Phase 4j: the graph softmax-tie is redundancy, not complementarity - #6
Merged
Conversation
… softmax? Adds the machinery to answer whether geometry is redundant with or complementary to softmax confidence (the open question after geometry ties softmax head-to-head on graph). - eval/evaluate.py: fit a softmax-only mapper and a [geometry ∪ softmax] combined mapper on the fit fold (invariant 7); add `softmax_learned` + `geom_softmax` scores and `delta_aurc_geom_adds` = ΔAURC(softmax_learned − geom_softmax) (positive w/ CI>0 => geometry adds conditional signal) + `geometry_adds_over_softmax`. - analysis/aggregate.py: aggregate the field (+ has_geom_adds guard for pre-4j rows). - analysis/make_tables.py: T4_complementarity table (per task × reasoner: head-to-head vs geometry-adds). plotting: F2 gets softmax_learned + geom_softmax lines. - tests: complementarity assertions in the evaluate smoke test + aggregate field/fallback. 95 green. Metric only; opt-in/additive, non-breaking. The 4 seed sweeps are being re-run to populate it; verdict (complementary vs redundant per cell, esp. graph-IRED) to follow. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…e_cell hardening) Answers whether geometry ties softmax on graph out of redundancy or complementarity, via a combined [geometry ∪ softmax] mapper vs a [softmax] mapper. - aggregate.headline_cell: dedup each (K, n_test) group by seed (latest timestamp) and prefer the newest field-carrying cell — fixes a re-run bug where adding config fields changed the config_hash, so old+new rows for one seed both survived and polluted the headline cell. - make_tables: T4_complementarity emitted + wired into the paper; F2 gains the combined line. - docs/paper (abstract already tempered; experiments/conclusion) report the redundancy finding. RESULT (5 seeds/cell): geometry adds conditional signal over softmax ONLY where it wins head-to-head (arithmetic-IRED, 4/5, dAURC_adds +0.007). Everywhere else, incl. the crux graph-IRED cell, the combined mapper ≈ softmax alone (0/5, +0.001) — so the graph tie is genuine redundancy: a well- calibrated softmax already captures what descent geometry would add there. Clean summary: geometry always beats scalar energy; it beats/adds-to softmax only under a learned landscape on arithmetic. 95 tests green, ruff clean; paper rebuilds. Additive; default reasoner + T1 unchanged. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
|
Bugbot is not enabled for your account, so this pull request was not reviewed. Enable Bugbot in the Cursor dashboard to get automatic reviews on future PRs. |
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.
Resolves what "geometry ties softmax on graph" means, via a combined-mapper test.
Method
Fit a nonconformity mapper on [geometry ∪ softmax] and compare to one on [softmax] alone (both on the fit fold). Positive ΔAURC ⇒ geometry adds conditional signal over softmax.
Result (5 seeds / cell → T4)
Geometry adds over softmax only where it also wins head-to-head (arithmetic-IRED). On graph-IRED the combined mapper is no better than softmax alone — so the graph tie is genuine redundancy, not a complementary-but-losing signal. Clean summary: geometry always beats the EBT scalar energy; it beats/adds-to softmax only under a learned landscape on arithmetic.
Also
headline_cellhardened: config-schema additions changed the config_hash, so a re-run's rows didn't dedup against the old ones (a 10-row K=12 headline group). Now dedups each (K, n_test) group by seed (latest timestamp) and prefers the newest field-carrying cell; T1 stays canonical basin-center.🤖 Generated with Claude Code