run_encounter gains llm_policy and per-run policy_factory (mutually
exclusive with policy/llm_policy); replay.py logs every (observation,
action) keyed by call_seq to JSONL and replay() reproduces the recorded
battle bit-identically (wins/draws/rounds/transcript) with mismatch,
exhausted, duplicate and out-of-order guards. P0 frontier complete: the
recorded run replays identically.
Compact render is 15,691 bytes (~3.9k tokens) = 24.4% of the raw 64,204-byte
to_dict() snapshot on the representative 7-combatant encounter, hitting the
~2-4k token target; grids are sparse and determinism is pinned in tests.
candidates() derives plausible actions from the pairs with no second legality
engine (swift/free/full_round excluded).
Measures: observation 64 KB (~16k tokens) for 7 combatants / 22 KB for a 2v2;
124 policy calls per battle, i.e. 124k calls per 1000-run sweep cell -> a
per-turn LLM in a sweep is infeasible (D9 validated) and a compact view is
mandatory. Danger map (default vs focus-fire monsters, 100 runs): the default
policy understates danger (PC deaths 13%->24% and 0%->19%, TPK 0%->2%) and the
win rate alone misleads (rises to 98% while deaths rise) -> D12/D13 validated.
Notes the spike limits and the remaining P0.2-P0.5 items.
All recommendations accepted by the user except prompt location: prompts live
outside this repo (separate pack, configurable path, versioned; prompt_version
feeds the cache key and the replay record; a pinned pack for integration, a
minimal local fixture for unit tests, no silent fallback). Also records the
repo-scope decision (code only: calculation + LLM + report) with the final
detachment of data/sheets/prompts as a backlog item and the design constraint
to avoid hardcoded data paths in the new code. Priority: cartography first.
The execution tiers answer 'how to run a tactic', not 'which tactic is best'
nor 'where is the danger'. Adds two decision layers: an optimizer over
StrategySpec (objectives incl. tail risk, coarse-to-fine search, LLM as
candidate seeder) and an adversarial cartography (AdversarySpec + alternating
min-max + scenario sweeps + danger frontier + sensitivity), giving the danger
view for BOTH sides. Reframes the LLM as hypothesis generator / red team /
explicator / reporter, never the optimizer or per-turn executor in sweeps.
Adds D12-D14 and phases P3 (optimizer) / P4 (robustness).
Key correction: the LLM must not sit on the per-turn hot path of a Monte
Carlo sweep (runs x rounds x combatants calls, no cross-run cache reuse).
Adopts a two-tier design - a compiled, parameterised deterministic Policy for
quantitative sweeps (tier A) and a per-turn NLPolicy for a handful of showcase
battles (tier B). Also fixes: 'engine unchanged' -> rules unchanged/additive
boundary; replay keyed by a monotonic call_seq (delay re-queues, summons);
composite response-cache key incl. model id; ObservationView as a structured
object; candidates derived from pairs instead of a second legality engine;
per-run PolicyFactory; cost guardrails in P1; concurrency; D1 to settle in P0.
Design-only document (no code): target architecture (compact versioned
observation view, legal-actions API, LLM client + cache, NL policy with
parse/validate/repair/fallback, strategy sweep, replay/eval harness), the
concrete gaps in the current runner/CLI wiring, the decisions to settle and a
phased implementation plan (P0 foundations, P1 client+policy, P2 strategy
sweep, P3 provider+report) with exit criteria and risks.
ctan