Phase 4l: certificates on firm footing (multi-seed guarantees + graph certification) - #8
Merged
Merged
Conversation
… certification)
Puts the project's core contribution — the distribution-free certificates — on a
5-seed footing, and shows the graph certificate is valid at its honest operating
point. Additive: no change to geometry features or the falsification test;
include_ood defaults False so existing sweeps are byte-identical.
Halting (CRC), 5 seeds (T6): new experiments/run_halting_sweep.py +
aggregate.{halting_rows, aggregate_halting_cell}. Arithmetic saves 57.1%±1.7%
(single-seed H1 was 57.8%), graph 80.8%±5.2%; both halting risk <= alpha,
within budget 5/5.
OOD stress, 5 seeds (T7): [sweep] include_ood=true in run_sweep + an OOD
aggregation block. The selective guarantee breaks under shift over 5 seeds
(arith ID sel-risk 0.077 -> OOD 0.764, 0/5; graph 0.010 -> 0.168, 4/5).
Graph certification: a full-fold run (n_calib=1500) certifies only 1.1% coverage
at alpha=0.1 but 22.5% at 0.15 and 48.5% at 0.20, with achieved risk <= target
at every alpha. The old "certifies zero coverage" was an operating-point artifact
of graph's ~30% error floor, not a broken certificate.
Bugs fixed: halting aggregation dedups by seed (the stale Phase-4b H1 seed-0 row
survived under a drifted config_hash -> n=6); plotting._selective_row/_halting_row
prefer arithmetic so the new full-fold (n_test=1500) graph rows don't flip the
arithmetic figures F2/F4/F6 to graph.
115 tests green, ruff clean, paper rebuilds via tectonic with T6/T7 wired in.
Co-Authored-By: Claude Opus 4.8 <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.
What & why
The geometry-vs-softmax science is well-characterized (Phases 4e–4k); the project's actual contribution — the distribution-free certificates — had three soft spots. This phase puts all three on a 5-seed footing and shows the graph certificate is valid at its honest operating point. Additive: no change to geometry features or the falsification test;
include_ooddefaults False so existing sweeps are byte-identical.Results
Adaptive halting (CRC), 5 seeds — T6. New
experiments/run_halting_sweep.py+aggregate.{halting_rows, aggregate_halting_cell}(halting rows were excluded fromselective_rows).OOD stress, 5 seeds — T7.
[sweep] include_ood=true+ an OOD aggregation block. The selective guarantee breaks under shift over 5 seeds: arithmetic ID sel-risk 0.077 → OOD 0.764 (0/5 within budget); graph 0.010 → 0.168 (4/5).Graph certification. A full-fold run (
graph_cert_seeds.toml, n_calib=1500) certifies only 1.1% coverage at α=0.1, but 22.5% at α=0.15 (achieved risk 0.078) and 48.5% at α=0.20 (0.157), with achieved risk ≤ target at every α. The old "certifies zero coverage" note was an operating-point artifact of graph's ~30% error floor, not a broken certificate.Bugs fixed
config_hash(gave n=6).plotting._selective_row/_halting_rownow prefer arithmetic, so the new full-fold (n_test=1500) graph rows don't tie/flip the arithmetic figures F2/F4/F6 to graph.Tests
115 green (offline, CPU-only), ruff clean. New
test_run_halting_sweep.py(monkeypatched — no training, no ledger writes); extendedtest_aggregate(OOD + halting + fallback),test_run_sweep(include_ood),test_make_tables(T6/T7). Paper rebuilds viatectonicwith T6/T7 wired in; abstract's halting number updated to the 5-seed range.🤖 Generated with Claude Code