forge-report

When this fires

Phase position: REPORT — last phase in the sequence. Runs after TEARDOWN. The two MD files it writes are what the operator reads in the morning.

What it does

1. Reads every artifact in slm-forge/.runs/<run-id>/:
   • plan.json, analysis.json, state.json
   • eval-report.json, plan-fit-report-final.json
   • card-validator-report.json, smoketest-report.json, publish-report.json
   • samples.md (if EVAL ran)
2. Computes the final verdict:
   • READY-FOR-DEMO — all gates passed cleanly
   • PUBLISHED-WITH-CAVEATS — published but one of PLAN_FIT / CARD_VALIDATOR / SMOKETEST had a soft fail
   • FAILED-DO-NOT-USE — fundamental failure (e.g. forged perplexity ≥ baseline)
3. Writes after-action.md with:
   • Final HF URLs
   • Plan vs. reality (estimated vs actual cost + wall-clock)
   • Eval summary (perplexity forged vs baseline, delta)
   • PLAN_FIT summary (axis 1 in-domain %, axis 3 Q/A grader mean)
   • Input recap (source, format, domain, base model, budget)
   • Phase ledger with ✅ per completed phase
   • First 3 sample generations (best-of)
4. Writes qa-report.md with gate-by-gate details + final verdict line

Inputs

• $1 = run-id

Outputs

• slm-forge/.runs/<run-id>/after-action.md
• slm-forge/.runs/<run-id>/qa-report.md
• Stdout JSON: {"status":"completed","after_action_report":"...","qa_report":"...","verdict":"..."}

Usage

bash slm-forge/skills/forge-report/run.sh <run-id>