces 0. Per parent §3 this is one quantity — corruptionStage = ArcState.threatStage — so Slice 2 and the Slice-4 render share the single helper. peakStage = max(threatStage) over the shift. Decision (resolves the calm ambiguity): because calm/reduced-motion forces threatStage to 0, peakStage stays 0 in reduced-motion → divertOnThreat never fires → reduced-motion always reaches ending.default. This is a deliberate, parent-consistent a11y flattening: the story still varies by arc selection / pacing / mentor, only the threat-spike diversions are suppressed. Slice 2 consumes threatStage only for peakStage/divert; the visual render is Slice 4.
• Ending resolution (resolveEnding): at PANIC→SILENCE, the first divertOnThreat rule with atStage ≤ peakStage wins, else ending.default (single-step on the selected arc's spec — a diverted ending does not re-evaluate its own diverts). Resolved EndingId → that ending's farewells/lastWords set. arc=undefined ⇒ the normal ending set (= today's pools), which is what keeps the additive-invariant golden byte-identical. Recording to endingsReached is Slice 3 (persistence).

---

Testing (Slice-2 subset of parent §6)

1. Additive-invariant golden — same seed, arc=undefined, corruption off ⇒ beat sequence / timing / threat curve / persona byte-identical to today (the existing golden tests, kept green). Web now; native when Slice-2-native lands.
2. arcAllows truth table — include/exclude/empty/untagged matrix, identical both sides.
3. Ending determinism — (arc, peakStage) → EndingId deterministic; divertOnThreat covered at several peak timings (incl. stage-3 diverts to cold-path/captured).
4. Arc selection + routing parity (LCG) — pinned seed, frozen ring buffer ⇒ identical arc choice, pickArc filtering, phase routing. Cross-platform test runs when native lands.
5. script-bank-parity re-pin — re-pinned after tagging; reviewed as an intentional change.

---

Decisions log (2026-06-28)

1. Next step after Slice 1 → Slice 2 web (first user-visible payoff; autonomous).
2. Content ambition → lean + tagged (mechanism + templates + endings + tag existing beats; no new antagonist prose this slice).
3. 5-arc bible (weights, antagonists, mentor availability, routing/pacing, endings, diverts) → approved as the table above.
4. Tag vocabulary (drift/trace/injection/wraith + calm/aggressive/technical), untagged = universal, guard re-pinned after tagging → approved.
5. Ending content (6 sets; normal = today's, 5 new in canon voice) → approved (full sets above).
6. Scope boundary: corruption render (4), scenes (7), threatCurve (8), persistence (3), false-all-clear backward routing → all OUT of Slice 2; web/native split at the 1.3 boundary.