AI Directors and Drama Management, v3¶

Third independent pass on the same ground as ai-directors-and-drama-management.md and ai-directors-and-drama-management-v2.md. Companion to 21-skyrim-quest-injection-machinery-v3.md, which covers this same report's Skyrim-specific half. This pass is the most quantitatively rigorous of the three on the LLM-narrator-drift question — it cites specific 2024–2026 papers with hard adherence/ conflict-rate numbers rather than general claims — and adds one entirely new, load-bearing lesson (Sid Meier's rubber-banding) absent from both prior passes.

What's genuinely new here (not in the first two passes)¶

[RISK — new, load-bearing] Perceived fairness is not the same thing as statistical fairness — Sid Meier's own account of Civilization's combat-odds backlash. Meier recounts that players couldn't accept losing a 3-to-1 battle roughly a quarter of the time, "past certain odds, people expected to win no matter what" — so Civilization Revolution deliberately tweaked the underlying math to make the player win more often than the "fair" odds would produce. Conversely, Civ 3's fair, fixed-random-seed combat resolution made players feel cheated when a reloaded battle replayed identically — a genuinely fair system read as rigged. Direct implication for any future Chronicle GM/director layer: budget for tuning to player expectation, not just to correct probability. A grudge-driven reprisal that is statistically well-calibrated against tracked belief intensity can still read as unfair if it violates what the player expected going in — this is a distinct failure mode from the "invisible decisive variable" problem already filed in the first pass, and needs its own tuning pass, not just legibility.
[BUILD-ON] Storylet casting is named explicitly as "your highest-value transfer" — sharper actionable framing than the first pass's generic KoDP/Wildermyth coverage. The concrete instruction: treat every GM intervention as a King-of-Dragon-Pass-style storylet — a precondition-gated situation with role slots — and cast the actual sifted entities (the real grudge-holder, the real target, the real faction) into those roles. This is what makes generated content read as authored, and it's the mechanism that directly fixes Radiant hollowness (the same "provenance for free" thesis from reports 15/17/19, now stated as a concrete storylet-role-casting recipe rather than a general principle).
[BUILD-ON] Robertson & Young's "Finding Schrödinger's Gun" gives a formal answer to which kind of intervention is more believable when a player's action threatens a planned arc. Revising the world's possibilities (retroactively establishing that a hidden option existed) is empirically more believable than overriding or silently failing the player's choice. This sharpens the Mimesis/mediation finding already filed in the second pass — it's not just "accommodate over intervene," it's specifically "if you must reframe, reframe the world's prior state, never the player's just-taken action."
[RISK — quantified, new] Hard numbers on LLM-narrator drift from a named 2026 benchmark, sharper than anything filed so far. NCP-Bench (Ma et al., arXiv 2608.08160) frames interactive narrative as a long-horizon constraint-satisfaction problem (Invariant/Ordering/ Achievement constraints) and stress-tests narrators with an adversarial player. Its results, verbatim: "GPT-5.2 has the highest average interaction length (32.92 turns), but its survival rate is only 42% after 20 turns. Fact conflicts are the most frequent failure mode, occurring in 40%–68% of runs across models." Across all models tested, fewer than 14% of narrative commitments are satisfied, and only ~3.5% of interactions reach 100 turns without an explicit conflict. Root cause named precisely: "LLMs lack explicit state-tracking mechanisms; they must implicitly reconstruct world state from the dialogue history at each turn, leading to drift," and models "yield rather than maintain established facts" under adversarial pressure. This is the single most citable piece of evidence yet filed for why Chronicle's symbolic-sim-as-truth architecture is necessary rather than merely prudent, should an LLM-narrated GM layer ever be built.
[BUILD-ON] Three named 2024–2026 systems solving exactly this problem, each a closer analog to Chronicle's own architecture than anything filed before:
Orchestrated Reality / WorldLines (Huang, Li, Fang; arXiv 2606.16014) formalizes an LLM game world as a Parameterized-Action POMDP where state is "a tree of canonical JSON entities" owned by a single orchestrating agent; the LLM only observes a narrative projection and proposes changes through a Plan–Diff–Validate–Apply pipeline that "commits schema-validated, content-hashed JSON deltas." Its own diagnosis of the field: current systems suffer "unvalidated writes: the model can assert any world change, so state silently drifts," whereas a validation guard (schema ∧ permission ∧ rule) "prevents prose from silently mutating state." Explicitly work-in-progress — even its own authors note un-schematized details still drift (an NPC's missing finger "swapping between left and right hand across a 24-turn session").
Neuro-symbolic TSL automata (Rothkopf, Zeng & Santolucito; arXiv 2402.16905, AAAI 2025) synthesize a correct-by-construction automaton from Temporal Stream Logic that decides which prompt modifier the LLM receives each turn, so long-horizon constraints are enforced by the automaton, not the model — "the automaton serves as the memory of the generative agent rather than the LLM itself." Hard number: "our approach using TSL achieves at least 96% adherence, whereas the pure LLM-based approach demonstrates as low as 14.67% adherence" (75 games × 20 turns), with the neuro-symbolic agent making zero arithmetic errors vs. 33.67% for the pure LLM.
Slice of Life (Treanor, Samuel & Nelson; FDG 2024/2025) is the closest published analog to Chronicle's own committed design: it pairs the Ensemble/ESP social simulation (a deterministic, symbolically-represented "social record"; character actions chosen by summing weighted first-order-logic social considerations) with an LLM used only for surface-text realization. Its stated principle, near- verbatim Chronicle's own doctrine: "keeping the simulation state deterministic and free from influence by the LLM in order to maintain authorial control," and "the dialogue does not feed back into the symbolic simulation state." Its cautionary finding matters too: an over-engineered prompt granting the LLM more latitude reintroduced hallucination (a character "invented that it is Tuesday") — a reminder that the validation boundary must stay tight even when it seems safe to loosen.
Function-calling as a hard validity gate (Gallotta, Liapis & Yannakakis, IEEE CoG 2024, system LLMaker): the LLM emits function calls into a constraint-enforcing back-end rather than free-form artifacts. Result: function calling never produced an invalid (parser/domain-failing) output, while baseline free-form prompting "never complete[d] a test case without failing." This is the concrete mechanism (not just the principle) for letting a future Chronicle GM act on the sim — spawn a quest, trigger a faction move — without the LLM layer ever being capable of writing an illegal world state.
[DESIGN-INPUT] A concrete four-stage build sequence for a future GM layer, useful as a starting draft if that layer is ever scoped. (1) Build the intervention layer as prerequisite-gated, entity-cast storylets with no LLM at all — benchmark: playtesters can unprompted explain why an event happened. (2) Add one legible, decaying pacing signal with published self-limits (what the director will never do) — benchmark/red-flag: if players discover a suppression exploit or report "the world feels out to get me," the causal anchor is too hidden. (3) Introduce an LLM GM as a one-way, validated author only — reads sifted state, writes surface text, never mutates belief/grudge/faction state except through a schema∧permission∧rule validator; benchmark: run an NCP-Bench-style adversarial stress test on your own content and count fact-conflicts per 20 turns. (4) Only after 1–3 feel solid, add PaSSAGE-style player-type weighting, driven by observed behavior rather than a hidden inferred vector.

Not repeated here¶

Façade's beat manager, the four-stage universal pipeline, Left 4 Dead's pacing FSM and RimWorld's wealth-exploit, Shadows of Doubt's provenance-anchored case generator, PaSSAGE's segment-limited results, DODM's Anchorhead transfer failure, Mimesis/mediation's accommodate-over-intervene doctrine, Daggerfall's Template v1.11 format, and AI Dungeon/Hidden Door's consumer-record failures are already filed across the first and second passes and substantially overlap this report's coverage of the same ground — not re-filed here.

Caveats¶

The 2026 papers (NCP-Bench, Orchestrated Reality) cite near-future model names (GPT-5.2, DeepSeek-V3.2) consistent with their stated dates; treat the architectural claims (validated writes, constraint-satisfaction framing) as robust regardless of which model generation is tested, but treat the specific percentages as tied to those model generations, not as durable constants.
Orchestrated Reality is explicitly work-in-progress with no reported effect sizes — cite its architecture, not results.
RimWorld's raid-cadence and wealth/combat-point figures are drawn from the community RimWorld Wiki, not a developer spec — well-corroborated but unofficial (same caveat as the first pass, restated here since this report repeats those figures independently).