Skip to content

The 256-seed ordering count both figures print is computed from the ensemble CSV - #60

Open
aurascoper wants to merge 3 commits into
fix/melanin-figure-hashfrom
test/ordering-count-from-csv
Open

aurascoper wants to merge 3 commits into
fix/melanin-figure-hashfrom
test/ordering-count-from-csv

Conversation

@aurascoper

Copy link
Copy Markdown
Owner

Stacked on #59 (fix/melanin-figure-hash at e20478f). This PR's head is eccf9ab.

Fig 2's label in biofilms_potts.jl and the WIDER constant in diagnostics/melanin_ensemble/figure.jl both print that the ordering fails in 19 of 256 seeds. Both were typed from ledger row SEED42-SCOPE-01 and no test read them against an ensemble. calibration/tests/test_ordering_count.py now computes the count from the committed diagnostics/melanin_ensemble/radial_n40_p6_seeds42-297.csv, taking the full arm at MCS 100 and testing per seed whether mean_melanin of species 3 (C. sphaerospermum) exceeds species 1 (C. neoformans), which exceeds species 5 (A. niger), strictly. That gives 237 holding and 19 failing over 256 seeds, with no ties.

Four locations must each print exactly one count equal to that: the two source strings and the two .txt sidecars extracted from the figures' PDFs. The sidecars are included so that an ensemble re-run which moves the count fails until both figures are re-rendered, not only until the strings are edited. A location that prints no count is refused rather than passed.

The negative controls follow AGENTS.md rule 1. A scratch copy of the CSV swaps C. sphaerospermum and C. neoformans melanin for the lowest seed whose ordering holds, selected from the file rather than recalled; the test asserts the copy differs from the source at those two rows and no others and that exactly that seed joins the failing set, and the mismatch is then reported at all four locations. Four further controls cut the count out of one location each and require the refusal. With the comparison weakened to never report anything, all five controls failed while the main check stayed green, so the comparison is load-bearing. The three-way ordering is what produces 19: C. sphaerospermum over C. neoformans alone fails in 7 seeds and C. neoformans over A. niger alone in 12, with none failing both, so a reader that dropped either pair would not reproduce the printed count.

Ledger row MEL-ORDER-COUNT-01 records this, with one code citation, so N_CITATIONS goes from 163 to 164. The test does not check which seeds fail, since neither figure prints them, nor the range 42:297 printed beside the count in figure.jl.

Run locally with the system pytest and calibration and contract on PYTHONPATH, since no coupling/.venv exists in these worktrees: the new test, test_claims_citations.py and test_claims_ledger.py give 98 passed. The full calibration suite gives 659 passed, 5 failed and 5 skipped. The five failures are in test_guide_citations.py (two) and test_overdamped_regime.py (three), and they fail identically at e20478f without this change, checked in a scratch worktree. The skips are Dryad .nd2 datasets absent locally. The Julia, coupling and contract suites were not run; nothing they read is touched.

🤖 Generated with Claude Code

…nsemble CSV

Fig 2's label and WIDER in the melanin ensemble figure printed "fails in 19 of 256" as a
constant copied from SEED42-SCOPE-01 and read against nothing. test_ordering_count.py
computes it from radial_n40_p6_seeds42-297.csv (full arm, MCS 100, strict CS > CN > AN)
and requires both source strings and both PDF sidecars to print exactly that count, so a
re-run that moves it fails until both figures are edited and re-rendered.

Controls: a scratch CSV with one holding seed's CS and CN swapped, selected from the file
and asserted to differ at those two rows only, is reported at all four locations; each
location with its count cut out is refused. With the comparison weakened to report
nothing, all five controls fail. Ledger row MEL-ORDER-COUNT-01; N_CITATIONS 163 -> 164.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@chatgpt-codex-connector

Copy link
Copy Markdown

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.
To continue using code reviews, you can upgrade your account or add credits to your account and enable them for code reviews in your settings.

aurascoper and others added 2 commits November 13, 2026 20:46
… named producers

The loader counted whatever the CSV held: len(m) stood in for "of 256", a
duplicate (seed, species) row was silently overwritten by dict assignment, a
missing producer row died as a bare KeyError, and the 3>1>5 indices were tied
to their species names only by a comment, never by the CSV's own species_name
column. A regenerated ensemble over 256 wrong seeds, or a remapped species
index, would have counted as the same claim.

melanin_at_100 now refuses, naming the offender: a producer row whose
species_name is not what ORDER stands for, a duplicate (seed, species) row, a
seed set other than exactly 42..297, and a seed missing a producer row. Four
negative controls draw their known-bad from the committed CSV, show a
faithful copy still gives the committed answer, show the mutation applied and
at which lines, and require the named refusal. Each refusal was weakened to
trivially-true and its control re-run: three then accepted the known-bad
(DID NOT RAISE) and the fourth reproduced the bare KeyError the named
refusal replaces, so all four are load-bearing.

Addresses the completeness and species-mapping findings from PR #60's
external review; the reproducibility-versus-consistency finding stays
bounded out by the docstring, which now also records the widened scope.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…ut did not render

Two stale-record failures, both left by the 2026-09-24 manuscript
restructuring and failing in CI since eccf9ab, before PR #60's test commit.

The overdamped paragraph moved from lines 1244-1257 to 1259-1272 of
modeling_radioresistance_and_radiotropic_fitness.tex when ad50fca inserted
lines above it; the paragraph itself is unchanged, the content anchors and
every accounted literal re-verify, and the window follows it. Third entry in
the window's recorded history.

calculus_in_this_code.md was re-anchored by ad50fca to cite
radiodialysis_model.tex:132 and :180, and its .tex was synced in the same
commit, but the render step was skipped: the committed PDF still carried the
old in-manuscript line numbers and the .sha256 recorded the 955ab9b source.
Re-rendered through tools/render_guide.py, which verified all 19 citations
in the new PDF's text layer and only then rewrote the record, as its
inversion requires. The .sha256 was not edited by hand.

calibration suite after: 668 passed, 5 skipped, 0 failed.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant