ForgeFlow / Reference

Lean Evidence

The examples use Claude Code command syntax. Other hosts can use the corresponding runtime helpers. Keep observed task results separate from fixtures and empty benchmark scaffolds.

Forgeflow treats lean guidance as advisory until there is local evidence.

Use this path when you want proof instead of intuition:

Run /forgeflow-lean-prime --prime-task "<work item>" --write-report.
If you did not prime the task in step 1, run /forgeflow-lean-decision --task "<work item>".
Build and validate the change.
Run /forgeflow-lean-report --write.
Run /forgeflow-lean-review before normal review when the risk is over-building.
Use /forgeflow-lean-lab --task-pack <json> --results <json> for repeatable local comparisons.
Use /forgeflow-lean-benchmark-runner --write and validate model-backed results with /forgeflow-lean-benchmark-results --promptfoo raw-results.json --out normalized-results.json, then /forgeflow-lean-benchmark-results --results normalized-results.json. Model-backed runs also write a local benchmark run ledger for dashboard/readiness evidence.
Check the benchmark evidence grade before making claims. Publishable evidence needs provider/model metadata, at least three tasks, at least two comparison arms, at least three iterations, required metrics, and the session-cost caveat.
Use the runner's historical task pack when you want Forgeflow-specific replay coverage for helper contracts, dashboard readiness, and release-gate work.

Run /forgeflow-stale-artifact-plan after commits when latest project guidance or insights are stale; its post-commit aftercare commands keep Lean Prime and dashboard guidance current.

Do not publish performance claims from single samples, failed validation, missing correctness gates, or results without the session-cost caveat.

Host CLI proof is separate from binary detection. /forgeflow-lean-host-cli-probes --write-template creates a local evidence template; mark probes verified only after manually running the listed command and recording a timestamp, note, and optional output digest.

Release readiness now shows advisory warnings for missing benchmark evidence, missing verified host probes, missing failure-digest aftercare, and RTK command-policy alignment. These warnings are local release evidence gaps, not publish actions.