AI Directors and Drama Management, v2

Second independent pass on the same ground as ai-directors-and-drama-management.md, from the general-field part ("Prompt A") of a combined report whose Skyrim-specific part is filed as 20-skyrim-quest-injection-machinery-v2.md. Same naming pattern as 17/18. This pass goes considerably deeper into the academic drama-management lineage (declarative optimization, plan-based mediation) and the LLM-era consumer record (AI Dungeon, Hidden Door) than the first pass — file separately rather than merge given how much is genuinely new.

What's genuinely new here (not in the first pass)

• [RISK] The Declarative Optimization Drama Management (DODM) failure record is the single most load-bearing new finding: the optimizer is the least important part of a drama manager. Nelson and Mateas re-applied Weyhrauch's search-based drama management (which had worked on a simplified Deadline) to Anchorhead and found the results did not transfer; switching to offline reinforcement learning didn't help either. Their conclusion: the available DM actions were "simply too weak to affect measured story quality" — an authorship problem, not an optimization problem. With a synthetic "maximally powerful" action set (a causer/denier/re-enabler for every plot point), even a weaker algorithm performed well. Direct implication for any future Chronicle GM: invest in the intervention vocabulary (courier notes naming a real employer, hired thugs traceable to a real grudge, faction mobilizations anchored to tracked resources) before investing in clever selection logic — a dumb picker over a rich, provenance- carrying action set beats a sophisticated picker over a thin one.
• [RISK] A director that optimizes a single quality metric can destroy player agency by forcing its own "best" story. Roberts et al.'s response — Targeted Trajectory Distribution MDPs — aims the director at an author-specified distribution of experiences rather than one optimal trajectory. Generalizes directly to RimWorld's Randy Random persona (bounded-but-undirected randomness is a legitimate design point, not a lesser one) and argues against ever tuning a future Chronicle GM purely to maximize one legibility/drama score.
• [BUILD-ON] Plan-based mediation (R. Michael Young's Mimesis lineage) supplies a formal license for retroactive reframing that matches Chronicle's own no-invisible-decisive-variables doctrine exactly. When a player does something the authored plan didn't anticipate, the mediator either intervenes (blocks/undoes the action) or accommodates (replans so the action is absorbed). The documented failure: repeated intervention makes players "realize their behavior is limited" — i.e. accommodate, don't block, whenever possible. The 2013 extension is the load-bearing piece: giving the mediator a model of character knowledge widens the accommodative search space by replanning past events, but explicitly "only as long as the planner maintains story-world consistency by restricting modifications of past events to aspects of the world unobserved by the user." This is the formal, citable version of "a rumor can retroactively reveal what really happened, exactly where the player couldn't have observed the truth already" — directly reusable if Chronicle's GM layer ever needs to justify a retcon-shaped reveal (the widow's grudge existed all along; the player just never saw the belief update happen).
• [BUILD-ON] Daggerfall's actual quest format is now precisely documented, correcting the vaguer "QRC/QBN" gloss in the first pass and in report 19/20. 227 shipped quests (241 with the CompUSA edition), gated by guild/social-class membership and reputation — the same "Domestic Squabble" template means something different at reputation rank 0 vs. rank 7. Donald Tipton's Template v1.11 decompiles the binary format into readable source ("more like a complex INI file"): resource declarations (Item, Person, Place, Clock, Foe) plus task blocks driven by startup/timer/condition triggers. Daggerfall Unity's open-source reimplementation runs this as a documented QuestMachine — quests instantiated from text source at runtime, parsed by regex-matched ActionTemplate factories, ticked at 10 Hz, serialized to JSON save state, with third-party-registerable actions (150+ community quest-pack quests exist in this format). The longevity lesson stated explicitly: Daggerfall's radiant content works because templates are embedded in a reputation economy, not because the templates themselves are clever — corroborates report 19's "provenance for free" thesis from the oldest possible source in the lineage.
• [BUILD-ON] The storylet academic thread names the closest published architecture to a sifter-fed GM. Breault et al. built a quest engine that generates quests from world state at generation time, rated by human judges as comparable to human-authored quests, with the significant property that "the creation of possible quests increases as the game progresses" — richer world state yields richer quest material. This is a direct academic endorsement of Chronicle's core bet (quests anchored to live simulation state scale with the simulation instead of exhausting a fixed template space), independent of any Skyrim-specific argument. Caveat from mixed-initiative studies (Questgram): designers gave generation and suggestion features a "mixed response," mostly over abstract actions and random placement — treat generative quest tooling as a creativity scaffold for human authors, not a fully autonomous author, at least as currently studied.
• [RISK] The AI Dungeon / Hidden Door consumer record is the sharpest cautionary evidence yet against unconstrained LLM narrative generation — new in this pass entirely. AI Dungeon's failures read as a checklist of what a symbolic substrate prevents: characters "forgot who they were after a few exchanges," NPCs changed gender mid-scene, a frozen vampire would bite the player again the very next turn, and players had to hand-maintain a "pin" memory that "felt like learning a programming language" — with memory drift remaining the category's #1 complaint as of 2026 reviews. Hidden Door (a more sophisticated attempt — structured plot beats, explicit setup/payoff data structures, a game-engine layer tracking every character/item/location) still failed on ungroundedness: "the story is written only as far as it's been presented to the player… the treasure chest doesn't have traps, and doesn't not have traps" — so choices retroactively change the present, not just the future, which "undermines autonomy" even when players can't articulate why. The reviewer's fix list is close to a ready-made Chronicle design checklist: commit to and track a world state, outline consequences before presenting choices, let failure consequences actually disrupt world state, and generate more material than the reader consumes. Also newly documented: turn-cycle latency kills pacing (90–120 seconds per cycle in one research director; 10–20 seconds with "thinking" disabled, at the cost of more hallucination), and Retail Mage's team had to deliberately make their AI dumber because NPCs that optimally self-organize "leave players with nothing to do" — a direct rediscovery of the old Sims-design lesson, worth flagging for Chronicle's own NPC-agency tuning once an LLM tier is added (currently deferred per docs/architecture.md's build order).
• [BUILD-ON] 2024–2026 research has landed almost exactly on Chronicle's own hybrid architecture — new, specific citations. Drama Llama (2025): an LLM-powered storylets framework where authors write only 3–4 natural-language "pivot point" triggers and the LLM improvises within them — explicitly positioned on the same authorial-intent vs. character-autonomy axis this literature keeps returning to. The Interview (prompt-engineering study of generative NPC dialogue) formalizes Symbolically Scaffolded Play: no single prompt design guarantees better play, over-constraining "actively diminishes improvisation," and the working pattern is role-differentiated scaffolding — rigid symbolic structure for quest-giver roles, loose structure for improvisational suspects, with fuzzy parameters persisted in a shared JSON memory schema. A 2026 AI-native-games survey's stated open problems — "persistent memory, world-state tracking, and long-term consequence management remain fragile," inference cost/latency must shape mechanic design, generated content must survive verification against rules — are named as still-open in the general literature, which the report explicitly notes Chronicle's symbolic sim "solves by being the substrate" rather than needing to solve at the LLM layer. Its closing claim is worth carrying forward precisely: nobody has yet shipped an LLM director whose interventions are verified against a live belief/grudge/faction simulation inside a commercial-scale RPG — that is the specific uncovered front Chronicle would occupy if it ever builds this layer.
• [DESIGN-INPUT] The seven documented ways directors break player trust, consolidated across the whole literature in this report's own synthesis — a more complete taxonomy than the first pass's trust-hazard list. (1) Visible railroading — repeated action-nullification. (2) Ungrounded generation — retroactively rewriting the world at presentation time. (3) Agency-destroying optimization — forcing one "best" trajectory. (4) Memory inconsistency — forgotten facts turn play into prompt maintenance. (5) Deceptive provenance — generated content disguised as authored content, or forced without consent. (6) Static-world exposure — sending players to just-cleared locations, spotlighting that the world didn't actually change. (7) Unsolvable/ unrewarding generated content — completion or reward paths that silently break. Each pairs with a named design countermeasure in the source (accommodate over block; sim-first framing; target distributions not optima; symbolic sim as sole truth; opt-in/legible generated content; track intervention history to avoid repeats; failsafe doctrines plus native completion paths) — worth lifting as a checklist wholesale if Chronicle ever drafts an ADR for a GM/director layer.

Not repeated here

Façade's beat-manager mechanics, the universal four-stage pipeline table, RimWorld's wealth-exploit and Left 4 Dead's pacing FSM, Shadows of Doubt's provenance-anchored case generator, PaSSAGE's player-modeling trade-offs, and DeepMind Concordia's grounded-validation loop are already filed in ai-directors-and-drama-management.md and substantially overlap this report's coverage of the same systems — not re-filed here. King of Dragon Pass and Wildermyth mechanics remain filed in this folder's Crusader Kings files, as noted there.

Caveats

• Single-source addition; citations to specific papers (Nelson & Mateas, the 2013 Mimesis extension, Drama Llama, The Interview) were not independently re-read by this session — treat as this report's synthesis, not primary-source verification.
• The 2026 AI-native-games survey and consumer-product claims (AI Dungeon/Hidden Door reception, turn-cycle latency figures) are point-in-time and will drift.