Key the SSI swap cap on the selection arm, not the export-mode flag#455
Merged
Conversation
The #452 dense cap keyed on dense_default_dataset - but that is an EXPORT-MODE flag the sparse rmloss100 script also passes, so sparse runs inherited the dense 0.12 bar: the 32605c9 sparse rerun gated its swap delta (710,076 = 10.9%) at 0.12 and exited reconciliation at pass 1, where the sparse 0.10 bar would have forced further passes (the certified attempt-16 run converged to 7.1% by pass 2). Caught by the recorded national_swap_sanity_cap_ratio in the rerun's manifest before anything shipped; the certified default (98bf731) predates #452 entirely and the dense arm's behavior is unchanged under either discriminator. The arm identity is the frozen-selection identity: reconcile now takes sparse_selection_arm (main passes selection_source_manifest presence) and keys the ratio on that. Constants document the hazard so the export-mode flag is never reused as an arm discriminator. Tests: the fresh-pair reconcile test is parametrized over both arms - the sparse-arm-in-dense-export-mode leg is the leak regression and fails without this fix; the dense-arm leg pins 0.12. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Follow-up defect in #452, caught by the same-commit sparse rerun's own manifest (
national_swap_sanity_cap_ratio: 0.12on a sparse run) before anything shipped.The bug: #452 keyed the cap ratio on
dense_default_dataset— an export-mode flag the sparse rmloss100 script also passes. Every sparse run at #452+ inherited the dense 0.12 bar, and it changed behavior, not just labels: the 32605c9 sparse rerun's delta (710,076 = 10.9%) passed at 0.12 and exited reconciliation at pass 1, where the 0.10 sparse bar forces further convergence (certified attempt 16 reached 7.1% by pass 2).Blast radius: zero shipped impact — the certified default (98bf731) predates #452; the dense arm behaves identically under either discriminator; the only affected artifact is the local rerun, which gets relaunched on this fix.
The fix: the arm identity is the frozen-selection identity.
_reconcile_ssi_take_up_and_refittakessparse_selection_arm;main()passesselection_source_manifest is not None; the constants block documents the export-mode-flag hazard.Tests: fresh-pair reconcile test parametrized over both arms — the
sparse-arm-dense-export-modeleg is the exact leak and fails without the fix (verified); dense-arm leg pins 0.12. Full suite green.🤖 Generated with Claude Code