feat/bot-baseline-inference #54

Merged

hellsingoff merged 2 commits from feat/bot-baseline-inference into main 2026-11-14 22:28:20 +00:00

Conversation 0 Commits 2 Files changed 76 +4153 -80

hellsingoff commented 2026-11-14 22:03:42 +00:00

Owner

Copy link

Изменение

Проблема и получившееся поведение; ссылка на PLAN/ADR.

Проверка

Команды и результаты, ограничения проверки, способ отката.

• make check, make test, make smoke проходят.
• Для protocol/security/storage изменений выполнены правила review.
• Измененные contracts, fixtures и документация согласованы.
• Секреты и приватные игровые данные отсутствуют в diff и логах.

## Изменение Проблема и получившееся поведение; ссылка на PLAN/ADR. ## Проверка Команды и результаты, ограничения проверки, способ отката. - [ ] `make check`, `make test`, `make smoke` проходят. - [ ] Для protocol/security/storage изменений выполнены правила [review](../docs/OWNERSHIP.md). - [ ] Измененные contracts, fixtures и документация согласованы. - [ ] Секреты и приватные игровые данные отсутствуют в diff и логах.

hellsingoff added 2 commits 2026-11-14 22:03:42 +00:00

Play bot turns through baseline policies, inference and a fallback ... f73d8f37ef

Implement P6.2: bots answer each decision through a bot job that is
resolved exactly once, by the inference service or by a fallback policy.

- Policies score candidates and TemperatureSampler picks with the bot
  level's temperature; RandomLegalPolicy works for every game and Durak
  adds DurakHeuristic through the optional BotHeuristicPolicy
- Bot levels (easy, medium, hard) set memory and temperature; the memory
  limit is an explicit `memory` field of the shared observation, Durak
  state records the round each remembered card was seen and forgotten
  cards count as unseen again; fingerprint baseline recomputed for @1
- One bot job per decision: request ID and command ULID derive from game,
  epoch, state version, turn and seat; the unique index now includes the
  version, and admission idempotency on that ULID is the job resolution
- Worker opens jobs and inference requests, writer stores requests as
  separate outbox rows, relay sends them only to epicly.bots and drains a
  bounded backlog per pass; bot seats get no turn notifications
- BotResults accepts an answer only for the current decision, version and
  budget; BotScheduler answers with the heuristic after the budget
- Python inference consumer (services/ai/inference, profile `ai`) with a
  SQLite inbox, deadline budget and confirmed publication
- 1000 baseline games on Durak and Pebble finish; P6.2a plans per-seat
  level selection and inference for a bot's first turn

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Plan training one model for every bot memory level ... b5d29d8957

Self-play does not wait for the level choice in the interface: the memory
limit is a model input, so the simulator takes it per seat and training
draws it at random from a wider set than the offered levels. Evaluation
reports each level and checks that strength grows with memory.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

hellsingoff scheduled this pull request to auto merge when all checks succeed 2026-11-14 22:04:13 +00:00

hellsingoff merged commit 88257f09db into main 2026-11-14 22:28:20 +00:00

hellsingoff deleted branch feat/bot-baseline-inference 2026-11-14 22:28:20 +00:00

hellsingoff referenced this pull request from a commit 2026-11-14 22:28:22 +00:00

Merge pull request 'feat/bot-baseline-inference' (#54) from feat/bot-baseline-inference into main

Sign in to join this conversation.

Reviewers

No reviewers

Labels

Clear labels

No items

No labels

Milestone

Clear milestone

No items

No milestone

Projects

Clear projects

No items

No project

Assignees

Clear assignees

No assignees

1 participant

Notifications

Subscribe

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference

hellsingoff/epicly-unstable-app!54

No description provided.