AI Directors and Drama Management

Single-source survey of drama-management and AI-director architectures across academic interactive-narrative research (Façade, narrative planning, Experience Management, Concordia) and shipped commercial systems (Left 4 Dead's Director, RimWorld's Storytellers, Shadows of Doubt's case generator, PaSSAGE). Filed here rather than in the main numbered docs/research/ series because — like the Crusader Kings and real-time-spatial-sim files already in this folder — it's cross-game comparative material informing Chronicle's reactivity/GM design, not Skyrim-substrate research tied to an accepted ADR. Its companion piece, 19-skyrim-quest-injection-machinery.md, covers the Skyrim-specific mechanics for actually wiring a GM layer into the game; this file covers the design patterns for what that layer should decide, independent of engine.

This is ahead-of-need research: docs/architecture.md's Build order lists "prospective drama management" under Defer — no GM/director layer is in Chronicle's current build order. Some ground here already overlaps this folder's existing Crusader-Kings coverage (King of Dragon Pass and Wildermyth are already covered in depth in ck-postmortems-and-design-lessons.md and ck-failure-modes-and-reactivity-loop.md) and the spatial-sim files (Shadows of Doubt is already covered there from a witness/memory-decay angle) — not re-filed here where it overlaps; see "Not repeated here" below.

Findings

• [BUILD-ON] The universal four-stage director pipeline, cross- validated against report 19's independent Skyrim-specific framing. Every architecture surveyed — from Façade's Beat Manager to Concordia's LLM Game Master — decomposes into: (1) input signals (live state: player stress, colony wealth, NPC belief graphs, faction assets), (2) recognition/sifting (evaluative logic — pacing FSMs, Prolog constraint satisfaction, storylet matching — that isolates dramatically interesting configurations), (3) intervention selection (a concrete structural choice: adjust a spawn budget, draw a storylet, instantiate a quest template), (4) presentation/surfacing (multimodal render: UI, dialogue barks, journal entries). This is the same shape report 19 arrives at independently for Skyrim's IntelEngine — worth treating as a validated architecture for Chronicle's own eventual GM layer, not a framing coincidence.
• [DESIGN-INPUT] Façade's four documented failure modes are a direct checklist against any future beat/storylet-driven Chronicle layer. (1) Shallow character self-awareness — autonomous characters execute beat parameters without genuine internal state, i.e. they're reactive to the director, not to each other; Chronicle's belief/grudge graph is explicitly the fix for this, since NPCs would act from their own tracked beliefs rather than beat-assigned behavior. (2) Spatial/ environmental constraints — Façade's beat manager only worked in a tightly confined single room; expanding physical space "dramatically expands the state space" and breaks coordination — a direct warning against a monolithic beat-manager approach for an open, ~1,000-NPC Whiterun-scale world. (3) Absence of inciting incidents — pure reactivity to player input means players get dropped in without clear framing; any Chronicle GM layer needs an explicit "why does this matter now" framing step, not just reactive triggering. (4) NLP reductionism — nuanced input collapsed to a binary affinity score reads as jarring; relevant if Chronicle ever parses free-text player/NPC dialogue into scalar state changes (the same "format-valid ≠ content-valid" caution already flagged in 03-hybrid-llm-symbolic-architecture.md).
• [RISK] RimWorld's Storyteller wealth-scaling is a documented, exploitable failure mode directly relevant if Chronicle's economy tier (deferred to v0.4, see 16/18) ever feeds a pacing/intervention system. Threat Points scale with raw colony market value rather than defensive capability, so players learn to deliberately destroy or gift away surplus wealth to suppress raid intensity — "wealth management" as an exploit. The generalizable lesson: any numeric metric driving GM interventions must track a cause the player can't trivially game by manipulating the metric itself, distinct from the systemic reality the metric is meant to proxy. If a future Chronicle director ever uses, say, aggregate grudge-intensity or rumor-spread-volume as an intervention trigger, this is the specific exploit shape to design against.
• [BUILD-ON] Left 4 Dead's four-state pacing FSM (Build Up → Peak → Sustained Peak → Relax) is the cleanest shipped model of legible intensity pacing, and its fairness mechanism is explicit and transferable: state transitions are always multimodally signaled (ambient audio shifts, distinct musical stingers per threat type, character audio barks) before the intervention lands, so a spike in difficulty never reads as untelegraphed. Its one documented exploit — players lingering in "safe zones" to manipulate the stress counter — is a second instance of the same "player games the metric, not the underlying threat" failure class as RimWorld's wealth exploit.
• [DESIGN-INPUT] Shadows of Doubt's "Case Generator," read as provenance-anchored intervention rather than witness/memory mechanics (the angle spatial-sim-shadows-of-doubt-nemesis-kenshi.md already covers). A killer is selected by archetype (Corporate Cutthroat, Dating Service Stalker, Apartment Sniper, etc.), then physically executes the crime using in-engine systems, leaving evidence determined entirely by their actual simulated history and actions — a business card and graffiti for the Corporate Cutthroat (targeting a real career rival from the simulated employment graph), lipstick and a note for the Dating Service Stalker (from the simulated in-world dating app), bullet trajectories from an apartment the sniper actually occupies. No clue is ever spawned from a template disconnected from simulation state — evidence has "complete physical provenance." This is the load-bearing precedent for report 19's "provenance for free" quest-blandness fix and for Chronicle's own inspectability requirement (docs/decisions/0007-inspectability.md): a GM intervention should be generated from, and fully traceable back to, actual tracked simulation state — never an independently-rolled event merely decorated with simulation-flavored text after the fact.
• [BUILD-ON] PaSSAGE (Thue et al., Neverwinter Nights) is the concrete prior-art implementation of player-modeling-driven content selection, useful if Chronicle's GM layer ever wants to vary what kind of intervention surfaces per player rather than just whether one fires. Player actions are pre-annotated with modifiers to a five-element style vector (Fighter/Power Gamer/Tactician/Storyteller/Method Actor, after Robin Laws' RPG archetypes); the next encounter is chosen by dot- product match between the player vector and each candidate encounter's annotation vector. Documented trade-offs, directly relevant to scoping: authorial annotation overhead (every branch/dialogue/item needs manual multi-dimensional annotation — multiplies content cost), vector drift (mechanically-motivated actions, like smashing crates for a key, get misread as playstyle signal), and pacing flattening (pure preference-adaptation creates a feedback loop that strips contrasting beats, e.g. no quiet conversational content ever reaches a player who tested as a Fighter). None of these are solved problems — file as known costs of adding player-modeling to a future GM layer, not as a recommended must-have.
• [BUILD-ON] DeepMind's Concordia is the most directly transferable academic pattern for an LLM-narrated, symbolically-grounded GM, and it's the same hybrid-neurosymbolic shape Chronicle's own architecture already commits to (docs/decisions/0003-substrate-choice.md, 03-hybrid-llm-symbolic-architecture.md). Its loop: agents emit natural-language intent strings; the GM intercepts and validates each intent against grounded variables (physical presence, inventory, spatial bounds) — if valid, it updates grounded state and synthesizes a natural-language "Event Statement" describing the outcome; if invalid, it tells the agent why the action failed; either way, the Event Statement is pushed into nearby agents' memory logs. The concrete, reusable idea: an LLM should never be allowed to directly mutate world state — it emits intent, a grounded validator (symbolic, deterministic) accepts/rejects and performs the actual mutation, and only the validator's outcome gets narrated back. This is a slightly more mechanized version of the same principle report 19 independently states for IntelEngine ("LLM restricted to narrative framing... symbolic state governs engine logic").
• [RISK] The "unconstrained generation breakdown" failure catalogue is a direct warning against ever letting an LLM freely narrate Chronicle world-state changes without a validation layer. Documented failure modes of autoregressive-LLM-only narrative systems: hallucinated goals/ items/characters that don't exist in game state (quest progression stalls), state amnesia (items vanish, dead characters reappear), and causal/world-logic collapse (physically impossible actions destroy mechanical agency). All three are avoided, per this report's own conclusion, only by the hybrid neurosymbolic pattern above — symbolic layer as absolute truth, LLM as reader/narrator/validated-API-caller only. Chronicle's belief-facet store already commits to this separation; this is corroborating literature, not a new design requirement.
• [DESIGN-INPUT] Symbolic narrative planning (IPOCL and successors) is academically important but not directly actionable at Chronicle's scale — filed for completeness, not as a build target. Classical planners (STRIPS/PDDL-style) treat story generation as search over character/environment actions from an initial state to an authored goal state; IPOCL's contribution was forcing every individual action to be motivated by that character's own goals (preventing "suicidal puppet" actions purely to satisfy the global plot). Its documented failure modes — combinatorial state explosion with added agents/locations, precondition rigidity breaking on unscripted emergent player actions, and heavy knowledge-engineering overhead — are exactly why storylet/precondition-filtering systems (King of Dragon Pass, Wildermyth — already covered in this folder's CK files) won out commercially over full symbolic planning, and why Chronicle's own design (sparse-graph, provenance-tracked events + a future precondition-filtering sifter) already tracks the commercially-proven branch of this lineage rather than the academically-elegant one.

Universal director pipeline table (from this report)

System	Input	Recognition	Intervention	Presentation
Façade	Discourse acts, spatial vectors	Tension level + zero-sum affinity score	Select next beat from authored library	Joint ABL dialogue + physical animation
Left 4 Dead Director	Damage, health, mobility, incapacitations	Intensity threshold + FSM pacing stage	Zombie population caps, item drop queues	Music cues, horde spawns, ambient barks
RimWorld Storytellers	Colony wealth, adaptation timer, colonist count	Threat-point budget + event cooldowns	Roll event category (raid/trade/eclipse/disease)	Warning envelope, alert sound, spawn vectors
Shadows of Doubt	NPC routines, phone logs, footprints	Killer-archetype trigger via relationship graph	Execute physical crime, drop physical clues	In-world crime-scene discovery, news reports
King of Dragon Pass	Faction state, magic, season, clan mood	Prolog constraint satisfaction	Priority-band filtering, storylet draw	Storylet modal + advisor commentary
PaSSAGE	Tracked player actions	Vector dot-product playstyle match	Select next encounter module	Instantiate encounter + customized dialogue
Concordia GM	Agent NL intent strings	Grounded variable validation	Update grounded variables, advance clock	Natural-language Event Statement to memory

Not repeated here

King of Dragon Pass's and Wildermyth's mechanics (storylet precondition-filtering, priority banding, character-pair bindings, legibility-through-visual-interface) are already covered in depth in ck-postmortems-and-design-lessons.md and ck-failure-modes-and-reactivity-loop.md; Shadows of Doubt's witness/memory-decay/provenance-graph detective mechanics are covered in spatial-sim-shadows-of-doubt-nemesis-kenshi.md and spatial-sim-legal-boundaries-and-witness-propagation.md. This file only adds the Case Generator's provenance-anchoring lesson (new framing on already-known ground) plus genuinely new material: Façade, formal narrative planning, Left 4 Dead, RimWorld, PaSSAGE, and Concordia.

Caveats

• Single-source report — no independent cross-check exists yet for this ground.
• Several citations (Façade failure modes, PaSSAGE trade-offs) are to a thesis/paper this session did not independently re-read; treat as this report's synthesis of that literature, not a primary-source verification.