Repository navigation
Unify native inference and game-owned action evidence - #248
Conversation
Co-authored-by: GPT-6 <noreply@openai.com>
Co-authored-by: GPT-6 <noreply@openai.com>
Co-authored-by: GPT-6 <noreply@openai.com>
Co-authored-by: GPT-6 <noreply@openai.com>
Co-authored-by: GPT-6 <noreply@openai.com>
Co-authored-by: GPT-6 Codex <noreply@openai.com>
Keep both current private-wire contracts and served-inference engine-binding checks. No production activation. Co-authored-by: Codex GPT-6 <noreply@openai.com>
Keep one authoritative decision serializer. Game scoring stays distinct from generated tokens and private player envelopes; current transport and u8 execution contracts remain. Co-authored-by: Codex GPT-6 <noreply@openai.com>
relh
left a comment
There was a problem hiding this comment.
Code review of 87c40eaa18229a3497ac68c15cc83345747e6e57.
Ready for library code merge; qualify consumers before publishing or claiming native training readiness.
Reviewed immutable engine records, selected-action equality, explicit fallback attribution, private wire exclusions, candidate-only scoring evidence, inference capture, and existing transport/u8 contracts. No must-fix finding in these paths.
Fresh validation: all 21 tests/test_decision_trajectory.nim cases pass using exact locked dependency source snapshots and one compiler worker. This includes rejecting player-supplied scoring evidence and retaining existing private byte/reader contracts.
The HTTP helpers do not themselves populate every exact-byte/header/join field required for fully qualified native serving evidence. Callers and the independent archive join must supply those facts; absent fields must remain unqualified. Linux-image and current-revision consumer qualification remain open, as documented.
[Codex GPT-6, acting on behalf of Richard Higgins]
The game now writes private native decision evidence after engine action application. Captures retain exact provider attempts, rejected replies, fallback attribution, private observations, and terminal outcomes. Public replay remains separate. The source-owned exporter replays the seeded engine and verifies prompts, parsed actions, execution, and scores. The shared campaign importer must additionally join independent provider archives and obtain content-bound reviews. Missing decoder metadata never qualifies checkpoint evaluation. Depends on Metta-AI/bitworld#248, pinned at `3f0266d0d9a18661b9e65244adc84871215bef81`. Validation: the native game binary compiled, then accepted, retry, and fallback HTTP/WebSocket fixtures passed against this committed source revision. Each fixture checks private file permissions, absence of operator prompts in public replay/logs, and game-owned replay verification. These are bounded synthetic CPU infrastructure checks, not ordinary hosted gameplay or approved teacher data. Publication and hosted image qualification remain pending. Keep this PR unmerged for review.
One authoritative native decision serializer now owns served-inference capture, engine-selected action joins, and game-owned scoring evidence. Consolidates #262. Scoring remains candidate-only with null generation response and no sampled-token/log-probability claims; private player envelopes cannot supply scoring evidence.
Integrated current master
c336f54e10e9028d9cfbda540ee085ac7627fbb3. Resolved both test conflicts by preserving current transport/private-wire/u8 contracts and the source proposals' distinct behavioral checks. No extra serializer, schema adapter, or model backend was introduced.Validation: all 21
tests/test_decision_trajectory.nimcontracts pass under Nim 2.2.10, with--parallelBuild:1. Git whitespace and current-master ancestry pass. Historical Goofspiel/Hanabi/Connect Four/Negotiation fixture results remain attached to5d5edde; this integration does not rerun or recertify those games. Their frozen dependency pins remain source-specific.The combined diff is four files, 316 additions. Code review is ready. Pinned Linux-image and game consumer qualification remain separate; synthetic HTTP fixtures establish no serving authenticity or learning strength. Source branches, failed attempts, and historical evidence remain preserved. No runtime was activated.
[Codex GPT-6, acting on behalf of Richard Higgins]