Add binding-aware meaningful-program evaluation (SLM-105, EFS0-03)#328
Conversation
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
|
Warning Review limit reached
Next review available in: 38 minutes Enable usage-based reviews in Billing to review now. Otherwise, wait until the next included review is available. How can I continue?After more reviews become available, a review can be triggered using the To avoid repeated limits, reduce automatic review volume by pausing incremental auto-reviews earlier, using label-based review opt-in, excluding WIP or generated PR titles, or requesting reviews manually when the PR is ready. If your team needs uninterrupted high-volume reviews, an organization admin can enable usage-based reviews. How do review limits work?CodeRabbit enforces per-developer PR review limits for each organization. Most developers receive the normal plan review availability. For paid Pro and Pro+ PR reviews, CodeRabbit uses adaptive limits for sustained high-volume activity. When a developer's recent PR review activity reaches the 95th percentile or higher among CodeRabbit users, additional reviews become available more gradually as earlier reviews age out of the rolling window. Please refer docs for additional details. Review details⚙️ Run configurationConfiguration used: defaults Review profile: CHILL Plan: Pro Plus Run ID: 📒 Files selected for processing (20)
✨ Finishing Touches🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
A torch-free binding-aware meaningful-program evaluator (v2) plus a
frontier-replay + anti-gaming audit CLI, wired into the eval runner and
ship-gate reporting as a RECORDING-ONLY candidate metric.
- evals/meaningful_program.py + scripts/audit_meaningful_program.py +
resources/evals/{meaningful_v2_frontier_replay.json, meaningful_v2_gaming.jsonl}.
- ship_gates.py: adds a provenance-only MEANINGFUL_METRIC_POLICY and records
binding_aware_meaningful_v2_* alongside the gated primary. The v2 metric
carries thresholds=None (candidate_pending_calibration) and is never gated;
DEFAULT_SHIP_GATES, the decision loop, and every threshold are byte-identical.
meaningful_program_v1 remains the active gated primary.
- eval_runner.py / train_loop.py: append the new metric keys only.
- quality.py: additive helpers required by the eval (preserves E292's
semantic-contract work and the existing judge).
No ship gate weakened or relaxed; no checkpoint promoted; meaningful-parse v1
remains the gated primary. No-new-breakage verified against a clean-main
baseline (identical failure sets; +40 new passing tests).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017NWyjZwGMUPHXZhtqQVzUJ
7e2f560 to
297a76f
Compare
Tyler-R-Kendrick
changed the title
2 participants
SLM-105 (EFS0-03): binding-aware meaningful-program evaluation
Adds a torch-free binding-aware meaningful-program evaluator (v2) + a frontier-replay & anti-gaming audit CLI, wired into eval reporting as a recording-only candidate metric — no ship gate is added, relaxed, or weakened.
What's here (20 files, on current
main)evals/meaningful_program.py(765) — binding-aware evaluator;scripts/audit_meaningful_program.py(935) — frontier-replay + gaming audit CLI; eval resources (meaningful_v2_frontier_replay.json,meaningful_v2_gaming.jsonl); tests (28 + 8 + 3) + provenance docs.ship_gates.pyadds a provenance-onlyMEANINGFUL_METRIC_POLICYand recordsbinding_aware_meaningful_v2_*alongside the gated primary. The v2 metric carriesthresholds: None(candidate_pending_calibration) and is never gated;DEFAULT_SHIP_GATES, the per-metric decision loop, and every threshold are byte-identical.meaningful_program_v1remains the active gated primary.eval_runner.py/train_loop.pyappend the new keys only.quality.py— additive helpers required by the eval; preserves E292's semantic-contract work and the existing judge byte-identical.Dropped the branch's
dsl/solver/*re-adds, dashboard/web clobbers, and doc clobbers already superseded on main.Honesty
No ship gate weakened or relaxed; the v2 metric is calibration-pending and recording-only. No checkpoint promoted; meaningful-parse v1 remains the gated primary (not claimed improved).
Verification
mainbaseline (bridge + AgentV SDK installed): failure sets byte-identical (271 failed both sides); +40 new passing = exactly the new tests. Consolidated new + wiring tests: 60 passed.ruff/repo_policy/git diff --checkclean. Torch tests (none new) run in CI.🤖 Generated with Claude Code