docs/hugging-face-release.md

# Hugging Face release procedure

Shiftedx Bench and evaluated model cards use immutable, two-revision provenance. Never benchmark or cite a moving `main` branch.

## Publish the benchmark

1. Freeze the version, configurations, schemas, CLI help, and changelog.
2. Run the complete test suite and `shiftedx-bench validate-release --root .` against a clean source tree.
3. Publish the clean tree to `Shiftedx/shiftedx-bench` in one release commit.
4. Fetch the repository anonymously, record the full commit SHA, and confirm the version and public file list from that immutable revision.

## Qualify a model

1. Record the model's immutable weight-bearing Hugging Face commit.
2. Serve byte-identical weights from that revision using the intended public runtime contract.
3. Run the lightweight gate with both `--hf-revision` and `--benchmark-revision` set to full immutable commit SHAs.
4. For Shiftedx Agent Harness claims, run complete paired `baseline` and `shiftedx-harness-v1` treatments under the same model, runtime, sampler, reasoning mode, tools, case order, and budgets. Use at least three complete trials for a promotion claim.
5. Generate aggregate summaries only after every expected row and manifest exists.
6. Export public ledgers and manifests with `export-public-results`; retain raw JSONL privately.
7. Render the standard model-card block from the complete post-publication run. Report baseline model scores separately from harness-assisted deployment results.

## Publish the model-card enrichment

1. Stage the README plus sanitized aggregate/ledger artifacts in a clean directory.
2. Run the public-release scanner on that directory.
3. Commit only metadata and sanitized benchmark artifacts; do not modify weight files.
4. Verify the rendered card anonymously at the new commit.
5. Confirm that all weight LFS hashes still match the recorded weight-bearing revision.

Do not publish access tokens, raw model responses, prompts, transcripts, local paths, request/session identifiers, private host names, or interrupted maximum-window outputs. A watchdog or out-of-memory event is a host operability limit, not an incorrect model answer.