Start

Operations and Release Evidence

START uses the filesystem as its operational record. This keeps the current single-user scope inspectable without introducing a database.

Run records

Each canonical run lives below the configured curriculum .runs directory and contains a manifest plus stage checkpoints. The manifest records:

Inspect history with:

uv run start-run-history --root data/written_curriculums/.runs --json

Retention is plan-first. The command never removes anything unless both --prune and --apply are supplied, and it only considers manifest-bearing run directories:

uv run start-run-history \
  --root data/written_curriculums/.runs \
  --keep 10 --older-than-days 30 --prune --json
uv run start-run-history \
  --root data/written_curriculums/.runs \
  --keep 10 --older-than-days 30 --prune --apply --json

Review the first command’s JSON before applying the second. Never prune a run that is the only copy of evidence required for a publication or incident review.

Cost and failure review

Usage is normalized to prompt tokens, completion tokens, total tokens, estimated cost, observed provider cost, and request count. Estimates are not spend. A budget refusal is persisted as a failed preflight manifest, and an observed over-budget run cancels downstream work and refuses publication.

For machine-readable operational output:

START_LOG_FORMAT=json uv run start-curriculum \
  --non-interactive --offline --dry-run --run-id ops-check --json

Logs redact common API-key and bearer-token forms. Prompts and provider error body text are not used as exception messages.

Release evidence sequence

Run the following in order and preserve the outputs for review:

uv run start-gate types
uv run start-gate audit
uv run start-gate artifacts
uv run start-gate docs
uv run start-gate numbers

The arms are the same commands the manual release workflow runs; see Gate entry point. Publication floors that only apply to a release stay in .github/workflows/release-gate.yml.

The public-release gate additionally requires publication-mode content validation, a complete provenance review, security sign-off, and human review of any authorized live pilot. No local green check substitutes for those human or external decisions.