The sim → improve → sim loop — one game at a time

The owner's brief, 2026-08-01: 「Primary: improve our world to a level that every game on the shelf can be played by human users without obvious bugs. Secondary: make sure our harness is sound… and we are using the most cost-effective LLMs for each role. Greatly improve the kits first: search the web for authoritative rules → internalize → write the kit for the GM — the GM is the reader. Then simulate with 1..N LLM users, analyze from the user's POV and the transcripts, fix, loop until no more issues. One game at a time, starting with 20Q.」 This page is the plan, for review before execution, and will become the scoreboard as cycles run.

1 · The goals, ranked

2 · The loop

3 · Known entering defects — the loop starts mid-story

The 20Q cycle will meet the first of these immediately — the loop's first fix stage is expected to build the row-lag cue against a reproducing sim, which is the right order: the defect reproduces on demand before the fix is written, and the same sim proves the fix.

4 · The model economy — the secondary goal, made a table

Each cycle's report prints the room's per-role spend and wall time, so the economy is measured continuously rather than audited occasionally. The FP/prop cache-shape pass (the proven −38% playbook applied to the two remaining volatile-prefix calls) rides whichever cycle first touches those prompts.

5 · The games, in order — the scoreboard

6 · 20Q · cycle 1 — the concrete first execution

1. Rules research. Authoritative sources for Twenty Questions (the classic parlor rules; the animal/vegetable/mineral opening taxonomy; the handling of ambiguous answers; multiplayer turn conventions; the guess-cost rule) — cross-checked, forks noted as DIALS.
2. The kit rewrite, GM-as-reader. The current kit is a good outline; the rewrite makes every step mechanical and every obligation single-homed: the answer ladder (是/不是/说不清) with worked ambiguity examples; the board row as PART of the answer (kept), plus what to do when rows are found missing (backfill, from the record); the count as the board's business with the exact tick pairing; the endgame's two doors (right guess · the twentieth) each as a numbered sequence ending in the reveal; multiplayer turn order as a declared DIAL. Length is not the goal — reachability is: each rule sits in the section the GM reads at the moment it applies (the dilution law governs placement).
3. Sim configs. N=1 (the solo guesser — the commonest real shape), N=2, N=4 (turn order + the free-for-all DIAL). The troupe's 20Q script grows: patient watch (done) · per-question board-row assertion (done) · a deliberate wrong guess mid-game (costs one question, game continues) · the twentieth-question exhaustion path in one config · the reveal-before-verdict order checked from the event ledger.
4. Analysis. The four tests + 20Q-specific checks: count honesty (已问/剩余 vs the actual question count), answer consistency (rulings vs the sealed word — the referee reads the record), the row law, the reveal ritual. The flash judge's player-POV read on every run; flags human-verified.
5. Expected first fixes. The row-lag cue (world) is the anticipated cycle-1 build — reproduced first, then built, then proven by the same sim. Harness soundness items as they surface (exception counts printed per run).
6. Exit. Two consecutive clean runs at N=1 and N=4 → the scoreboard row flips green → 卧底.