jax ile: endpoint gap off on the 6-D field; refuse narrow windows at parse time (follow-up to #288) - #299
Merged
Conversation
Refine the complex primitive kappa(t) first, then apply the same distance quadrature on the refined nodes, under the existing fail-closed certificates. Crop to the integrated window before reducing: same value, 6.50 -> 2.56 GB. The angle-marginalized wrappers still refuse, and now say why they do. Numbers in DESIGN_jax_bandlimited_distmarg.md; 14 tests wired into the gate. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…ax-ile-check) The CI job installs jax but not flowMC; pytest.importorskip makes this a real skip there. A non-flowMC mode was tried and dropped: blind full-sky prior draws trip the reflected-FFT convergence check (fail-closed by design) on about a quarter of draws for this injection, so an IS-type mode is not a reliable carrier for this test. That fragility is flagged for a follow-up. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…ver names failing rows Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…ment The distance-marginalized field has a floor (far-distance prior mass), so the fixed-distance kernel's 15-nat endpoint gap rejects every low-contrast row, converged or not: 90 of 256 blind prior draws at a 20 ms window, all within 1e-4 nat of the independent reference. Every prior-seeded driver mode stopped on its first chunk. `_time_marginalize_reflected_primitive` takes `endpoint_log_gap`; the distance path passes None, the 6-D kernel keeps the default. Doubling, resolution and guard certificates are unchanged. The driver still stops on an uncertified row and now counts them, prints the first three, and names a window narrower than the detector arrival shifts (2 R_earth/c ~ 43 ms) as the usual cause: with the gap off, 36 of 256 rows still fail at 20 ms and 0 of 256 at 50 and 75 ms. Six tests: blind census with and without the gap against the reference, the 6-D default pinned, the driver message, and prior-mc / map end to end under --distance-marginalization bandlimited. The flowMC driver test is deselected in CI (it would skip). EXPECTED_TESTS 647, measured on the merged tree. Numbers: DESIGN_jax_bandlimited_distmarg.md, "The endpoint certificate". Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…windows at parse time The fixed-distance bandlimited kernel returned NaN on 92 of 256 blind full-sky prior draws at a 20 ms half-window (rift_O4d d84597c), so every prior-seeded driver mode without distance marginalization stopped in its first chunk. 57 of those rows failed only the 15-nat endpoint gap and agree with an independent reference to 1.1e-4 nat; the gap certifies a row's amplitude, not the quadrature. The other 35 have an arrival peak beyond the window edge and vanish at 50 ms. - core: endpoint_log_gap defaults to None on both fields (constant kept). - driver: prior-mc, laplace-is, map, nuts + bandlimited are refused at parse time when the half-window is below 2 R_earth / c. - test/jax/test_jax_bandlimited_6d_blind.py, gated; EXPECTED_TESTS 647 -> 657 (ldas-grid, ~/.cache/jaxci_venv). Evidence: DESIGN_jax_bandlimited_distmarg.md, "The fixed-distance kernel". Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
oshaughnessy-junior
deployed
to
private-review-dispatch-rift
September 9, 2026 00:17 — with 
GitHub Actions
Active
GitHub Actions
Active
…x-6d-bandlimited-blind Resolves the EXPECTED_TESTS comment tail (their measured 647 line kept, my seventeenth block after it, 657) and the DESIGN record (their driver-run table and rho caveat first, my fixed-distance subsection after). Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
oshaughnessy-junior
changed the base branch from
claude/jax-bandlimited-distmarg
to
rift_O4d
September 9, 2026 00:23
oshaughnessy-junior
deployed
to
private-review-dispatch-rift
September 9, 2026 00:23 — with
GitHub Actions
Active
GitHub Actions
Active
oshaughnessy-junior
deployed
to
private-review-dispatch-rift
September 9, 2026 01:44 — with
GitHub Actions
Active
GitHub Actions
Active
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 to #288 (branched from its head 8e928ca, so the diff shows #288's commits until it merges). Pre-existing on rift_O4d: the 6-D bandlimited kernel returned NaN on 92 of 256 blind full-sky prior draws at a 20 ms half-window (35+30 Msun at 900 Mpc, H1L1), so
--mode prior-mc / laplace-is / mapwithout distance marginalization stopped in the first chunk.Cause: the 15-nat endpoint gap. A row's contrast is bounded by its own amplitude, so a far or wrong-sky draw can never show 15 nat. Rows it rejected alone match an independent reference to 1.1e-4 nat, 54 to 128 nat below the batch maximum. The other 35 rows at 20 ms have an arrival peak past the window edge; at 50 ms there are 0.
Changes:
endpoint_log_gapdefaults toNoneon both fields. A behaviour change on the opt-in quadrature: RO's call.prior-mc,laplace-is,map,nutswithbandlimitedat parse time below a 2 R_earth/c (42.6 ms) half-window.test_jax_bandlimited_6d_blind.py(10 tests); EXPECTED_TESTS 647 to 657, measured on ldas-grid on the merged tree.Memory (pcdev13, peak RSS): driver prior-mc 1.25 GB bandlimited vs 1.05 GB Simpson; map 1.70 vs 1.27 GB; the pytest process with both kernel fixtures 2.15 GB.
Landing: after #288 (and #298, whose sharding this needs on the CI runner). Conflicts with #295, #289, #286, #266 are the EXPECTED_TESTS tail and adjacent driver lines. The full gate on ldas-grid aborts in
test_joint_anglemarg_peaklocal.py(the 20 GiB scan #295 reduces), before this file; the two bandlimited files pass there, 30 of 30.🤖 Generated with Claude Code