Skip to content

Version 2.0: the radial ensemble, and the manuscript written to its answer - #61

Open
aurascoper wants to merge 9 commits into
fix/linear-rise-and-fig2from
claude/determined-mayer-noqhni
Open

aurascoper wants to merge 9 commits into
fix/linear-rise-and-fig2from
claude/determined-mayer-noqhni

Conversation

@aurascoper

Copy link
Copy Markdown
Owner

Stacked on #58 (fix/linear-rise-and-fig2 at 33c77e9), which this branch merges in, so the diff here is the paper work alone. This PR's head is 24cd3ec. PR #59 is stacked on this branch at 56d3908, and #60 on #59.

What the branch does, in commit order. 43b2381 registers the radial ensemble before any run: 256 seeds at N = 40 with six parcels per species, read at MCS 100, under four arms of the acceptance rule (shipped, direct term off, melanin production off, both off), with two primary tests declared in PREDICTION_radial_n40_p6.md and as constants in radial_analyse.jl. 29c8af2 restructures the manuscript around what the CPM runs, with appendices and a model table, and moves the radiodialysis model out to a companion paper. de8d88e runs the ensemble; the answer and its receipts are recorded in ledger row RAD-ENS-01. ad50fca writes the paper to that answer: results, abstract, conclusion and a new Figure 1 rendered from the committed CSV alone. 56d3908 adds a seeded consortium panel and video, the full model beside both terms off (RAD-FIG-02, RAD-VID-01).

Two commits were added on 2026-09-24 after review. 30f576b re-anchors the overdamped guard's window to lines 1259 to 1272. ad50fca had moved it to 1244 to 1257, fifteen lines short of where its own tex put the paragraph, and the anchor assertion refused it, failing three tests here and on every PR stacked on this branch. 24cd3ec qualifies the ensemble's null as holding at the shipped constants in the abstract, the section heading and the conclusion. It names the positive control that already exists, tests/deterministic_radiation.jl, whose amplified direct term moves C. neoformans inward and S. oneidensis outward on average over five seeds at a smaller configuration, and says the melanin path has no such control. It also corrects the abstract's list of executed simulations and the README's first lines, and adds calibration/tests/test_figure_palette.py, which requires the Python renderers' species colours to match FIG_COLORS in biofilms_potts.jl. That test reads both files as source text, because the renderer imports h5py and the calibration job has none, and both of its controls drive the same comparison the main test uses. The wording of 24cd3ec was approved by the author. Every number in it was checked against biofilms_potts.jl and tests/deterministic_radiation.jl.

Expected CI state. The calibration suite at 24cd3ec, run locally with no h5py as in CI, gives 658 passed, 2 failed and 5 skipped. The two failures are test_guide_citations.py on calculus_in_this_code: the guide's PDF and render record have to be regenerated with tools/render_guide.py on a machine where fontspec can find Latin Modern Roman by name, which the Mac cannot. That render is pending on the GPD and will land as its own commit. The skips are Dryad .nd2 datasets. The Julia, coupling and contract suites were not run locally for these two commits.

Open items, not in this PR yet: the guide render above; re-running the consortium frames from a clean tree so the manifest and caption stop recording a dirty build; and the shared document's checklist line that says the model does not express radiotropism, which the new conclusion makes stale and which no guard in this repository can reach.

🤖 Generated with Claude Code

claude and others added 8 commits September 24, 2026 18:40
…ed tests

The manuscript's only radiotropism result is seed 42 alone, and it says of
itself that the exact null must come from replicate simulations. This adds
the replicates as instrumentation, and runs nothing yet.

radial_sweep.jl runs the same seeds under four arms of the acceptance rule,
by zeroing CPMParams vectors only: full, beta_ion = 0, alpha_M = 0 (so the
melanin field stays 0 and dH_mel is 0 on every move), and both. Every arm
of a seed starts from the same placement, checked on every seed. It records
take_snapshot's per-parcel mean_r, and writes julia_version and Sys.MACHINE
into the header. With --reference, the full arm's melanin at the read MCS
must print identically to the committed sweep's for every shared seed, and
a reference sharing no seed is refused as a check that never ran.

radial_analyse.jl carries the two primary tests as constants: P1, under
full minus no_mel, C. neoformans and C. sphaerospermum move by a paired sign
test at alpha 0.05 AND |mean delta r/R| >= 0.05; P2, under full minus
no_rad, every species is equivalent within +-0.05 r/R by two one-sided
tests. PREDICTION_radial_n40_p6.md states the same thresholds, what had
been seen when it was written, and the predictions: P1 both inward, P2 all
equivalent. test_statistics.jl asserts the prose and the constants agree.

The sign test, paired difference and p formatter move from analyse.jl to
stats.jl so both analyses share one implementation.

Controls added, 75 assertions: the producer's header and column contract
through the real reader; an arm with production zeroed prints exactly 0;
the MCS 0 seeding frame is identical across arms; the arms diverge by MCS 4;
one digit of one reference value, selected from the artifact, makes the
producer refuse; a mismatched-configuration reference and a disjoint one
are refused; the analysis's paired means on a synthetic file; a mutation
of one value moves the paired mean by exactly delta/n and changes the
report; the reader's refusals.

Ledger row RAD-ENS-01 registers the question as needs_verification at the
moment of asking. N_CITATIONS moves to 161 for its one code citation.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AJo8eWDeh16NtQxUt1s7RG
…l table, radiodialysis out

The body now carries only executed work. Everything specified and not
executed moves, unchanged, into three appendices of the same file, because
every guard reads this one path and none follows \input: Appendix A holds
the PSDE, symplectic integration with its overdamped-regime paragraph, the
adaptive nutrient term, the kNN functional, the damage equation, the
mechanics and the planned feedback; Appendix B the per-species radiation
record; Appendix C the literature priors. The overdamped guard's window is
re-anchored to the paragraph's new lines in the same file.

The radial membrane-transport model leaves the PDF. Its section, its two
Methods paragraphs, its Results subsection with the membrane and
contaminant figures, and the five references cited only there move to
preprint/radiodialysis_model.tex, which builds standalone and is compiled
by the manuscript-build job so the withdrawal cannot leave it broken. A
stub subsection keeps the coinage disclaimer at the first surviving mention
of the word and says where the model went and why. The four figure files
stay on disk because ledger rows name them.

The parameter table is the executed model. Table tab:params lists every
CPMParams field and build_J_matrix entry in model units, with provenance
and a column saying whether each value reaches compute_delta_H_terms or
mcs_step! directly, through a field, or not at all; the melanin coupling
and seconds_per_mcs = NaN have rows. A new testset reads the table back
against the struct, the adhesion matrix, main_coupled's configuration and
the source of the reaches column, and its controls move one value, drop
one listed pair, and mislabel one reaches cell. The old longtable is
Appendix C as tab:priors.

Also: the two Langevin programs become one Methods paragraph; the
motivation moves from Chernobyl and Hanford to spent-fuel-pool biofilms
with the three references already in the bibliography; the title and
version line change; six ledger rows record where their subjects moved.
Every reference resolves, no numeral cross-reference remains, every
bibitem is cited, and the coinage, phase-locking, PP-RECIP-01, low-Z,
adaptive-feedback and ceiling-vocabulary guards pass.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AJo8eWDeh16NtQxUt1s7RG
256 seeds, four arms, N = 40, six parcels per species, read at MCS 100,
produced by Julia 1.12.7 on x86_64-conda-linux-gnu; the CSV header records
both. The reference check bound the full arm to the committed Mac-generated
melanin sweep: 112 of 112 shared rows print identically at six decimals.

Against the pre-registered tests in PREDICTION_radial_n40_p6.md:

P1 failed. Under full minus no_mel the paired mean delta r/R is +0.0001
for C. neoformans (129 of 256 seeds positive, p = 0.95) and -0.0004 for
C. sphaerospermum (125 of 256, p = 0.75). The prediction that the melanin
path moves both inward was wrong.

P2 held. Under full minus no_rad every species' 90% CI lies inside the
+-0.05 r/R bound; the widest is [-0.0029, -0.0006].

Exploratory: full minus null is within +-0.002 r/R for every species, with
within-run r above 0.98, so the four arms track one another. The drift
from the seeding frame (0.53 r/R) to about 0.51 r/R by MCS 100 appears in
the arm with both terms off, so it is adhesion-and-volume relaxation. The
C. neoformans minus B. subtilis separation is -0.006 r/R in every arm with
sd 0.11 over seeds; seed 42's +0.15 was one draw from that spread.

RAD-ENS-01 records the verdict and moves to keep; RM-KR-07, RM-KR-08 and
PP-65-02 to PP-65-05, which asked for these replicates, record them.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AJo8eWDeh16NtQxUt1s7RG
…on, Figure 1

The radial subsection now reports the ensemble in place of the single
seed: design, the two pre-registered tests, the result (P1 failed, P2
held), and what it establishes, with a reading of the mechanism marked
as a reading. The abstract and conclusion say that neither
radiation-derived term moves a parcel at this configuration and that the
prediction was wrong; the Discussion's single-seed sentence is replaced;
the OpenMC subsection says the ensemble runs under the analytic field and
the interface contributes no number. A version 2.0 note in Software and
Data names the four published numbers the ensemble bears on and says no
other published number is computed from a radial position. An Author
Contributions section is drafted under CRediT and marked for the authors
to confirm. This revision is prepared against de8d88e.

Figure 1 is the ensemble figure, rendered from the committed CSV alone by
radial_figure.py with matplotlib: per-arm means with intervals beside the
seeding frame, the two pre-registered contrasts against the bound, and the
per-seed separation with seed 42 marked. It writes the PDF, the pdftotext
sidecar, the sha256 and the PNG in that order, reads the retracted figure
vocabulary from its one source, and refuses a sidecar that carries any of
it. The old figure files stay on disk because ledger rows name them.

Also: README's Figure 1 paragraph, Key-results radial row, stale-figure
notice and preprint title; the prediction file's post-run record; the
ensemble README's answer; ledger row RAD-FIG-01; the overdamped guard's
window re-anchored again; and the two calculus-guide rows that cited
radiodialysis sentences now cite preprint/radiodialysis_model.tex, with
the guide's .tex synced by hand. The guide's render record still needs
tools/render_guide.py run where tectonic can fetch its bundle, which this
container cannot.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AJo8eWDeh16NtQxUt1s7RG
…ide both terms off

The manuscript had no picture of the model, and the animations in the
tree could not be used: the consortium GIF recorded no seed and was N = 20
with two parcels per species, and the 2024 GIF and MP4 files carry no run
record and predate the radiation-sign fix.

consortium_frames.jl runs one seed under the ensemble's arms through the
same project-free loader as radial_sweep.jl and writes every occupied
site per MCS as gzipped CSV, with a manifest naming the seed, the
configuration, the arms as CPMParams fields zeroed, the git commit, the
Julia version and the machine; it checks that the MCS 0 frame is one
frame across arms. Seed 42, N = 40, six parcels per species, full and
both-off, MCS 0 to 100, on the machine and tree that produced the
ensemble CSV.

tools/render_consortium.py draws the frames as tools/render_label_trajectory.py
draws voxels, in that file's palette, which is FIG_COLORS with white for
medium: a 2 by 3 panel at MCS 0, 50 and 100 for the paper, with sidecars
in the order the guards read them, and every frame of both arms side by
side as an MP4 for a supplement and a GIF for the README, assembled with
ffmpeg, each with a sha256 receipt. Reproducing them takes Julia, Python
and ffmpeg; no ParaView.

The figure joins the radial ensemble subsection with a caption saying
every reported number is the ensemble's. The README's front-page
animation is the seeded one; the unseeded GIF stays in assets/ and is
embedded nowhere. Ledger rows RAD-FIG-02 and RAD-VID-01.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AJo8eWDeh16NtQxUt1s7RG
ad50fca moved the window from 1159-1172 to 1244-1257 after adding lines above
the paragraph, but in ad50fca's own tex the paragraph starts at 1259. The
anchor assertion refused the window, and the three tests that read it failed
in CI on this branch and on every PR stacked on it. At 1259-1272 all 19 tests
in test_overdamped_regime.py pass. The window's history comment records the
miss.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… control, bind the palette

The abstract, the radial-ensemble heading and the conclusion say "at the
shipped constants" where they said neither radiation-derived term moves a
parcel. The abstract and results name the positive control that already
exists, tests/deterministic_radiation.jl: beta_ion = -20 for C. neoformans
and +20 for S. oneidensis (4e5 and about 270 times the shipped values), N =
24, one parcel per species, 60 MCS, five seeds, with the mean drift inward
and outward respectively. It covers the direct term only, and the text says
the melanin path has no positive control. The abstract states the null as
consistent with a coefficient too small to act at this configuration within
100 MCS, not as a measurement of one. The conclusion drops "at this
configuration the model does not express it" for the direct term moving
parcels along the dose gradient in the direction its sign sets when
amplified. The abstract's list of executed simulations is corrected. The tex
stays at 2017 lines, so the overdamped window at 1259-1272 holds.

README's first lines drop the stale branch name and the old title.

calibration/tests/test_figure_palette.py reads FIG_COLORS from
biofilms_potts.jl and COLOURS from tools/render_label_trajectory.py as source
text, since the renderer imports h5py and the calibration job has none, and
requires them to agree species by species through one function,
palette_mismatch. Both controls call that function on a copy with one colour
changed at an asserted index. With palette_mismatch weakened to return
nothing, both controls fail; with one colour changed in the real renderer
file, the main test fails.

Wording approved by the author. Content prepared in a separate review
session as paper-branch-content-fixes-v3.patch; numbers checked here against
biofilms_potts.jl and tests/deterministic_radiation.jl.

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.

ad50fca edited docs/guides/calculus_in_this_code.md and synced its .tex by hand,
but could not run tools/render_guide.py, so the committed PDF and the .md's
render record were stale and test_guide_citations.py failed on this branch and
on every PR stacked on it. Rendered with tools/render_guide.py on the GPD
(ok-cyberdeck, Linux x86_64, tectonic from ~/.local/bin), where fontconfig
resolves the preamble's "Latin Modern Roman" by name; macOS registers those
files only as LMRoman10 and the like, so the render fails there. The script
verified all 19 of the .md's citations in the new PDF's text before writing
the record, 1bb9953b for the .md.

The old and new PDFs embed the same eleven fonts (LMRoman 8/9/10/12,
LatinModernMath, DejaVuSansMono and their weights), have 8 pages each and
share the producer, so no fallback font entered. test_guide_citations.py:
5 passed.

Co-Authored-By: Claude Opus 5.5 <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.

2 participants