validate-run

name	validate-run
description	Internal Harness instruction source for validate-run. Route through visible Harness aliases or hook contracts instead of invoking directly.

Validate Run

References

Read these first:

• ../../../.agents/references/workflow-guide.md
• ../../../.agents/references/contract-gating-rule.md
• ../../../.agents/references/ubiquitous-language.md
• ../../../.agents/references/language-policy.md
• ../../../.agents/references/run-artifact-contract.md
• ../../../.agents/references/documentation-evidence-rule.md
• ../../../.agents/references/documentation-style.md
• ../../../.agents/references/reviewer-independence.md
• ../../../.agents/references/review-tracing.md
• ./references/review-checklist.md
• ./references/validate-run-report.md
• ../../../project_map.json
• ../../../docs/Technical_Spec.md
• ../../../docs/Implementation_Roadmap.md if it exists
• ../../../docs/20_facts/Project_Glossary.md if it exists
• ../../../docs/20_facts/Codebase_Map.md if it exists
• ../../../docs/30_evidence/Validation_Table.md if it exists
• ../../../docs/10_contract/Evaluation_Contract.md if it exists
• ../../../docs/10_contract/Baseline_Contract.md if it exists
• ../../../CLAUDE.md

When To Use

Use this skill for WF9 when the user wants to verify the codebase is safe to enter the iteration loop.

Required Work

1. Read the new implementation, baseline references, and technical spec.
2. Run the canonical semantic review:
   • data pipeline equivalence
   • model and rendering equivalence
   • loss equivalence
   • metric equivalence
   • common ML bug checks
3. Review slice completion against docs/Implementation_Roadmap.md: slice trace, planned files, public interfaces, test/smoke command, and downstream validation doc.
4. Review language, boundary, and complexity:
   • identifiers, configs, metrics, tests, and errors follow docs/20_facts/Project_Glossary.md when present
   • public APIs and dependencies match module boundaries
   • complexity budget changes are explicit
5. Run the canonical smoke test:
   • 100-step training
   • checkpoint save and load
   • evaluation run
   • wandb check
   • git_snapshot check
   • run artifact bundle check
6. Classify the result as:
   • PASS
   • REVIEW
   • FAIL
7. Write docs/Validate_Run_Report.md with evidence sources, raw log paths, run artifact bundle paths, review trace paths, commands, and verdict.
8. Create or refresh docs/30_evidence/Validation_Table.md with the human-readable Conclusion Evidence behind the validation verdict: reviewed slices, smoke commands, raw log paths, review traces, failures, and open validation questions.
9. Clean up temporary smoke-test artifacts.
10. Report a gate ledger for semantic review, smoke test, report write, and any workflow-state check run before WF10 readiness.
11. After docs/Validate_Run_Report.md or docs/30_evidence/Validation_Table.md is finalized, invoke $docs-site or report docs_site_boundary_report / docs_site_render_or_NOT_RUN. Do not render after temporary draft edits.
12. In workflow-supervisor $build, include a final Gate ledger entry with command validate-run verdict:
    • PASS only when the semantic review, required smoke/eval/training checks, checkpoint or dry-run substitute explicitly allowed by the roadmap, run-artifact bundle checks, and dynamic-context gates passed.
    • FAIL when the runnable entrypoint, config, evaluator, training dry-run, checkpoint path, run-manifest bundle, or required tests are absent.
    • NOT_RUN only with a concrete blocker and follow-up action. REVIEW must not be reported as PASS unless the operator has explicitly approved proceeding with the recorded review findings.

Codex Adaptation

• Treat natural-language requests as the canonical $validate-run [config_path] flow.
• Preserve the canonical two-part gate: semantic review plus smoke test.
• Preserve validate-run verdict as the supervisor-readable readiness gate. Do not mark build_ready_for_iterate from a report that says foundation-only, partial, FAIL, REVIEW without approval, or smoke checks NOT_RUN.
• If the canonical workflow routes to a fix step, use $code-debug.
• Use ../../../.agents/references/language-policy.md for reply language and for localizing natural-language review notes and verdict summaries; keep checklist item names, commands, status labels, and identifiers in English.

Execution Rule

Follow the local validation sequence and language policy instead of reducing this stage to a quick smoke test only. Do not report WF9 as PASS unless the semantic review and smoke-test evidence are listed with commands/log paths or explicitly marked NOT_RUN.