Add cost-to-go energy scorer for the verified solver (VSS3-02, SLM-70)#347
Conversation
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
|
Warning Review limit reached
Next review available in: 20 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 (7)
✨ 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 |
48e1401 to
ad67509
Compare
The first learned model for the verified solver: a small MLP over (state, hole, candidate) features whose only runtime authority is to ORDER the exact live candidates from the CandidateRanker seam. Low energy is a ranking hint, never a correctness claim — it never defines membership, certifies support, suppresses UNKNOWN, or bypasses the verifier. - models/solver_energy.py: CandidateEnergyScorer (kept as a separate adapter so existing checkpoints load unchanged) + CandidateEnergyRanker, which validates exactly one finite energy per live value and returns a permutation; any defect (count mismatch, duplicate, NaN, inf) triggers a counted deterministic identity fallback and never alters the hard domain. v1 work target log1p(expanded_nodes + w_v*verifier_calls + w_b*backtracks + w_d*decisions) over VSS3-01 candidate_cost rows; Huber regression only where cost_observed, pairwise ranking over same-state/hole pairs; UNKNOWN/censored masked. - TwoTowerConfig: 7 disabled-by-default solver_energy_* / solver_ranker fields; old configs load with defaults. - 12 guarantee tests + a deterministic tiny fixture overfit. Fixture overfit is wiring evidence only — no checkpoint trained, no ship gate touched, MODEL_CARD unchanged. Torch tests run in CI. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_017NWyjZwGMUPHXZhtqQVzUJ
a7fc982 to
37237ed
Compare
Tyler-R-Kendrick
changed the title
2 participants
VSS3-02 (SLM-70): cost-to-go energy scorer for the verified solver
The first learned model for the verified solver. Its only runtime authority is to order the exact live candidates from the
CandidateRankerseam — it never defines membership, certifies support, suppressesUNKNOWN, or bypasses the verifier. Low energy is a ranking hint, never a correctness claim.What's here (7 files, +636/−0, on current
main)src/slm_training/models/solver_energy.py—CandidateEnergyScorer(a separate adapter, not baked into the base state dict, so existing checkpoints load unchanged) +CandidateEnergyRanker(aCandidateRanker) that validates exactly one finite energy per live value and returns a permutation; any defect (count mismatch, duplicate, NaN, inf) triggers a counted deterministic identity fallback and never alters the hard domain. Versioned v1 work targetlog1p(expanded_nodes + w_v·verifier_calls + w_b·backtracks + w_d·decisions)over VSS3-01candidate_costrows; Huber regression only wherecost_observed; pairwise ranking over same-state/hole pairs;UNKNOWN/censored masked; features are decision-time only (no witness/final text).src/slm_training/models/twotower.py— 7 disabled-by-defaultsolver_energy_*/solver_rankerconfig fields inserted after the VSS1-03 solver-certificate group; old configs load with defaults.tests/test_models/test_solver_energy.py— 12 guarantee tests + a deterministic tiny fixture overfit.verified-scope-solver.md,research-lineage.md(adapted EDLM/EBM, fidelity: adapted), andquality-experiment-matrix.md, plus a dated fixture note.Reduced to exactly the in-scope VSS3-02 change and rebased on current
main(which already carries VSS3-01'ssolver_supervisioncorpus this consumes) — the earlier branch's out-of-scope re-implementations ofdsl/solver/*, arity, dashboard assets, and README/MODEL_CARD/pyproject/CI edits were dropped.Honesty
Fixture overfit is wiring evidence only — no checkpoint trained, no ship gate touched,
MODEL_CARD.mdunchanged. Meaningful-parse is the primary metric and out of scope for this increment.Verification
main:ruffclean,compileallOK,repo_policyok,git diff --checkclean.tests/test_models/test_solver_energy.py) run in CI (torch is installed there). ConsumedCandidateCostRowfields (remaining_expanded_nodes/remaining_verifier_calls/remaining_backtracks/remaining_decisions,cost_observed,censor_reason) verified present in main'ssolver_supervision.py.🤖 Generated with Claude Code