[WS2][RFC] Clarify cross-config logprob drift contract#207
[WS2][RFC] Clarify cross-config logprob drift contract#207CyberSecurityErial wants to merge 2 commits into
Conversation
📝 WalkthroughWalkthroughThis PR adds a single new RFC documentation file defining the WS2 cross-config logprob drift contract, covering the dlogp measurement definition, comparability requirements, primary failure signal, diagnostics, tolerance sourcing/interpretation, drift source taxonomy, decision rules, a layered ablation strategy, and a follow-up test matrix scope. ChangesWS2 Drift Contract Documentation
Estimated code review effort: 1 (Trivial) | ~5 minutes Possibly related issues
Poem A rabbit sat with quill in paw, 🚥 Pre-merge checks | ✅ 5✅ Passed checks (5 passed)
✨ 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 |
Flink-ddd
left a comment
There was a problem hiding this comment.
Excellent work! LGTM, Here's a small suggestion: could you add a design sketch to this document?
Summary
Part of #111. Builds on #108.
This PR clarifies the WS2 RFC tolerance model: the #108 per-dtype threshold is a numerical infrastructure contract for pre-update selected-token logprob alignment, not a universal proof of algorithmic harmlessness.
Changes
max_abs_dlogpsatisfying the [WS1] Ground-truth harness + numerical contract for batch-invariant ops #108 per-dtype threshold.Validation
Results:
git diff --check: passmkdocs build --strict: not run; local Python environment does not have mkdocs installed (/usr/bin/python3: No module named mkdocs)Notes
Documentation-only RFC update. This does not add a second tolerance table, runtime gates, tests, or implementation changes.
Summary by CodeRabbit