Skip to content

Add run-1 decision kit: §4 analyzer and live spend kill-switch - #69

Draft
Ian-Tharp wants to merge 1 commit into
developfrom
claude/dynamic-subgraphs-run1-kit-yfc9kv
Draft

Ian-Tharp wants to merge 1 commit into
developfrom
claude/dynamic-subgraphs-run1-kit-yfc9kv

Conversation

@Ian-Tharp

Copy link
Copy Markdown
Owner

Opened from a finished Claude Code cloud session branch (the session pushed but didn't open a PR).

Commits

Add run-1 decision kit: §4 analyzer and live spend kill-switch

  • dynamic_subgraphs/_bench_verdict.py applies the frozen §4 rule to a
    bench.jsonl, keyed by arm name (A1/A2 are both invented): floor, per-task
    paired Pareto on full and worker-only cost, 60% share + exact one-sided
    sign test, seeded 10k bootstrap CI, lambda=0.05 cut, and
    KEEP/KILL/INCONCLUSIVE/INCONCLUSIVE-leaning-template-cache per sub-arm,
    headline and stratum. Every open choice is fixed in code before paid data.
  • --analyze <bench.jsonl> writes verdict.json + verdict.md.
  • SpendGuard: stop before spend passes spend_cap_usd ($8, top of the
    registered envelope; --max-spend-usd overrides) or when a STOP file
    appears; live runs need a cap and abort on unpriced spend; spend.json
    records the outcome.
  • Golden tests for all four verdicts plus the bar arithmetic; runbook §6
    documents the choices and the win bars for owner sign-off.

Co-Authored-By: Claude Opus 5.5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_019KQ4WiXCSkZA3wXfnj3BSo

🤖 Generated with Claude Code

- dynamic_subgraphs/_bench_verdict.py applies the frozen §4 rule to a
  bench.jsonl, keyed by arm name (A1/A2 are both `invented`): floor, per-task
  paired Pareto on full and worker-only cost, 60% share + exact one-sided
  sign test, seeded 10k bootstrap CI, lambda=0.05 cut, and
  KEEP/KILL/INCONCLUSIVE/INCONCLUSIVE-leaning-template-cache per sub-arm,
  headline and stratum. Every open choice is fixed in code before paid data.
- `--analyze <bench.jsonl>` writes verdict.json + verdict.md.
- SpendGuard: stop before spend passes spend_cap_usd ($8, top of the
  registered envelope; --max-spend-usd overrides) or when a STOP file
  appears; live runs need a cap and abort on unpriced spend; spend.json
  records the outcome.
- Golden tests for all four verdicts plus the bar arithmetic; runbook §6
  documents the choices and the win bars for owner sign-off.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019KQ4WiXCSkZA3wXfnj3BSo

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants