Practice AI replay verification — investigation

Goal

Replace trusted client fields (won: true) in reportPracticeAiMatch with server-derived outcomes from deterministic replay.

Current flow

local-game-page.tsx launches with { config, seed } (Date.now()).
bridge-table.tsx runs the match locally (applyAction + decideGameActionAsync).
On game.phase === 'complete', client calls reportLocalAiMatch({ won, skill, ... }).
Cloud Function reportPracticeAiMatch applies Elo — does not verify the match happened.

What we already capture (but do not send)

Seed — in launchSession.seed
Config — LocalGameConfig (objective, AI tiers, modules, house rules)
Action log — actionLogRef in bridge-table (human + AI + auto), included in debug export

Proposed server payload

interface ReportPracticeAiMatchV2 {
  config: LocalGameConfig;      // JSON-serializable
  seed: number;
  humanActions: GameAction[];   // human captain only, chronological
  advisorUsed: boolean;
  displayName?: string;
  // decisionPct / decisionGrade — optional metadata, not used for TEI
}

Server derives won, skill, and objective from config + replay. Client never sends won.

Replay strategies

Strategy	Client sends	Server does	Blocks `{ won: true }`?
2a Full log	Entire `actionLog`	Replay all `applyAction`	Yes (must forge full legal log)
2b Human + server AI	`humanActions` only	Run AI from same seed	Yes (strong for standard AI)

Recommendation: 2b for production. Smaller payload; AI moves cannot be spoofed.

Spike code (this repo)

File	Purpose
`apps/Warp12/src/game/verify-local-ai-replay.ts`	`replayLocalAiActionLog`, `replayLocalAiHumanActions`, `extractHumanActions`
`apps/Warp12/src/game/simulate-local-ai-match.ts`	Headless match simulation for tests
`apps/Warp12/src/game/verify-local-ai-replay.spec.ts`	Proves replay matches simulation

Run:

cd apps/Warp12 && npx vitest run src/game/verify-local-ai-replay.spec.ts

Spike result (2026-06-29): All three tests pass. Two engine issues must be handled in production replay:

Inter-round shuffle — scoreRound() defaults to Math.random when recycling tiles. Replay must use the same seeded stream as the client (createMatchRoundReshuffle(seed) in verify-local-ai-replay.ts, matching self-play’s mulberry32(seed ^ 0x9e3779b9)).
AI decision side effects — decideGameAction reads through observe(state); run AI on structuredClone(state) so the authoritative state matches the action log. Bridge UI should do the same before dispatching AI moves.

Gaps before production

Client

Persist seed on BridgeTable (pass from launchSession — already available on parent).
Filter actionLog to extractHumanActions(config, log) at match end.
Send payload to new/updated Cloud Function instead of won.
Drop academy startingTei client writes (separate security item).

Cloud Function

Add warp12-engine dependency to functions/ (or move replay into shared package).
Port replayLocalAiHumanActions, applyMatchAction, and createMatchRoundReshuffle (or import from built bundle).
Memory: 512MB default; Class I* may need 1GB + 120s timeout.
Reject Class I* for v1 verified rated play, or benchmark ISMCTS cost first.
Long-term: teach applyAction / scoreRound to derive inter-round shuffle from match seed so client and server cannot drift.

Not covered by replay

Advisor-assisted matches — keep advisorUsed: true as unrated (no TEI change), skip strict replay or replay without rating.
Drop-to-impulse AI challenges — spike driver auto-END_ROUND only; production driver must mirror bridge-table off-turn AI logic.
External assistance — solver during play still possible; replay only proves move legality + outcome.

Cost estimate (unchanged from planning)

Standard AI replay: ~2–15s CPU, ~$0.00005–0.0002/match at Gen2 pricing. Class I* higher.

Next implementation step

Run spike tests locally and note timing for ensign/lieutenant/commander.
Wire client payload (human actions + seed + config).
Extend reportPracticeAiMatch to replay before writing localAi.