TRAINING-PREFLIGHT-006: explicit local-agent handoff - #19
scrimshawlife-ctrl wants to merge 2 commits into
Conversation
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: afbc3513be
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| python -m athanor.training_rights \ | ||
| --atoms ~/.athanor/corpus/atoms.jsonl \ | ||
| --candidate-manifest-sha256 <CANDIDATE_MANIFEST_SHA256> \ |
There was a problem hiding this comment.
When the local corpus snapshot differs from the candidate pack, this command only attaches a caller-supplied manifest string to every atom in ~/.athanor/corpus/atoms.jsonl; audit_training_rights() never joins those atoms to pack provenance or proves complete source coverage, while settle_training_rights() accepts any nonempty list of cleared sources. The sequence can therefore produce G3 PASS from an unrelated or incomplete source review. Derive the queue from the exact pack provenance and reject missing or extra reviewed source IDs.
AGENTS.md reference: AGENTS.md:L14-L15
Useful? React with 👍 / 👎.
| --rights-evidence /private/provenance-rights.json \ | ||
| --gold-evidence /private/independent-gold.json \ | ||
| --baseline-evidence /private/untrained-baseline.json \ | ||
| --config-evidence /private/training-config.json \ |
There was a problem hiding this comment.
A repo-wide search for these evidence kinds finds no producer or substantive validator for the GOLD, baseline, or config sidecars. training_preflight._evidence_gate() merely checks the envelope's kind/status/binding fields and that evidence_sha256 looks like 64 hex characters; it neither recomputes the digest nor inspects the claimed evidence. Thus arbitrary four-field JSON envelopes supplied here can make G6-G8 PASS without an independent review, baseline run, or frozen configuration. The handoff should keep these gates on HOLD until authenticated producers and validators exist.
AGENTS.md reference: AGENTS.md:L14-L15
Useful? React with 👍 / 👎.
| Use the existing candidate-pack/freeze workflow documented in | ||
| `docs/TRAINING-PREFLIGHT-001.md`. Do not invent source grouping identities. If a |
There was a problem hiding this comment.
docs/TRAINING-PREFLIGHT-001.md only documents auditing an already-existing candidate-preparation/1 pack; it contains no pack creation or freeze workflow. A repo-wide search also shows src/athanor/dataset.py::prepare_dataset() immediately exits HOLD, while athanor.adapter_freeze emits a different freeze receipt rather than the four-file pack consumed by training_preflight. A local agent without a preexisting compatible pack therefore cannot execute this authoritative sequence; identify a real builder or explicitly require an operator-supplied pack.
AGENTS.md reference: AGENTS.md:L14-L15
Useful? React with 👍 / 👎.
1 participant
Adds an explicit executable handoff for the agent on the machine holding ~/.athanor. Covers synchronization, G2-G5 freeze/leakage evidence, G3 rights audit and independent model-training review, G6 GOLD, G7 untouched baseline, G8 config freeze, final receipt assembly, stop conditions, and return packet. AGENTS.md now routes all training work through it. No training authorization is granted.