RESUME — merging completed work back to `main`

Written: 2026-08-09 · Branch when written: feat/revision1-validation-groupA @ a0b3887

This file exists so a crash, a closed laptop, or a new session loses nothing. Every step is copy-pasteable and has a verification command. Work top to bottom; tick the boxes as you go.

Latest session — 2026-09-02: v0.3.1 — patch bump for post-0.3.0 fixes/features

pyproject.toml and CITATION.cff (version + date-released) bumped 0.3.0 → 0.3.1. 37 commits had landed on main since the 0.3.0 tag without a version bump: compute_quality_flags() (per-observation reconstruction QC), lazy Zenodo auto-download for the WOA23 ocean_prop_ds NO₃ field, screening.py now defaults outlier detection to the published calibration domain instead of the user's own data, plus lockfile/env/CI/docs fixes and a README GDGT-2/3 notation fix. Still open: tag v0.3.1, push, cut a GitHub release, and decide whether PyPI / the software Zenodo DOI need a matching publish (data record and case-id files are unaffected — this bump is code-only).

Previous session — 2026-08-27 (desktop): v0.3.0 resubmission prep — the Zenodo swap is staged

The revised manuscript's Zenodo upload is prepared end to end. The data record's new version replaces the additive (GHEA) files with the T₀-shift refits under case-id names; the additive files stay on the v0.2.0 version of the record (Zenodo keeps every published version), and download.py still resolves their legacy names — pinned to that old version.

Done this session

Version → 0.3.0 (pyproject.toml). CITATION.cff fixed: cff-version 1.5.0 → 1.2.0 (1.5.0 is not a real CFF schema), doi now the software concept DOI 10.5281/zenodo.19671664 (it pointed at the data record), and a comment names the data concept DOI 10.5281/zenodo.19666744.
fit_predict added to MahalanobisOutlierDetector — Appendix C documents df["flagged"] = detector.fit_predict(df) and the method did not exist. Pinned by TestFitPredict in tests/test_screening.py.
download.py registry rewritten to the nine case ids (6× GHEB, 2× GHPU, GCDU). Legacy v0.2.0 entries kept but pinned via "record": to the old version and excluded from the no-argument download; tx.GHEA.* case ids alias onto them. ⚠️ ZENODO_RECORD_ID still says 20032542 — update it to the new draft's id the moment zenodo_upload.py prints it, before tagging (a draft's record id is final before publication).
scripts/prepare_resubmission_archive.py (new; replaces the deleted prepare_review_archive.sh, which staged legacy 050126 names that no longer exist). Stages review_archive_v0.3.0/: 9 forward posteriors, 6 compiled global-coretop reconstructions (b01–b07 concatenated; row order = coretop_maps_sites.csv), 40 paleo invT files (30 GHEB + 8 GHPU + the two Fig. 13 ud_draws), 4 data files, MANIFEST.csv (117 runs — the "archive manifest" of SI Text S3.3, warm-up lengths included), and a README written from Appendix C (this is the fix for R1C1's "I checked the Zenodo site and can't see this information"). Dry-run default; --apply verifies every .nc. Built and verified this session: 0.48 GB.
scripts/zenodo_upload.py retargeted: version/archive dir from pyproject, file list collected from the staged tree with count guards, invT zip name versioned.
scripts/add_retained_flag.py (new, applied): the R2R promised (R2C13) that retained warm-end samples are flagged in the archived calibration dataset — no flag column existed. The gridded CSV now carries mahalDist_TEXRI_cren3, mahal_outlier_090, warm_end_retained (27 warm-corner rows; ellipse fitted on the RAW sample-level coretops, lowAbundance-screened, G2/3 ≤ 5 — fitting on the gridded set itself gives a tighter ellipse and 150 false flags; one Arctic row sits marginally outside from gridding and is deliberately NOT flagged as retained).
Notebook cleanup: SI_code2_TEXAS_analysis and SI_code3_paleo_showcases (additive-era) → notebooks/superseded/ with a README; SI_code1_PreProcessing_finalized → SI_code00_PreProcessing so that "SI Code 1", which the SI cites for run_variance_partitioning(), resolves uniquely to SI_code01_t0shift_variance_partitioning. Live references (docs, skills, figure-sync agent) repointed; checkpoints and a scratch png removed. SI_code04 moved out of manuscripts/ (2026-08-27): its subject -- WAIC, 5-fold random and spatially blocked CV, elpd, Moran's I, paired bootstrap -- appears in neither main.tex, si_template_2019.tex, nor R2R_main.tex (verified by concept words AND signature numbers: 24.3, 0.750, 0.343, "elpd", 1298 all absent), and its two figures are in no SI. R2C4 was answered a different way -- everything labeled in-sample plus the 21-basin breakdown (Fig. S16, from SI_code02) -- and the letter says outright "not an out-of-sample validation ... we do not make any out-of-sample claims anywhere in the revised manuscript", which publishing the CV table would contradict. It now lives in notebooks/reviewer_response/ with a README recording why, so the Open Research claim "the notebooks that generate every figure in this manuscript" describes manuscripts/ exactly. SI_code02a figure numbers fixed: it wrote figS16_mcmcBudget_convergence_curves and figS17_mcmcBudget_heatmaps_bothArms, but the SI renumbered them +1 and includes figS17_/figS18_. The manuscript repo holds byte-identical copies under BOTH numbers, so the stale pair would have silently survived a re-run. Notebook now writes S17/S18; figS18_uncertainty_calibration (in no SI, and colliding with the real S18) became figS_uncertaintyCalibration in the notebook and in scripts/plot_uncertainty_calibration.py.
Figure folders cleaned in BOTH repos (2026-08-27). Manuscript repo: 22 byte-identical duplicates under superseded numbers deleted, 8 unique older versions moved to figures/*/superseded/, figures/README.md added; all 44 \includegraphics references across main / SI / track-change / initial-submission verified to resolve. TEXAS repo: finalized/ now holds exactly the 34 figures the paper prints, under the printed numbers; .svg masters moved to figures/manuscript/sources/, stale output to superseded/, scratch posterior_check.png deleted. The same stale-number bug was found in SI_code02, not just SI_code02a: it wrote figS15_basin_skill (printed S16), figS14_residual_difference (printed S15) and figS13_TEX86_..._t_sf2tc (printed S14). All corrected. Note the first cleanup pass archived two live figures because their names did not match the paper — the trap one level up; they were restored under the printed numbers. fig1 and fig2 are hand-finished from sources/*.svg, so re-running the notebook does not update what prints. Second pass completed the ladder: SI03_paleo_showcases_modelswitch → SI_code03_paleo_showcases and SI_code2a_… → SI_code02a_…, with every repo reference updated (including tests/test_manuscript_refit_config.py / test_param_sensitivity_config.py, which load the notebooks by path — both pass). The manuscripts folder now reads SI_code00 · 01 · 02 · 02a · 03. ⚠️ The text repos still say "SI_code2a" in a \note{} (revised main.tex ~755) and OVERLEAF-EDITS mentions the old SI03 name — margin notes only, but sweep them when editing there.
docs/preprint_additive_archive.md no longer promises the GHEB archive "at acceptance / v1.0.0" — it now says v0.3.0 and explains the version pinning.

The proplot -> ultraplot migration had not reached environment.yml

Found 2026-08-27 while verifying the locks before tagging. environment.yml still carried matplotlib<3.5, which was a proplot constraint. ultraplot 2.4 needs matplotlib>=3.9,<3.11, so the solver could not install it and fell back to ultraplot 1.0 — which pyproject.toml forbids (>=2.4.0) and which cannot even be imported against matplotlib 3.4.3 (matplotlib.cm.ColormapRegistry arrived in 3.5).

Confirmed inside the built image, not just theorised:

ultraplot metadata version: 1.0
matplotlib metadata version: 3.4.3
import ultraplot -> AttributeError: module 'matplotlib.cm' has no
                    attribute 'ColormapRegistry'

The SI notebooks guard with except ImportError, which does not catch an AttributeError, so a notebook cell crashes outright rather than degrading to plot = None. In other words the recommended Docker image could not run the SI notebooks, and texas-env on this machine has no ultraplot at all — the figures were in fact being drawn in base (ultraplot 2.4.0, matplotlib 3.10.9). uv.lock always had it right, because pyproject.toml sets no upper matplotlib bound; the conda pin was the anomaly.

Fix: matplotlib<3.5 -> matplotlib>=3.9,<3.11 and ultraplot -> ultraplot>=2.4.0 in environment.yml, re-solve conda-lock.yml, re-render the four platform locks, rebuild the image. This moves the shipped environment toward the one the figures were actually made in.

Done 2026-08-27. All four platforms now lock to matplotlib 3.10.9 + ultraplot 2.6.0, and the rebuilt image passes the acceptance test that the old one failed:

matplotlib : 3.10.9
ultraplot  : 2.6.0 -> import OK
uplt.subplots() OK

conda-lock is flaky on a cold cache. The four-platform conda-lock lock failed twice and then succeeded on the third attempt with no edits in between. Locking each platform singly worked every time, and plain mamba solved every platform's spec list directly. If it fails, retry before hunting for a dependency conflict.

Do not restore the matplotlib upper-bound-below-3.5 pin. It is only correct for proplot, which this project left behind.

CI: `actions/checkout` does not fetch Git LFS

The docs deploy failed on the v0.3.0 push with KeyError: 'model' from build_sampler_budget.py. Not a data problem: .gitattributes routes *.csv through LFS, and actions/checkout@v4 leaves pointer files unless given lfs: true. pandas then read version https://git-lfs.github.com/spec/v1 as the header row.

It could not be reproduced locally in any environment, because a working copy has the real files smudged in — the CI log (gh run view <id> --log, or gh api .../actions/jobs/<job>/logs) was the only way to see it. Reach for the log before theorising.

Fixed by lfs: true on the docs checkout. Only docs.yml needs it; tests, lint and the Docker build pass without LFS content. If another workflow ever reads a *.csv, *.nc, *.npz, *.pkl, *.h5, or *.xlsx, it needs the same flag.

v0.3.0 RELEASE STATUS (2026-08-27)

step	state
main pushed	✅ `f748db5` (5 later tooling commits still local)
tag `v0.3.0`	✅ pushed, points at `f748db5`
data record	✅ published: 10.5281/zenodo.22131367 (v0.3.0, 16 files, 516 MB)
concept DOI 19666744	✅ resolves to it — this is what the manuscript cites
v0.2.0 (additive)	✅ still live, still citable
PyPI	✅ texas-psm 0.3.0 live, cold-install verified
Docker image	✅ builds; GHCR publish fires on the tag
GitHub Release	⬜ not yet published — software DOI not minted (pre-release is fine: v0.2.1 was `prerelease=True` and still got a DOI)
push tooling commits	✅ pushed through `32d91bb`
CI (tests / lint / docs / docker)	✅ all green; docs site serves the revised API and concept DOI

Verified against the live record: download_posteriors(['tx.GHPU.sst.sri03.p0']) fetched 282,453 bytes; a PyPI install reports 0.3.0, carries both bundled posteriors and fit_predict, and compute_scaledRI reproduces the quickstart's committed values (0.477 / 0.540).

Zenodo API lessons, all learned the hard way today. The upload script had three separate defects: --publish re-ran the whole upload instead of publishing the existing draft; update_metadata round-tripped Zenodo's legacy metadata shape (creators: [{name, affiliation, orcid}], resource_type: {title, type}) into the InvenioRDM API, which silently dropped both fields; and PUT-ing the whole draft object back returned HTTP 500. Publishing also needs publisher, which the parent record never had. The required set for a Zenodo DOI is: resource_type, creators, title, publication_date, publisher. scripts/zenodo_fix_draft_metadata.py repairs a draft that is missing any of them.

Release sequence (order matters — the record id must be in the tag)

~~Commit everything.~~ Done 2026-08-27 — eight commits, tree clean.
~~Merge feat/revision1-validation-groupA → main.~~ Done 2026-08-27, LOCALLY: fast-forward, 175 commits ahead of origin/main, 0 behind, 302 tests pass. Currently checked out on main; the feature branch points at the same commit. NOT PUSHED — deliberate (decision 2026-08-27). The push waits until the Zenodo draft record id is in download.py (step 4), so main, the tagged release and the docs site tell one story with no window where the published docs name a record id that then changes. The push also fires .github/workflows/docs.yml, redeploying https://paleolipidrr.github.io/TEXAS/ — the URL the manuscript sends readers to, today still serving pre-revision content. When ready: git push origin main.
python scripts/prepare_resubmission_archive.py --apply (if not fresh).
ZENODO_TOKEN=… python scripts/zenodo_upload.py → draft; put the printed draft id into download.py::ZENODO_RECORD_ID; commit.
Tag v0.3.0, push tag → GitHub release → software Zenodo version DOI.
zenodo_upload.py --publish; verify TEXAS.download_posteriors() cold.
Manuscript: restore the Open Research availability paragraph (deleted in _draft — the track-change \remove{} at main-track-change.tex:1164–1177; SI still cites rattanasriampaipong_2026_ZenodoSoftware). Final wording agreed 2026-08-27 (see the pre-flight artifact, card B1): v0.3.0 accompanies the submission and the version of record will be v1.0.0 at publication — so the software bib entry should cite the CONCEPT DOI 10.5281/zenodo.19671664 (survives the bump; update only version= at proofs). Fix the bib: …_ZenodoData is a byte-copy of …_ZenodoSoftware — retype as @dataset, doi 10.5281/zenodo.19666744.
At acceptance: tag v1.0.0, publish v1.0.0 versions of both Zenodo records (re-run prepare_resubmission_archive.py + zenodo_upload.py), update version= in both bib entries and the prose at proofs.

Previous session — 2026-08-23 (desktop): the default calibration now ships, and the nitrate token is readable

Start here. Three things changed that a reader of the paper will touch on day one: a reconstruction no longer needs a download, N10 no longer reads as "nitrate = 10", and posterior attrs no longer carry two names for one model. Nothing on the science side moved.

1. The full multivariate T₀-shift calibration ships inside the wheel

predict_T_from_proxyObs with no fwd_posterior now uses tx.GHEB.sst.sri03.G23-N1p0 (or .thm. for temptype="thermoT"), and both ship in src/TEXAS/bundled_posteriors/. This is possible because the archived posterior's 81 MB is entirely the EIV per-site latents (true_*, 1513 sites x 4000 draws), which the inverse model never reads: dropping them gives 0.37 MB with every parameter the forward and inverse paths touch present and unmodified. scripts/make_bundled_posteriors.py rebuilds them; --check verifies.

Cache still beats the bundle in load_posterior, so a local refit is never shadowed.
*.nc is LFS-routed here, so .gitattributes exempts the bundled dir — an LFS pointer would install as a broken posterior for anyone who has not run git lfs pull.
Wheel and sdist verified to contain the real NetCDF (wheel 823 KB total).
New: a warning when a multivariate calibration is used without gdgt23ratio. Zero is not "off" for G23 — it asserts a ratio of zero and biases the reconstruction cold by ~γ_G23 x (true ratio) ≈ 0.6 °C per unit.

2. `N10` → `N1p0` (and `temptype=` is now genuinely optional)

N + cutoff x10 was misreadable in the one way that matters: 10 is also the documented value for switching the NO₃ correction off, so the token read as the opposite of what it means. encode_predictors writes N1p0 (p for the decimal point); N10 still parses and is never written.

191 cache files renamed by scripts/rename_cache_files.py (dry-run default, refuses on clash, --revert, .npz siblings move together). Per-machine — run it on the Windows box too.
Every read path is spelling-tolerant in both directions first: resolve_posterior_path normalizes the token on both sides and tries both spellings on exact-path lookups, and naming.swap_no3_token() covers invT leaves, which carry a site and scenario after the case and so never parse as a case id. Verified: old id → renamed file, and renamed id → old cache.
temptype was inferred by string-matching the posterior name for "thermoT", which case ids spell thm — so every thermocline reconstruction was silently labeled unknown_temptype. It now comes from the calibration's own attrs, and a contradicting temptype= warns.

3. One fact, one attr

Posteriors carried two model names (stan_model_name, plus arviz's echo of CmdStan's model = the same string + "_model") and two versions (texas_version, plus version = extract_and_update_metadata()'s own default "1.0.0", set by no caller and read by nothing). That duplication is exactly why stan_model_name could say _t0shift while model still said _boundedT. Both duplicates are gone at the writer, and generated_by no longer says culRI-Bayesian — the project's name from before it was TEXAS.

scripts/normalize_posterior_attrs.py fixed the 35 forward + 173 inverse files in place through netCDF4 append mode: draws never rewritten, verified byte-identical (md5 over every data var), +87 bytes of header on an 81 MB file. superseded/ is skipped — an archive that gets edited is not an archive.

⚠️ The dry run caught a hazard before it wrote anything. For inverse posteriors case_from_attrs cheerfully encodes the invT model name into a compset and proposes case_id = tx.GTDA..., a calibration that never existed. An invT run is a member of its parent case and never has one of its own; the script now refuses to compute one. Unguarded it would have stamped 173 files with invented provenance.

Cache audit (this machine, 2026-08-23)

	files	size
forward	33	1.89 GB
`superseded/` (archive)	33	1.80 GB
inverse	173	0.01 GB
halo archive + loose npz	10	0.53 GB

Five calibrations are stored twice and the copies are different fits — e.g. GHPU sst: legacy 2026-05-01 t₀ = 34.709 vs case 2026-08-12 t₀ = 34.446. Addressing is deterministic today (case id → case file, legacy name → legacy file), but this is the 5C collision, and SI_code3 still asks for the May copy.
35 inverse posteriors have no fwd_case — parent calibration unrecoverable. Unchanged recommendation: do not migrate, let them age out.
~1.9 GB (45%) is archive that nothing reads. Keep until after submission.

Also done

Appendix C2.4 drafted (/tmp/.../scratchpad/appendixC_C2.4.md, copy-paste LaTeX, one line per paragraph) with a compset table and the full multivariate as the recommended default, plus corrections to C2.1–C2.3: cren_rings → cren_weight, and the NO₃ lat/lon lookup needs the gridded field passed in (no3_dataset=) — it is not automatic, and it interpolates bilinearly rather than snapping to the nearest cell.
docs/index.md Step 2 rewritten around the bundled default; the stale "Zenodo multivariate posteriors are additive-EIV" warning replaced with an accurate note on the two formulations.
Streamlit: the prediction page defaulted to gen_logi_fixed_culmesocore_thermoT, a posterior that exists in no cache, so the app failed out of the box.
Fixed in passing: predict.py's module docstring passed fwd_posterior_name= to a function that has no such parameter.

⚠️ Still open

download_posteriors() cannot fetch any GHEB posterior — the registry holds only the five preprint-era files. The bundled pair removes this from the default path, but the .G23 / .N1p0 singles and the archival copies need the re-deposit. docs/index.md is worded to promise only what is true today.
The SI notebooks were not re-run. All 11 token edits were comments or docstrings, so no result changes; a re-run would re-sample every cached inverse (the invT cache key changed spelling) and rewrite tracked figures for nothing.
The lookup_no3_from_woa name and docstrings say WOA23, but its default variable no3_sf2tc_avg is the CMEMS field. Decide which is authoritative.

HANDOFF — 2026-09-08 (laptop): all five plans merged; repo is release-ready

Picked up the desktop handoff below and finished it. Read this first.

Where things stand

main is green: 602 passed, 24 skipped, ruff check . clean, jupyter-book build docs/ succeeds with no warnings, and the published site at https://paleolipidrr.github.io/TEXAS/ is current (gh-pages deploys main's tip; verified against the deployed HTML, not just the build log).

plan	state	PR
01 repo hygiene	merged	#25
02 kriged cache folder	merged	#27
03 dependency audit	merged	#29
04 package API cleanup	merged	#28
05 docs alignment	merged	#30

Follow-on work merged the same day: #31 (CLAUDE.md/RESUME.md American English), #32 (src/ docstrings), #33 (myst_heading_anchors 3→4), #34 (declare shapely, drop the invalid channel_priority key), #35 (thermal-only fallback), #36 (docs for #35).

The baseline is environment-dependent — 602/24 here, and that is correct

This laptop has esmpy 8.9.1, so two tests/test_regrid.py cases exercise the absent-esmpy fallback and skip instead of passing. The desktop, without esmpy, reports two more passes and two fewer skips. Same total. Both are right. CI's pytest (pip, pandas 3.x) job runs Python 3.12 and reports a third figure again, because seven tomllib-gated dependency guards execute there and skip on 3.10. None of these three numbers is the "true" one.

What changed beyond the plans

environment.yml no longer pip-installs texas-psm from PyPI. That entry made conda-lock pin the published wheel and bake its dependency metadata into conda-lock.yml, reintroducing plotly — which the audit had just removed — under a version number identical to the working tree's. Nothing about the version signaled that the lock described a different package than the source. Two guards now prevent its return.
predict_T_from_proxyObs no longer fails on a bare proxy. It selects the thermal-only calibration, which now ships in the wheel alongside the multivariate one (four bundled posteriors, was two), and warns that it is a different calibration rather than the multivariate one with its corrections off. A missing GDGT-2/3 ratio now raises like a missing NO₃ already did.
R2_thermal is a prior scale, and the posterior is insensitive to it. Measured, not argued: seven GHEB refits over the full 0–0.99 grid. Over the plausible 0.4–0.9 range t0 moves 1.1% of its own CI width and gamma_NO3 0.4%. Written up as Text S4 in the manuscript placeholder repo. See memory/project_r2_thermal_sensitivity.md.

Eleven defects were found in the plan text, not in execution

The desktop found ten; this session found eleven more of the same kind. The pattern is consistent and worth carrying forward: the plans' anchors and reasoning were sound, but essentially every count, cell index and line number in them had drifted. Verify each at the point of use.

The two that would have caused real damage:

Plan 03 Task 7 told us to delete import os from SI_code02a. It is not unused — os.path.join builds the export paths in three fig.savefig cells. Deleting it breaks manuscript figure generation. The tell was arithmetic: ruff reports 7 F401 findings, not the plan's 8, and os is absent from them.
Plan 03 Task 8 omitted import plotly.express as px from SI_code03. Not a lint problem — ruff does not flag it, because cell 69 rebinds px to a numpy array. But plotly had just been removed from both declaration files, so leaving the import would ModuleNotFoundError on a fresh env, and no gate in the plan would have caught it. Hence the installability cross-check added below.

A gate the plans did not have, and should have

The notebook verification was F401 (unused) and F821 (undefined) only. Neither sees an import of a package that is no longer installable. After the audit stripped plotly/geopy/requests/statsmodels/odrpack from both declaration files, any surviving import of them was a latent crash. The check is cheap — parse every top-level import in every notebook against the declared dependency set — and it is what caught defect 2 above. All seven notebooks pass it now.

Two decisions, resolved

A temperature-only default calibration — declined. The circularity objection was answered empirically first (see R2_thermal above), so the objection is not what killed it; the revision deadline and the judgment that it does not change the paper's main message did. Consequence: the wheel still ships seven Stan models, constraint_type stays out of the public API, and docs/stan_models.md is correct as written. The thermal-only posterior now ships and is auto-selected (#35) — that is a packaging convenience, not the GHEU model, which was never written.
Bayesian R²/RMSE for reviewer 3 — already done, better than proposed. 8ed4316 in the manuscript repo reports R² 0.803 [0.800, 0.806] and RMSE 0.0508, recomputed per draw at the observed predictors. The two cached quantities (R2_full 0.813, bayesR2_full 0.874) are both latent-predictor values and the manuscript correctly rejects them. Do not "fix" this by swapping in the cached numbers.

Still open

.github/workflows/docs.yml has no pull_request trigger. tests.yml is path-filtered to src//tests//pyproject.toml/environment.yml, so a docs-only PR runs no checks at all. #30, #33 and #36 each merged with zero CI. #36 exists only because the staleness was noticed by hand. Fix is a pull_request: block mirroring the existing paths: filter, deploy left push-only. (An agent sandbox blocks edits to .github/workflows/; this needs a human.)
dist/ holds a stale wheel — built 2026-08-27, 17 Stan models including archived ones, no univariate bundle. rm -rf dist/ && python -m build before any release.
v0.4.0 cut and published for resubmission (2026-09-09, superseding the 2026-09-08 "version stays 0.3.2 through review" decision). That decision was made without noticing that Zenodo had received neither v0.3.1 nor v0.3.2 — the concept DOI still resolved to v0.3.0, so the manuscript's software citation did not match the archive. Publishing was therefore forced by the citation, not optional. A minor bump rather than a patch because the [Unreleased] API removals are minor-release changes under this project's own stated convention. v1.0.0 remains deferred to acceptance. The Phase C checklist further down this file is also stale and should not be worked from: download.py already defaults to GHEB, ZENODO_RECORD_ID is already 22131367, the README/docs already label the GHEA rows as the preprint archive, and the model count is 7, not 8. At acceptance what remains is the bump, the tag, the GitHub release, and the Zenodo/PyPI publish. Never git push --tags here — docker.yml fires on every v* tag and would ship a GHCR image for an unreleased version.

Two deploys can race

#32 and #33 merged minutes apart; both docs deploys ran at 19:07 and both force-pushed gh-pages. One lost the ref lock and the run shows as failed. The build was fine and the later run published correctly. If you merge several docs PRs at once, expect this and check the last run rather than the failed one.

HANDOFF — 2026-09-08 (desktop): three of five finalization plans merged

Written on the Linux desktop for a laptop pickup. (Superseded as the entry point by the 2026-09-08 laptop section above — all five plans are now merged. The search traps and the plan-staleness warning below still apply.)

Where things stand

main is green: 583 passed, 15 skipped, ruff check . clean, only main and gh-pages on the remote.

plan	state	PR
01 repo hygiene	merged	#25
02 kriged cache folder	merged	#27
04 package API cleanup	merged	#28
03 dependency audit	not started	10 tasks
05 docs alignment	not started	7 tasks

Plan 04 was run before 03, reversing the index order. That was checked, not assumed: nothing 04 removes appears in any notebook, and 04 edits no notebooks.

The plans do NOT arrive with a pull

docs/superpowers/ is gitignored. Extract from the scratch branch without checking it out — checking it out would drag the tree backward:

git fetch origin scratch/superpowers-plans
git checkout origin/scratch/superpowers-plans -- docs/superpowers/

They land untracked, which is intended. Start from docs/superpowers/plans/2026-09-07-00-INDEX.md: it holds the execution order, the settled decisions, and every correction found so far.

To refresh after editing a plan: cut scratch/superpowers-plans from current main again, re-add, force-push, re-extract.

Build an environment before anything else

Do not reuse the name texas-env. A collision with an existing environment makes conda report a misleading matplotlib/matplotlib-base conflict that looks like a broken environment.yml. The file is fine.

conda env create -f environment.yml -n texas-env-laptop
conda run -n texas-env-laptop pip install -e . --no-deps
conda run -n texas-env-laptop pip install arviz pytest "jupyter-book<2" ruff build pyflakes

Measure the baseline before starting and record it. It is environment-dependent: where esmpy is present, two tests/test_regrid.py cases exercise the absent-esmpy fallback and skip instead of passing. Both counts are correct. Never use a bare python/pytest if a broken base conda env is on PATH.

Run the cache migration — it does not travel

data/cache/** is gitignored, so the laptop still has kriged grids loose in the cache root. The loader reads them from there with a printed note, so nothing breaks, but tidy it:

python scripts/migrate_kriged_cache.py            # dry run, read the plan
python scripts/migrate_kriged_cache.py --apply

It moves live grids, and renames rather than deletes a superseded grid that has no replacement.

Four search traps that cost real time here

A path built with / operators does not match a grep for the slashed form. One such reference broke a docs build silently.
--include=*.ext cannot see extensionless files. A Dockerfile copying a deleted directory survived three sweeps. Use git ls-files -z | xargs -0 grep -Iln <pattern>.
git check-ignore does not apply ignore rules to tracked files, so a keep-list check using it is vacuous. Use --no-index.
A name in a notebook's stored output is not a reference. Scanning code cells only is what found a seventh unused export.

Plan text goes stale — ten defects came from the plans, not from execution

Including stale line numbers that would have put an edit 65 lines off target, a premise about code that did not exist, and an instruction that would have permanently broken two tests. Verify every premise and every line number at the point of use. Implementers refusing a bad instruction is the system working.

Open decisions — mine, not the assistant's

A temperature-only default calibration. With no predictors, the T0-shift and additive parameterizations reduce to the identical curve, so the gap is the estimator, not the parameterization. EIV covers three things and the third — the proxy's own analytical error entering in quadrature — survives into a univariate model, so the existing GHPU posteriors are not a drop-in match. Unresolved circularity objection: a univariate EIV model would take R2_thermal from a univariate non-EIV fit, i.e. a fit prior-informing its own near-twin. Answer that before writing the model.
Bayesian R² / RMSE ranges for reviewer 3. Feasible from saved draws, no Stan change. But the median of per-draw R² is systematically lower than the point R², so swapping it into a comparison table against other groups' plug-in values would penalize TEXAS for a definitional change. Recommendation: keep point estimates in the comparison, add the interval alongside, ship the distributions in the .nc. Note the paired-RMSE table already reports lo/hi/p.

Smaller things left standing

docs/why_plugin_p50_differs.md still explains the truncated-prior models, which are now archive-only. Plan 05 deletes the page.
docs/stan_models_explanation_v2.md is corrected but plan 05 replaces it.
data_list_extreme_example_{boundedT,eiv}.pkl survive in the posterior cache; SI_code03 with MODEL_VARIANT="t0shift" would look for a _t0shift name.

Session — 2026-09-07: repo finalization, and six branches archived as tags

Six unmerged remote branches and one local backup were tagged, the tags pushed and verified on the remote, and only then were the branches deleted. Nothing is lost; every commit is reachable from its tag.

deleted branch	tag	tip	what it held
`claude/gridT-gui-exploratory`	`archived-branch/claude-gridT-gui-exploratory`	`8f29f091`	12 commits, ~1476 lines: `predict_grid.py`, a Streamlit page, 121 lines of tests. Its own commits mark it exploratory, "no merge without co-author sign-off", and its tip says the GUI has an unfixed cold-end bug. Its BAYSPAR work is already on `main` via PR #21.
`revision/boundedT-si-figures`	`archived-branch/revision-boundedT-si-figures`	`088ca463`	4 commits from July: SI figures and `.stan` files under the old `boundedT` name, superseded by the `t0shift` rename.
`claude/repo-audit-docs-update-od2dsj`	`archived-branch/claude-repo-audit-docs-update`	`c51bbb6d`	3 commits. Its genuine fixes were recovered onto `main` first — see the recovery commit. Left behind deliberately: its edits to `docs/stan_models_explanation_v2.md` (that page is being replaced), its 91 lines of session narration, and its `data/README.md` lines (superseded by the README sweep).
`claude/gridt-inversion-characterization-15i183`	`archived-branch/claude-gridt-inversion-characterization`	`82554fa3`	2 commits improving the gridT explainer at the old `TEXAS-revision/` path, which this session moved to `archive/exploratory/gridT-inversion/`.
`claude/bounded-t-model-revisions-idrogu`	`archived-branch/claude-bounded-t-model-revisions`	`883dec99`	`manuscript_revision/` Section 6-7 draft plus a handoff note, 226 lines. Archived rather than merged because manuscript text lives in its own repositories.
`claude/version-bump-v0-3-1-pxkvas`	`archived-branch/claude-version-bump-v0-3-1`	`e0050f46`	RESUME.md handoff notes only.
`backup/revision1-groupA-prerebase` (local)	`archived-branch/backup-revision1-groupA-prerebase`	`0ac01441`	The pre-rebase backup kept during the PR #15 work.

To recover any of it: git checkout -b <name> archived-branch/<tag>.

The local-only v0.2.6 tag was deleted, not pushed. Its commit is on main. .github/workflows/docker.yml fires on every v* tag, so pushing it would have published a GHCR image for a version that was never released. Confirmed absent from the remote before and after.

Never run git push --tags in this repository for that same reason.

Latest session — 2026-08-22 (desktop): the gridded CV finally reached Part 1

Start here. figS17 was not the only thing left behind by the ungridded → gridded CV switch: all of SI_code04 Part 1 was. The notebook is now executed top to bottom against the gridded exports and every Part 1 number has changed.

The scope was bigger than "re-run one cell"

121c294 staged the gridded cv_* exports and added a meta["dataset"] == "gridded" guard to the load cell, but nothing after that cell had been re-executed since 2026-08-13. Stored outputs in cells 7–18 were all still the n = 2043 run, and SI_table_model_comparison.{csv,tex} on disk were dated Aug 17. So the notebook's saved state, its two committed tables, and figS17 were a consistent picture of a superseded dataset.

	n = 2043 (was)	n = 1513 gridded (now)
in-sample R², additive / T₀-shift	0.730 / 0.745	0.796 / 0.803
spatial-CV R²	0.686 / 0.699	0.748 / 0.750
spatial-CV RMSE	0.0716 / 0.0701	0.0574 / 0.0572
spatial-CV cov95	0.923 / 0.925	0.904 / 0.899
Δelpd (T₀-shift − additive)	+68.4 ± 27.6 (2.5 SE)	+24.3 ± 11.6 (2.1 SE)
fitted-mean floor, additive / T₀-shift	0.008 / 0.375	0.343 / 0.420
in-sample → spatial RMSE penalty	+8.7%	+12.9%

⚠️ Two claims must be softened in the response to reviewers

R3.1's headline weakens. "The additive fitted mean reaches essentially zero" is gone: on the gridded set it reaches 0.343. The claim still holds — 0.343 is below the curve's own lower asymptote b ≈ 0.412, a value the generalized logistic cannot produce — but it is a ~0.07 violation, not a ~0.40 one. Gridding removes the dense clusters of near-duplicate sites that were dragging the additive mean down.
T₀-shift's predictive edge all but vanishes under spatial blocking. ΔRMSE was 0.0015; it is now 0.0002. The case for T₀-shift is boundedness and elpd, not accuracy. Say that.

REVIEWER_MAP.md §2.4 carried the old numbers verbatim and is updated; §3.1's _boundedT.stan paths and fig7/11/12/13/14 numbering were rename/renumber leftovers and are fixed too. PROVENANCE.md was already correct.

What changed in the notebook

figS17 now renders n = 1513 and the five per-fold n's sum to it (361 + 365 + 312 + 254 + 221).
The centroid assertion fired, as predicted. The five regions came back intact but under permuted fold indices (old 1↔3, 2↔4); the boxes themselves needed no change. Labels re-derived from the new centroids, not loosened.
Color is now keyed to the region name, not the fold index (COLOR dict, cell 16), so the palette and legend order are byte-for-byte what they were. Without this the permutation would have silently recoloured the map — the Part 3 prose about "the red block" and "the green one" would have gone wrong with no error anywhere. Both those observations still hold on the new figure.
Cell 29 was sk, a leftover that referenced a name cell 30 defines. The notebook had never run top-to-bottom; it does now.
Cell 14 is data-driven. It reads the floor out of meta and states the b-asymptote comparison instead of hardcoding "essentially zero".
Stale prose fixed in cells 0, 18, 19, 28, 33 — there is no "third n" any more (Part 1 and the production calibration are both 1513).

What is dirty and why

SI_code03_paleo_showcases.ipynb and fig10–fig13, figSI_variant_comparison_t0shift_vs_eiv.pdf were rewritten at 14:21–14:22 by a live VS Code kernel, not by this session — presumably the σ = 10 RUN_EXTREME re-run listed below. Left untouched; check them before committing.

SI_table_inverse_skill.csv and SI_table_residual_structure.csv (Part 2–3) re-emitted with last-digit float noise only — verified against the LFS blobs in HEAD. figS18 likewise re-rendered with no numeric change.

Resolved: the untracked fig8 pair

fig8_comparison_all_calibrations_proxy_residuals_maps_..._and_scatterplots.{pdf,png} was a superseded orphan. The stem appears in SI_code02 (cell 71) and SI_code2 (cell 69) only inside commented-out fname= lines; no live cell emits it. Moved to ~/.texas-superseded/20260822-fig8-comparison-orphan/ with a WHY.md. Never tracked, so nothing left git.

Still open

Push. Seven commits sit on feat/revision1-validation-groupA ahead of origin, plus this session's work. Nothing else has a copy.
σ = 10 extreme cases — see "Two edits waiting on a re-run" below; may be what the live kernel is doing right now.
Surface prior_mu_t / prior_sigma_t as invT .nc attrs before that re-run, so σ = 10 and σ = 15 reconstructions stop being indistinguishable on disk.

Session — 2026-08-21/22 (desktop): rename finished, Fig. S15 landed

Three sessions' worth of uncommitted work (2026-08-20, -21, -22) is now in. Written as a handoff across a PC restart. (Superseded as the entry point by the 2026-08-22 section above.)

commit	what
`732434a`	finish `boundedT` → `t0shift`; regenerate SI_code04's tables + figS18
`cefd058`	SI_code02's new Fig. S15, prediction skill by ocean basin
`d164355`	this file
`b51f9c4`	SI03 extreme cases: σ back to 10, Co1010 μ stays 5
`c1318b7`	`compute_scaledRI`: `cren_rings` → `cren_weight`

What actually changed

The rename reached the notebook switch keys. They were left alone on 2026-08-15 as "runtime branches, not artifacts". Wrong for SI_code03_paleo_showcases, which interpolates MODEL_VARIANT into figSI_variant_comparison_{variant}_vs_{other}.pdf and data_list_extreme_example_{variant}.pkl — the branch key is an artifact name. Fixed, and the two outputs renamed on disk.
A live bug fell out of it. run_coretop_maps.py already looked for coretop_maps_t0shift_manifest.csv, but only the boundedT spelling existed, so a resumed --arm bnd run had been finding no manifest at all.
SI_code04's three tables and figS18 are regenerated, closing the "Not done here" note in 121c294. Numbers are unchanged to float noise; the row label is not. TEXAS (T$_0$-shift, G23+NO$_3$) named a parameterization that the neighbouring TEXAS (univariate) row does not share — that row is GHPU, not GHEB — so the contrast the table draws is the predictor set, and it now says so.
Fig. S15 (figures/manuscript/revision1/figS15_basin_skill_SST_scaledRI_cren3_t0shift.*): RMSE by ocean basin, all three TEXAS arms as separate rows. Headline is uniformity, not a lower mean — across-basin SD 1.1–1.2 °C for TEXAS against 2.1–3.4 °C for the published forms, and no arm above 6.9 °C in any basin where every published calibration fails somewhere by ~10 °C. G23+NO₃ improves four basins, the two big ones being the Red Sea (6.92 → 2.13 °C) and the Arabian Sea (5.18 → 2.52 °C) — found without any basin being named to the model. Number is provisional (S19 ± 3); stays in revision1/ until _draft catches up, same gate as figS17/S18.
.gitignore now exempts figures/manuscript/revision1/*.csv, the same evidence carve-out 9c76045 gave data/revision1. Each figure writes a companion table; that table is where a reviewer checks the caption's numbers.
compute_scaledRI(cren_rings=) is now cren_weight= (c1318b7). The old name described half the parameter: the value is the coefficient on cren and cren' in the numerator and the constant the index is divided by, so it fixes the scale every sample is expressed on — which is why a posterior calibrated at 3 cannot read an index built at 4. "Rings" also implied a count of cyclic moieties, and neither 3 nor 4 is that count; they are conventions. The old keyword still works and warns, with a test pinning that both spellings return the same number. The cren_rings dict keys in run_param_sensitivity.py and SI_code02a are each file's own config keys, never reach the function, and were deliberately left alone.

⚠️ figS17 is stale — RESOLVED 2026-08-22, see the section above

It was not "the one thing left undone": all of Part 1 was stale with it.

figS17_spatial_cv_folds still renders n = 2043, the ungridded CV, while cv_sites.csv, cv_folds_map.csv, and cv_waic_meta.json are the gridded n = 1513 run that 121c294 staged. The commit message predicted the figure would need refreshing and it never was — the 2026-08-21 re-run only reached the table cells.

# in SI_code04_model_comparison_cv.ipynb, re-run the figS17 cell and confirm:
pdftotext -layout figures/manuscript/revision1/figS17_spatial_cv_folds.pdf - | head -3
# must say n = 1513, and the five per-fold n's must sum to 1513, not 2043

Note the assertion guarding the region labels: it refuses to plot if a fold moves. With a different site set the folds will move, so expect it to fire — that is the guard working. Re-derive the labels from the new centroids rather than loosening it.

Picking this up after the restart

Nothing is dirty and nothing is pushed. Five commits sit on feat/revision1-validation-groupA ahead of origin. A restart does not touch them, but nothing else has a copy — git push first.
The Jupyter kernels die with the restart. SI_code02 had been run end to end and SI03 was open; both notebooks are saved and committed, so only the live namespaces are lost. Re-running is the only way back.
docs/_build/ is gitignored and disappears if you clean; rebuild with the three commands under "Reading the docs before they are published" below.

Two edits waiting on a re-run

Neither is in effect yet — both change inputs that are cached:

σ = 10 for the extreme cases (b51f9c4). RUN_EXTREME is False and the cached reconstructions are all σ = 15: tx.GHE[AB].sst.sri03.G23-N10.inv.{ODP1259,Co1010}.ud.{nc,npz} and both data_list_extreme_example_*.pkl, built 2026-08-21 12:44–12:48. Set RUN_EXTREME = True, re-run cell 77, then redraw fig13 and figSI_variant_comparison — both are committed from the σ = 15 pickles.
figS17 is still the ungridded n = 2043 CV — see the warning above.

The invT .nc attrs do not record prior_mu_t / prior_sigma_t, and the filename does not vary with them either. Nothing on disk distinguishes a σ = 10 reconstruction from a σ = 15 one, which is why the two edits above have to be tracked by hand. Both values already reach build_fwd_data's output dict, so surfacing them as attrs is a small change — worth doing before the re-run, so the new files carry their own provenance.

Reading the docs before they are published

https://paleolipidrr.github.io/TEXAS/ is not stale — it is main. .github/workflows/docs.yml deploys on push: branches: [main] only, and this branch carries ~5,150 lines of docs main has never seen (rewritten index.md, callmap.md, sampler_budget.md, preprint_additive_archive.md).

python docs/_scripts/build_callmap.py
python docs/_scripts/build_sampler_budget.py
jupyter-book build docs/      # jupyter-book<2 is in pyproject's dev extras
# open docs/_build/html/index.html

⚠️ The second command regenerates docs/_static/sampler-budget.data.json from this machine's posterior cache — on 2026-08-21 it added 16 entries the committed snapshot lacks (sri05, GHEA). CI publishes from the committed snapshot, so revert that file after building unless you mean to update the published table.

To put the new docs live: merge to main. Do not reach for workflow_dispatch from this branch — it deploys whatever ref it runs from and force_orphan: true replaces gh-pages wholesale, so it would publish unmerged revision-1 docs to the public site.

Housekeeping done

figXX_basin_* (9 files) removed from figures/manuscript/revision1/ — superseded drafts of Fig. S15 under its pre-rename name. The _SST_ skill pair was byte-identical to the committed figS15_; the basin_RMSE pair and the t_sf2tc_avg variant are regenerable by re-running the cell with sel_temp_param switched. Never tracked, so nothing left git. Copies kept at ~/.texas-superseded/20260821-figXX-basin/ (moved out of /tmp before the 2026-08-22 restart).
Still untracked, unresolved: main-text/fig8_comparison_all_calibrations_proxy_residuals_maps_scaledRI_cren3_SST_and_scatterplots.{pdf,png} (2026-08-18). Today's full SI_code02 re-run did not rewrite them, so no live cell emits that name — either a superseded orphan of figS12 or a cell that no longer runs. Decide before submission; do not commit blind.

Session — 2026-08-17 (desktop): everything committed and pushed

The 2026-08-16 work below is now committed and pushed in all three repos. Nothing was left dirty as of that date. (Superseded as the entry point by the 2026-08-21 section above.)

repo	branch	head	what landed
TEXAS	`feat/revision1-validation-groupA`	`aac9c48`	figure renumbering (`2363c20`), `.gitignore` + 30 evidence files (`9c76045`), SI_code04 (`68cc2b6`), REVIEWER_MAP corrections (`3cd36b5`), figS17 region labels (`aac9c48`)
working-repo	`main`	`0d3cab0`	CV/WAIC exports made portable
placeholder	`main`	`e5dd0af`	ODR move + renumbering (`c2d0d89`), SI label realignment (`e5dd0af`)

What this session did beyond committing

REVIEWER_MAP.md had four stale entries, not one. 2.4 (CV never run) was the known one. Also stale: 2.3 / 1.5 (PETM priors listed as NEW — closed without a run on 2026-08-14) and, worst, 3.4, which still reported the retracted "intervals are ~14% too narrow" result from the 200-site stress set. 3.4 now opens with a RETRACTED box. The "needs new compute" list drops from four items to the screening pair.
figS17 is relabelled by ocean region, and two labels came out wider than the Part 3 prose had them: the red block reaches into the equatorial E. Pacific (10% of its sites west of 93°W) and the green one includes the equatorial E. Atlantic. Labels are derived from each fold's spherical centroid, and an assertion refuses to plot if a fold moves — a reseeded run fails loudly instead of mislabelling itself.
SI slots confirmed: S17 and S18 are correct, against a placeholder SI that now runs S1–S16 continuous. Still staged, not promoted — see below.
The placeholder's SI labels were off by one on five figures after the ODR move (two of them both starting fig:s11-). Renamed to match what they render as; clean si→main build gives 66 pp / 32 pp, zero undefined refs.

`_draft` has moved on, and has 7 unresolved refs

The author already fixed the main.tex:525 dangling ODR ref that the previous session flagged — that item is done, and _draft's SI has independently renumbered s11→s12 and s12→s13 exactly as the placeholder now does. But _draft is mid-transcription and currently does not resolve. Hands off, so this is an author list:

file:line	ref	fix
`main.tex:679,694,699`	`S-fig:s11-residual-plots-existing-models`	→ `s12-`
`main.tex:694,696,699`	`S-fig:s12-thermoT-residual-maps-multivariate`	→ `s13-`
`main.tex:701`	`S-fig:S13-residual-diff`	the residual-difference figure is not in `_draft`'s SI at all yet
`main.tex:713`, `si:M-`	`sec:regression-dilution`	label exists in neither file
`si:342`	`M-fig:ODR-regression-dilution`	ODR is in the SI now → local `fig:s11-ODR-regression-dilution`
`si:483`	`tab:multivariate_results`	label exists in neither file
`si:490`	`sec:secondary-environmental-effects`	lives in `main.tex:519` → needs the `M-` prefix

(M-... and M-<label> also show as unresolved; those are comments in the template, not real refs.)

Why figS17/S18 are not promoted yet: _draft's SI figures stop at S13. It has not received the residual-difference figure or the two MCMC-budget figures that occupy S14–S16 in the placeholder. Promote out of figures/manuscript/revision1/ only once _draft carries those; until then the number could still shift by up to three.

Still to do

The _draft ref table above (author).
Promote figS17/figS18 to finalized/supplementary/ once _draft catches up.
The two genuinely unrun items: Mahalanobis threshold sweep (R2C12) and the refit without the >= 0.75 retention rule (R2C13, also closes R3C3).
The [TO BE CONFIRMED BY THE AUTHORS] box in R3C3.
Optional: spatially blocked refit of TEXAS (~3.5 h).

Notebook churn to keep reverting. SI_code00_PreProcessing.ipynb and SI_code2_TEXAS_analysis.ipynb show kernel-metadata-only diffs (3.12.8 → 3.13.5) whenever they are opened. Not ours; reverted, do not commit.

~~Re-running SI_code04 dirties four files it did not change.~~ Superseded 2026-08-21 — do not revert these any more. The advice was right while the only difference was float noise. It stopped being right at 121c294, which changed the inverse-skill row label from TEXAS (T$_0$-shift, G23+NO$_3$) to TEXAS (T + G23 + NO$_3$) in the notebook but did not regenerate the tables. figS18 and the three SI_table_*.csv now carry a genuine label change; they were regenerated and committed in 732434a.

Session — 2026-08-16 (laptop): figure renumbering + model comparison

✅ Committed and pushed 2026-08-17 — see the section above. The warnings below about uncommitted state are historical.

1. Figure renumbering — the ODR figure moved to the SI

The author moved fig6_ODR_regression_dilution out of the main text into the SI, so main text is now 13 figures and everything after fig5 shifts down one. _draft is authoritative and the TEXAS notebooks already matched it; the placeholder repo did not, and was compiling with stale figures.

repo	what changed
TEXAS	`SI_code01` writes `supplementary/figS11_ODR_..._t0shift.pdf` (was `main-text/fig6_...`); `SI_code02` cell 67 → `figS12_comparison_all_...`; cell 85 `save_dir` → `supplementary`. Renamed the PDFs rather than re-executing (filename does not affect the render). Deleted 13 old-numbered `_t0shift` orphans — each verified as a different render from its replacement first.
placeholder	ODR lifted from `main.tex` into the SI as `figS11` (label `fig:s11-ODR-regression-dilution`); main text fig7→6 … fig14→13; SI figS11→12 … figS15→16; all files renamed; 11 stale figures refreshed from TEXAS, incl. Fig 7 which now carries σ_crtp = 0.0391.
`_draft`	untouched (hands-off)

Cross-references fixed (all consequences of the move): \ref{fig:ODR-...} ×2 in main.tex → S- prefixed; \ref{M-fig:ODR-...} ×3 in the SI → local; \ref{sec:secondary-environmental-effects} in the moved caption → M- prefixed; dropped a now-false "of the main text" in Text S2.

Both documents compile clean: 66 pp main, 32 pp SI, zero undefined refs.

xr-hyper is circular — build si→main from clean 2–3 times. A halted run leaves a truncated .aux and the next build dies with File ended within \read. rm -f *.aux *.log and redo the loop; it is not a content error.

⚠️ _draft has a dangling reference I could not fix. main.tex:525 still has \ref{fig:ODR-regression-dilution} for the G23 citation; that label now lives in the SI as fig:s11-ODR-regression-dilution, so it renders ??. The NO₃ citation on the same line was already repointed. Author must fix.

2. `.gitignore` boundary fixed (the one flagged "backwards, NOT yet fixed")

Blanket *.csv / *.log were swallowing SI evidence. Added negations scoped to data/revision1/** only. This un-ignores 17 previously-untracked evidence files, incl. mcmc_budget_grid.csv and proxy_definition_summary.csv. data/spreadsheets/ and data/cache/ remain ignored — verified. The inert figures/ # ← uncomment… line was not touched.

3. WAIC / cross-validation — found, made durable, and extended

It had already been run in full (2026-08-13, smoke:false, 5 folds, n=2043, ~1 h). REVIEWER_MAP.md still says it never was — that entry is stale, fix it.

Code lives in working-repo/TEXAS-revision/, under its own uv env (pandas 3). Do not re-run it under texas-env; uv.lock is the provenance record. results.pkl will not unpickle under TEXAS's pandas 2 (StringDtype), so it cannot be the record — added scripts/export_cv_results.py → CSV/JSON, plus .gitignore negations so run.log, manifest.json, and the four exports are tracked. Copied to TEXAS/data/revision1/groupA/model_comparison_cv/ with a PROVENANCE.md.

New notebook notebooks/reviewer_response/SI_code04_model_comparison_cv.ipynb — 43 cells, executed clean, reads only CSV/JSON, never samples. Three parts:

Part 1 — TEXAS variants. T₀-shift beats additive on every metric; Δelpd = +68.4 ± 27.6 (2.5 SE). Spatial blocking degrades both (R² 0.743→0.699, cov95 0.944→0.925). Boundedness for R3.1: additive fitted mean reaches 0.008, T₀-shift floor 0.375. → figS17_spatial_cv_folds
Part 2 — vs existing calibrations, in °C (n=1298; 215 sites dropped for unmatched/ambiguous TEX₈₆ joins). TEXAS is in-sample here, the others are not. → figS18_inverse_skill_by_block
Part 3 — significance + residual structure. The key result.

model	RMSE °C	Moran's I	bias SD	bias range
TEXAS multivariate	4.38	0.423	0.86	2.25
TEXAS univariate	3.88	0.579	1.54	3.81
BAYSPAR	4.34	0.650	2.66	7.29
Schouten02	5.34	0.769	3.23	9.33
Kim2010 TEX₈₆^H	5.99	0.810	3.81	10.41

The argument for the paper (author's framing, now quantified): multivariate TEXAS and BAYSPAR are statistically indistinguishable on RMSE (Δ = +0.041 °C, p = 0.786) — so claim neither outperformance nor concede underperformance. Univariate TEXAS significantly beats BAYSPAR (−0.456 °C, p = 0.001) — safe to claim. And the residual-structure ordering is monotonic in how much mechanism each model represents, the reverse of the RMSE ranking: the multivariate model converts systematic regional error into random error. ΔMoran's I is robust (vs univariate −0.156 [−0.185, −0.128]; vs BAYSPAR −0.227 [−0.277, −0.179]). BAYSPAR's competitive global RMSE conceals a +5.7 °C Arctic-block bias — the signature of empirical regional recalibration. Southern Ocean confirms the mechanism: NO₃ 15.5 vs 2.3 µmol/L global, only 13.9% below cutoff, and multivariate improves bias (−2.34 vs −2.66) while worsening RMSE — variance, not bias. Pair with figS14.

Three different n in play — always state which: 2043 (Part 1 CV), 1513 (production), 1298 (Part 2/3 comparison).

figS17 and figS18 use DIFFERENT partitions — same construction (k-means on unit vectors, seed 20260727, k=5) but n=2043 vs n=1298, so fold indices do not correspond. The five geographic regions do line up. Part 3's markdown says so; figS17's legend still reads "Fold 1…5" and should be relabelled by region.

Still to do — all closed 2026-08-17 except the refit

Commit + push all three repos (TEXAS, working-repo, placeholder).
Fix the stale REVIEWER_MAP.md entry — four were stale, not one.
Relabel figS17 legend by region.
Confirm SI slots for figS17/figS18 — S17/S18 are correct; still staged.
_draft dangling \ref{fig:ODR-regression-dilution} — the author fixed it, but _draft has seven other unresolved refs now; see above.
Optional, not needed for the argument: spatially blocked refit of TEXAS (~3.5 h) so its side is genuinely held out. The 8.7% in-sample→spatial penalty from Part 1 currently stands in as a bound.
Unrelated churn in the tree: SI_code00_PreProcessing.ipynb and SI_code2_TEXAS_analysis.ipynb kernel-metadata diffs — reverted, not committed.

Latest session — 2026-08-15 (evening): rename finished, docs synced, all pushed

Pick up from the laptop — read this first.

What is pushed where (everything below is on the remotes)

repo	branch	head	contents
TEXAS	`feat/revision1-validation-groupA`	`72871376`	`boundedT`→`t0shift` rename (`6379e4de`), docs sync (`58ff1ef7`), case-id-first docs + naming explainer (`1baea32c`), preprint β archive page (`da76808d`), compset-B re-gloss (`fab01622`), re-executed notebooks + σ in panel (b) (`72871376`)
`AGU_PALO_TEXAS_PSM_revised_submission`	`main`	`0874a02`	"T₀-shift" terminology + figure renames; compiles clean
`R2R-report-...2026PA005459`	`main`	`e6c307e`	names the T₀-shift parameterization in R3C1 + the Bayesian-R² remark

The rename, in one paragraph

The variant token is now t0shift everywhere it names an artifact (Stan files, all 30+ figure files, SI_code01/02_t0shift_* notebooks, fit_t0shift_single_predictors.py); the manuscript term is the "T₀-shift parameterization". encode_compset maps both spellings to compset letter B — re-glossed as bounded-by-construction (user decision: letter names the property, paper names the mechanism; never migrate GHEB ids). Nothing in data/cache/ was renamed; legacy .nc/.pkl names and old stan_model_name attrs resolve unchanged. Notebook switch keys (MODEL_VARIANT = "boundedT") intentionally keep the old spelling. The Python curve function is generalized_logistic_fixed_upper_t0shift (old _bounded_t name aliased).

Still to do (laptop)

Re-run the two bottom-layer calibration cells in SI_code02_t0shift_TEXAS_analysis.ipynb (SST + thermoT): panel (b) now lists sigma_proxyObs_crtp (σ_crtp, coretop color, stack starts at y=0.31) but the committed figures predate the edit.
Re-run the fig10/fig11 cells in SI_code03_paleo_showcases.ipynb and eyeball the new code-drawn annotations (site_annot_dict, elbow leaders) against the hand-revised PDFs in ~/Downloads/*_revised.pdf; nudge xy/xytext as needed. The axs[-1,1] SubplotGrid bug that ate the age ticks is fixed (bottom_ax = axs[n_sites], compare with is).
_draft repo untouched (hands-off): still references _boundedT figure names; author transcribes from the placeholder, then its figures/ copies need the same _boundedT→_t0shift rename.
Public docs site updates only when this branch merges to main (gh-pages builds from main). New docs: preprint_additive_archive.md (Archive section), naming explainer in index.md + quickstart_demo.

Laptop environment caveats

LFS budget: clone/pull with GIT_LFS_SKIP_SMUDGE=1, then git lfs checkout selectively (see the LFS memory note). The rename commits carry no LFS objects (figures are plain git).
data/cache does not travel. The laptop needs the 2026-08-14 GHEB refits (tx.GHEB.{sst,thm}.sri03.G23-N10.fwd.nc) to re-run SI_code02 — restore from the LaCie sidecar (texas-sidecar-20260814.tar, see below) or refit. Two fits of the SST case exist on this desktop (legacy-named July 29 + case-named Aug 14); differences are MCMC noise only.
Renamed Stan models recompile on first use.

Latest session — 2026-08-15: `boundedT` → `t0shift` rename

The manuscript now calls the model the T₀-shift parameterization (the "bounded-T" name invited the question which temperature is bounded? — the bound is on the response). The code/figure token was renamed to match:

Stan files git mv'd to ..._priorApprox_eiv_t0shift.stan and invT_..._unconstrained_t0shift.stan; binaries recompile on next use.
All 30 figure files *_boundedT.* → *_t0shift.*; notebooks emit the new names (FIG_TAG/fig_tag = _t0shift).
Notebooks renamed: SI_code01_t0shift_variance_partitioning.ipynb, SI_code02_t0shift_TEXAS_analysis.ipynb. scripts/fit_t0shift_single_predictors.py.
Backward compatible: encode_compset maps both spellings to B; nothing in data/cache/ was renamed; notebook variant keys (MODEL_VARIANT = "boundedT") intentionally unchanged.
Manuscript placeholder + R2R updated to "T₀-shift" wording and the new figure names. _draft untouched — its tex still references _boundedT figure names and its own figures/ copies still carry them, so it stays internally consistent until the author transcribes.
Historical entries below this one keep the old spelling on purpose — they record what was true when written.

Latest session — 2026-08-14 (evening): the R2R is fully drafted

All 31 R2R responses are now written. The 15 [RESPONSES] placeholders are filled and main.tex compiles clean under xelatex (0 errors, 36 pages, was 28). Committed as 27a7ab5 in the R2R repo.

Three responses rest on evidence extracted for the first time this session:

R2C14 — the reviewer is right that coretops do not constrain the warm end, and more strongly than they put it. All 28 observations above the warmest coretop (29.80 °C) are culture (24) or mesocosm (4); not one coretop exceeds it. T₀ = 34.8 °C lies above the entire coretop range and the max-slope temperature (29.85 °C) sits at its very top edge. Also clarified: the upper asymptote is fixed at 1 by construction, not estimated — what culture and mesocosm constrain is the approach to it.
R2C13 — the >= 0.75 retention rule keeps 39 samples (1.9%) the 90% ellipse would drop, mean SST 27.7 °C against 16.3 °C for the rest. The warm-end leverage R2 suspected is real. (Reproduced on the processed coretop table, n=2024; the production chain grids to 1513, so treat as indicative.)
R1C5 / R2C3 / R2C18 — the PETM is demonstrably not prior-driven, so the planned prior-sensitivity refit is no longer needed. See below.

The PETM prior question is CLOSED — no run required

Every reconstruction used prior_sigma_t = 10 (verified in SI03 cell 61 for the PETM block and cell 77 for the extreme cases — not 15). Because prior and posterior share a scale, the posterior/prior variance ratio bounds the prior's contribution directly:

record	68% half-width	post/prior var	variance from data
South Dover Bridge (PETM)	2.2 °C	0.05	95%
ODP 959 (PETM)	3.8–4.1 °C	0.14–0.16	84–86%
ODP 1259 (upper asymptote)	5.4 °C	0.29	71%
Co1010 (Antarctic, lower asymptote)	5.9 °C	0.35	65%

A prior 4–20× wider than the posterior it produces is not setting the answer. The same table independently confirms the asymptote argument (prior influence rises monotonically toward both asymptotes) and confirms R1C5's Antarctic criticism — Co1010 is the most prior-influenced record in the study.

Nitrate-scenario sensitivity at the PETM sites is also second-order: −1.34 °C (SDB), −0.96 °C (ODP 959), both smaller than a single run's 68% half-width — against −4.5 to −5.3 °C at the Quaternary sites.

CORRECTION: the "intervals are 14% too narrow" result was wrong

It came from invt_budget_sites.csv, the 200-site stress set (tail-weighted by construction, and not held out). Over all 1513 sites the production calibration is calibrated and the bias flips sign — see the corrected entry further down this file. Population: bias −0.99 °C, RMSE 4.35, R² 0.824, cov68 0.664 vs 0.68 nominal. Univariate is better in the inverse direction (RMSE 3.87, R² 0.860), which is the evidence behind R2C4 and R1C2.

Section 6.2 / Section 7 drafting (2026-08-14)

§6.2's new closing paragraph (replacing the dissolved §6.3): refined draft in scratchpad/sec62_para3.tex. The fix was that it estimated β coefficients and then pointed at §7 for "posterior coefficients" — but §7 reports γ, a different quantity in different units. It now says the two parameterizations are not numerically comparable and that only the qualitative result carries forward. Verified: thermal R² 0.75, both G₂/₃ slopes, both NO₃ slopes, n=1513, and n=562 for no3_sf2tc_avg ≤ 1.0 all check out; divergence factor 2.48.
§7's sampler paragraph: corrected draft in scratchpad/sec7_sampler_para.tex. Four numbers were wrong — the budget sweep is Text S2 not S3; forward runs take 9.8–212.0 s not 11–186; agreement is within 0.06 posterior SD at the production 400/1000 cell (0.0677 belongs to a different budget, bounded-T's recommended 300/900); and median ESS across the seven fits is 1247.6, not 1341 (1340.9 is one fit's value). Also narrowed "All model fits" → "All calibrations reported here", because the SRI05 univariate comparator has R̂ = 1.01407.

Where M=300 came from — do not claim the sweep recommended it. It did not: recommended_invt_budget.json recommends 1000/1000 with M=500. Production runs 500/1000 with M=300, i.e. cheaper than the recommendation. The reason the recommender returned the reference cell is that its drift gate is the seed-to-seed floor (0.271 °C), measured by rerunning the reference at a new seed — but any cheaper cell carries a budget effect plus seed noise, so it essentially cannot come in under a pure-noise threshold. The gate is close to unpassable by construction, so "no cheaper cell passed" is not evidence that M=300 is inadequate. What the grid actually shows: across all eight cells bias 0.921–0.934, RMSE 4.511–4.529, cov68 0.590–0.610. The production cell drifts 0.339 °C against the 0.271 °C floor — ~13× smaller than the 4.5 °C RMSE. Defensible claim: the budget moves medians by ~0.3 °C, comparable to seed noise and an order of magnitude below the reconstruction uncertainty. Avoid "more than sufficient" for the inverse.

TWO manuscript repos — know which one you are in (2026-08-15)

repo	role	edit it?
`~/Documents/GitHub/AGU_PALO_TEXAS_PSM_draft`	the main project. `main.tex` and `si_template_2019.tex` are edited by the author, with active track changes; Overleaf-synced.	NO — hands off those two files
`~/Documents/GitHub/AGU_PALO_TEXAS_PSM_revised_submission`	working placeholder for Claude's text edits	yes

The names differ by one word, so check git remote -v before editing. An edit in the placeholder is a drafting aid and does not reach the submission — the author transcribes what they want into _draft. Editing _draft directly would collide with the author's track-changes work and with the Overleaf sync.

The manuscript work recorded in this file (the §6.2/§7 paragraphs, the Text S2/S3 restructure, figS14–S15) was committed to the placeholder repo (fcbed66). It still has to be carried into _draft by hand.

THE SIDECAR LIVES ON THE LaCie DRIVE (2026-08-15)

/media/rrattan/LaCie/Postdoc/TEXAS-PSM/texas-psm-export/
    texas-sidecar-20260814.tar     3.7 GB, 557 members
    texas-sidecar.md5              df997b86506723d41e860380bc8b3cc5

Verified with an O_DIRECT read back from the device, not from page cache — a cp of this size completes in seconds against cache and tells you nothing, so if you ever re-copy it, verify with dd if=<file> iflag=direct bs=4M | md5sum rather than a plain md5sum -c.

Local copies and the Google Drive parts have both been deleted — the LaCie drive is now the only copy outside git. On the desktop: clone, git lfs pull, check the --skip smudge filter, then untar this over the clone.

Text S numbering — the R2R was wrong and is now fixed (2026-08-15). The SI numbers them Text S2 = Frequentist Attenuation-Bias Analysis and Text S3 = MCMC Sampling Budget Sensitivity. The R2R had both swapped in all five places and was corrected in 589f19b. Anything written earlier in this file calling the budget sweep "Text S2" is wrong — it is Text S3.

Moving this work to another machine (the zip question)

Do not zip the whole repo. The working tree is 4.9 GB before .git (7.8 GB) and .venv (1.9 GB), and it contains 6 compiled Linux Stan binaries in src/TEXAS/stan_models/ that are actively harmful elsewhere (TEXAS detects the exit-127 cross-environment failure and recompiles, but only after confusion).

Clone + git lfs pull already carries everything tracked. What git does not carry is 459 files under data/, 3.9 GB:

what	size	needed?
`data/cache/TEXAS_posterior_cache/`	3.5 G	yes — the expensive thing
`data/cache/TEXAS_invT_posterior_cache/`	18 M	yes — 173 `.nc` + 175 `.npz`
`data/revision1/**` untracked (18 files)	<1 M	yes — SI evidence tables
`data/spreadsheets/` untracked (9 files)	small	yes — incl. `predT_bayspar_p16_p50_p84.csv`, which SI_code02 loads
`data/cache/kriged_grids_*.npz`	332 M	optional, regenerable (slow)
`data/cache/superseded_halo_cache/`	111 M	no

Build the sidecar (~3.5 GB; .nc/.npz are already compressed, so -z buys little and costs a lot of time):

cd /home/rrattan/Documents/GitHub/TEXAS
tar -cf ../texas-sidecar-$(date +%Y%m%d).tar \
    --exclude='data/cache/superseded_halo_cache' \
    --exclude='data/cache/kriged_grids_*' \
    data/cache data/revision1 data/spreadsheets

On the other machine: git clone, git checkout feat/revision1-validation-groupA, git lfs pull, check the --skip smudge filter (see below), then untar over the clone. Tracked files are overwritten with identical content, so the order does not matter. Do not copy src/TEXAS/stan_models/ binaries — let them recompile.

Do NOT push this to Google Drive through the gvfs mount (tested 2026-08-14)

The GNOME/gvfs Drive backend (/run/user/1000/gvfs/google-drive:host=...) is unfit for multi-GB transfers. Measured over an hour: a 64 MB probe ran at 15.4 MB/s, but sustained throughput collapsed to 1–2 MB/s, a direct cp of the 3.7 GB tar failed outright with Input/output error, and the backend wedged twice after roughly 1–1.5 GB — the second time taking the mount down with it. Only 3 of 8 512 MB parts landed in ~55 minutes.

Two traps if you ever debug this:

An in-progress upload is listed under its Drive file ID, not its filename. stat on the destination filename reads 0 while the transfer is actually running, so it looks stalled when it is not. Check du -sh on the folder instead. (I killed a healthy upload at 49% believing it was hung.)

killing gvfsd-google does not respawn it — it removes the mount. Recover with gio mount "google-drive://<account>@gmail.com/", which does not re-prompt for auth. Restarting the backend is what cleared the wedge, so if you must persist, remount proactively between parts rather than reactively.

Use the browser instead — drag the tar into Drive. It is resumable, chunked, and runs at your full uplink. Or install rclone (needs a one-time OAuth in a browser, so it cannot be set up unattended). Splitting is only worth it as a hedge against a transport with no resume:
split -b 512M -d -a 2 texas-sidecar-YYYYMMDD.tar texas-sidecar-YYYYMMDD.tar.part
cat texas-sidecar-*.tar.part* > texas-sidecar.tar   # reassemble
md5sum -c texas-sidecar.md5                         # verify BEFORE untarring
Verified 2026-08-14 that the split/rejoin round-trip reproduces the original md5 exactly, so parts are safe to rely on if you do go that route.

Still open

.gitignore boundary is backwards in one place, NOT yet fixed. A blanket *.csv ignores SI evidence (mcmc_budget_grid.csv, the Text S2 sweep; proxy_definition_summary.csv) while off-scope material is tracked: TEXAS-revision/ (9 files — gridT poster/explainer, zero references from main text, SI, docs, or package) and notebooks/current/IMOG_presentation.ipynb. Also note proxy_parameter_comparison_by_arm.csv is tracked but its sibling proxy_parameter_comparison.csv is not — almost certainly accidental.

⚠️ .gitignore:89 reads figures/ # ← uncomment if you want... — it is inert only because the trailing text makes the pattern match nothing. "Tidying" that comment would ignore all 92 tracked manuscript figures.
Two R2R red boxes remain, both genuinely unrun: the Mahalanobis threshold sweep (R2C12) and the refit without the >= 0.75 retention rule (R2C13). The latter also closes R3C3's screening-robustness request. The PETM prior box is gone — answered above.
The [TO BE CONFIRMED BY THE AUTHORS] box in R3C3 is still for you.

Latest session — 2026-08-14

Everything below is pushed. Branch feat/revision1-validation-groupA is level with origin; the manuscript and R2R repos are level with their main.

The manuscript restructured Sections 6 and 7. Eq. 10 (the multivariate forward model) no longer exists in §6.2. The bounded-T model is introduced once, in §7.1.2, as Eq. 13 — and the two duplicate statements of it that used to sit there (eq:bottomlayer-general, eq:bottomlayer-full) were folded into it. §6 is now confined to the S-curve, the temperature-only fit, and the evidence that the nonthermal effects exist. Equation numbers downstream of §6.2 all shifted; the R2R's noise-term table was updated to match (Eq. 11 → 10, Eq. 14 → 13). If you cite an equation number from memory, re-check it.

New Appendix C, "Running TEXAS on your own data" — answers R1C1's request for a practical "list of ingredients". Two facts in it were verified against the package, not assumed: compute_scaledRI is scale-invariant, so peak areas and fractional abundances give identical results and no normalization is needed; and nothing in the package consumes an age model, so chronological uncertainty is not propagated. The Open Research Section was trimmed to availability only.

arydshln removed from the manuscript preamble. It was loaded ahead of array/colortbl, which broke \hline, \midrule and \toprule alike — LaTeX did not fail cleanly, it ground for >10 min in an error cascade. Nothing in the document uses dashed rules. If a table ever stops compiling, check this first.

New notebook: SI_code01_boundedT_variance_partitioning.ipynb — the frequentist additive-vs-bounded-T comparison, one fitter with parameterization= and beta_fit= switches. Reproduces the submitted additive numbers and settles the fig6 NO₃ ODR question (see below).

Explainer: working-repo/TEXAS-revision/boundedT-explainer.html — plain-language derivation of the bounded-T model from the production posterior, written for the §6–7 rewrite.

The sampler budget is now recorded and reported. sampler.py stamps iter_warmup, iter_sampling, chains, and thin on every new posterior, read from the CmdStan fit so defaults are captured too. scripts/backfill_iter_warmup.py retrofits existing files only where a refit manifest names the specific file — a blanket stamp would be wrong, because the cache mixes 400 (refit script), 300 (SI_code02_boundedT) and CmdStan's default 1000, all indistinguishable in the file. 11 forward posteriors stamped and verified; the 173 inverse posteriors are deliberately left unstamped, since the manifest records no path for inverse runs. §7 now states the budget, and Text S2 in the SI carries the sweep.

Two more manuscript edits landed: the Conclusions now says explicitly that R² 0.75 → 0.80 is in-sample, and gives the cross-validated counterparts (0.74 random, 0.70 spatially blocked). R3C3's robustness half is drafted from the CV run and SI_code02a.

Section 6.3's methodology moved to Text S3; §6.3 now carries the argument and ends on a prediction §7.1.2 tests. Bounded-T is presented as the formulation, not as the winner of a comparison — the additive model appears nowhere in the manuscript or SI, only in the R2R. §7.2's results were still in β from the additive fit and are now γ from the production posterior, with 68% and 95% intervals.

R² convention decided 2026-08-14: report R2_full (1 − RSS/TSS), because published TEX₈₆ calibrations are mostly not Bayesian and report that quantity. Univariate 0.746, multivariate 0.813 — an internally consistent pair. Bayesian R² (0.874) is named once, cited to Gelman et al. 2019, and never quoted interchangeably. The Conclusions moved 0.80 → 0.81 to match.

The length argument is now on the record. The Introduction closes with a roadmap naming each section; the R2R has a full response to the Associate Editor (there was none before) arguing that the length is intrinsic to a methods paper making two structural departures, and that the answer is navigability rather than brevity.

Claim audit done 2026-08-14 — six moderated, two numbers reconciled. Anchored to what the reviewers actually objected to (R2C4 and R2's closing paragraph on in-sample metrics and forward-vs-inverse; R3 on "outperforming all existing"; the AE endorsing both), not to general adjective-hunting. Method: render the text as a reader sees it — resolve \change{old}{new} to the new half, strip \note{} — then scan for evaluative language and check each hit against the evidence actually in hand.

Two were factual inconsistencies, now fixed:

The abstract still said R² = 0.80 (RMSE 0.052) for the multivariate fit while §7.2 and the Conclusions had moved to the production posterior. The temperature-only pair (0.75, 0.058) already matched; only the multivariate one did not. Now 0.81 (0.050).
The inverse figures (R² = 0.82, RMSE 4.4 °C) were unlabelled as in-sample, which R2C4 and R3 both asked for explicitly. Now labeled, with the spatially blocked counterpart given.

Six were overstatements against the paper's own evidence or its own hedged wording elsewhere. All use \change{} except the heading:

where	was	now
Conclusions (Tasman)	"removes the persistent warm bias"	"reduces", plus an explicit statement that these are sensitivity tests conditional on assumed nutrient depletion
§7.2 heading	"resolves first-order spatial biases"	"reduces" — §8.1's heading had already been softened the same way
§7.2	"captures the universal, nonlinear physiological response"	"captures the nonlinear response"
§8.3 PETM	nonthermal controls are "essential for reliable reconstruction"	"illustrates how much reconstructed temperature can depend on the assumed nonthermal state"
Plain Language Summary	"more geologically reasonable … more reliable window"	"agree more closely with independent proxy and model evidence", uncertainty reported
§6.1 / §7.2	"superior ability"; maps "unambiguously"	"better represents"; maps onto

The Tasman one was the important one: §8.1 says TEXAS "can recover realistic SST amplitudes once the extent of past nutrient depletion is constrained" and labels those runs sensitivity reconstructions, so the Conclusions was arguing against the body. It was also the sentence most exposed to R2C2.

The R2R's AE response now enumerates all six rather than asserting that claims "have been moderated", so an editor can spot-check each against the revised text.

Two gotchas worth remembering. A section heading cannot take \change{} — soul markup in a sectioning command propagates into the ToC and the .aux; edit the heading directly and record it in a \note{}. And a naive \change regex fails on nested braces like [NO$_{\text{3}}^-$], so a verification script needs a real brace matcher or it will report the struck-through half as if it were live text.

Scope limit, so nobody assumes more than was done: the scan used a fixed term list (outperform, superior, resolves, demonstrates, universal, essential, unambiguous, guarantee, and similar). It is not a full read-through, and it did not audit numerical claims other than the two above.

Full read-through audit done 2026-08-14 (the earlier one was a term-list scan; this one read the whole manuscript as a reader sees it). Twenty findings, all fixed. Three were serious:

Eleven figures were still the additive fit — fig7–fig14 and all three appendix prior plots pointed at the superseded versions. Copied from the TEXAS repo and repointed. fig6 deliberately stays the parent: its _boundedT counterpart contains the additive-vs-bounded comparison, which must never appear in the paper.
§7.2's curve parameters were the additive fit's. T₀ = 36.2±1.6, b = 0.437, k = 0.20, ν = 2.8 matches tx.GHEA (additive) to three decimals; the production tx.GHEB gives 34.8±0.7, 0.412, 0.28, 4.0. Replaced for SST and Thermo-T, and the interpretation with them — T₀ is no longer "virtually unchanged" but 1.7/3.5 °C cooler, and b no longer shifts "marginally higher". Fig. 8's caption had the same stale R²/RMSE.
The manuscript addressed a reviewer by name in §7 ("which Reviewer 3 rightly noted") — my wording, the same fault already fixed in Appendix C.

Plus: §7.1.1 and §7.2 both said 1000 warm-up against §7's 400 (budget now stated once); three places called T₀ the inflection point when §6.1 says it is not; two cross-references pointed at figure labels while typeset "Eq." and "Table"; the acronym read "indeX for"; a doubled period; a lowercase sentence start; four grammar slips. One regression of my own — the claim pass had dropped "the nonlinear temperature response" from §6.1 — restored.

CI was red and is now green. Nine ruff F401/F841 failures on this branch, each verified dead individually rather than bulk-autofixed. ruff check passes, 277 tests pass.

Still open after this session

Two figure choices need your eye. fig11 and fig12 previously pointed at hand-edited _revised.pdf files (they have paired .svg); they now point at the freshly generated _boundedT.pdf, which is the right data but carries none of the manual annotation. Also fig1_..._boundedT.pdf exists and is unused — the manuscript still shows fig1_..._revised.pdf.
One sentence of R3C3 is for the authors — whether the screening criteria and the index choice were fixed independently of the final calibration results. It is a red box in the R2R titled [TO BE CONFIRMED BY THE AUTHORS]. Everything else in R3C3 is drafted.
Two manuscript edits the R2R promises but that are NOT made — flagged in a second red box in the R2R. (1) Move §6.3's per-site uncertainty assignments, OLS/ODR implementation and delta-method propagation to the SI; that becomes Text S3, since Text S2 is now taken. (2) The Conclusions half of this is done.
ρ for NO₃ — RESOLVED 2026-08-14, was never wrong. The draft's −0.38 is quoted at threshold 1.8 and reproduces exactly (−0.382, n=701). The −0.328 that looked like a discrepancy was computed at threshold 1.0, a number the manuscript never claims. G2/3 reproduces at −0.360 too. No manuscript change needed.
The Open Research Section cites Zenodo v0.2.1, but tags run to v0.2.5 and the installed package reports 0.2.6. Three releases stale.
The frequentist notebook is committed but not yet in the Zenodo record — it is on this branch only, and releases are tagged from main. Zenodo archives the whole repo tarball (no export-ignore), so it travels automatically once the branch merges and a tag is cut; no manual step. Its input CSV is tracked in LFS, so it is reproducible from the archive.

PLAN — resubmission, then the archive split, then v1.0.0 (written 2026-08-14)

Three phases, strictly ordered. Phase A blocks resubmission. Do not start Phase B until the revised manuscript is out — moving files while the SI notebooks are still being edited is how a figure quietly loses its source.

The line the split is drawn on

The initial submission (tag v0.1.10, 2026-04-24) fitted the additive-EIV model: the nonthermal predictors added an offset to the response, outside the logistic. Compset GHEA, Stan file ..._priorApprox_eiv.stan, coefficients beta_G23_crtp / beta_NO3_crtp.

The revision fits the bounded-T model: the predictors shift T₀, inside the logistic. Compset GHEB, Stan file ..._priorApprox_eiv_boundedT.stan, coefficients gamma_G23_crtp / gamma_NO3_crtp.

Everything that exists only to reproduce the GHEA arm is archive. Everything the bounded-T arm runs on is the package. That is the whole rule.

Phase A — finish the revision (blocking)

15 of 31 R2R responses are still [RESPONSES] placeholders. This is the largest single piece of remaining work, and nothing else on this list is close. Unwritten: R1C2, R1C3, R1C4, R1C5, R1C6, R2C3, R2C5, R2C9, R2C11, R2C12, R2C13, R2C14, R2C17, R2C18, R2C19. Note that R1C1, R2C1, R2C2, R2C4, R3C1–R3C4 — the structural ones — are written.
The [TO BE CONFIRMED BY THE AUTHORS] red box in R3C3 (main.tex:640): one sentence on whether the screening criteria and the index choice were fixed independently of the final calibration results. Not draftable from the repository.
Two stale thermoT uncertainty maps — both were rendered before their inputs finished and silently show a subset of the 1513 sites. Regenerate figXX_..._thermoT_... from SI_code2 and figXX_..._thermoT_..._boundedT from SI_code02, then confirm the site count.
AppendixA_culmesoT_prior_distributions_boundedT.pdf is stale — predates the four single-predictor fits.
fig11 / fig12 lost their manual annotation. They now point at the freshly generated _boundedT.pdf, which is the right data with none of the hand editing that _revised.pdf carried. Re-annotate, or accept the plain version.
Commit the four dirty SI notebooks. SI_code2 is the dangerous one — its run cells are uncommented and executing it overwrites the audited 400/1000 posteriors.
The Open Research Section cites Zenodo v0.2.1; tags run to v0.2.5 and the package reports 0.2.6. Cut a tag and update the citation as part of resubmission.

Done this session (all uncommitted):

The R2R now opens with a Summary of revisions table (ten rows, each keyed to the comments it answers). Builds under xelatex, 28 pages.
Every Stan header rewritten to read standalone. No file describes itself as a diff against a superseded design any more — the two bounded-T headers were the worst offenders (the forward one opened with "THE PROBLEM THIS ADDRESSES (reviewer comment)" and a measured indictment of the additive model). Six files that had no header, or a header naming the wrong file, now have one. All 17 parse under stanc 2.36.0; 279 tests pass.
Tutorial Module 3 rebuilt on the bounded-T design: slider defaults are the published posterior medians rather than illustrative values (t₀ 34.8, k 0.275, b 0.412, ν 4.0), and there are new G₂/₃ and NO₃ sliders that slide the curve sideways while the thermal-only curve stays put, so the boundedness is visible rather than asserted.
README and docs/index.md describe the γ-on-T₀ formulation, and the forward-calibration example now names ..._eiv_boundedT.

⚠️ docs/index.md carries a new warning that the two multivariate posteriors on Zenodo are the additive-EIV formulation. That warning is a stopgap for the download.py problem in Phase C, not a fix — remove it when the bounded-T posteriors are published.

Two defects found in passing

The inflection_point generated quantity had the wrong sign in gen_logi_fixed_hier_crtp_univ_priorApprox.stan and gen_logi_fixed_culmesocore.stan: both computed t0 + ln(v)/k. Setting d²f/dT² = 0 gives exp(-k(T-T0)) = v, hence t0 − ln(v)/k. Verified numerically against the production posterior — the true max-slope temperature is 29.85 °C against t₀ = 34.8 °C, i.e. 4.95 °C below t₀ and matching the 4.2–5.2 °C range CLAUDE.md already records. The old formula returned 39.75 °C, wrong by 9.9 °C and on the wrong side. Both are fixed and the variable renamed to max_slope_temp, which is what it actually is. Nothing in Python or the notebooks read the old variable, so this renames cleanly; cached .nc files carry the old name and value, so regenerate before quoting it.
SI_code1 repeats the same sign error in Python — generalized_logistic_inflection_point() returns x0 + np.log(v)/k, and it is plotted as a marker on the parameter-sweep panels. The marker therefore sits on the wrong side of x₀. Not fixed here: that notebook is deliberately uncommitted, and correcting it moves a published figure. Decide before resubmission.

The archive already half-exists inside the package

src/TEXAS/stan_models/ contains archive/ (16 tracked files) and archive_pre_annotated/ (4) — the earlier gen_logi_free_*, logistic_* and ensemble invT models, with their own README. The non-recursive ls *.stan used when Phase B was drafted missed them. They are already excluded from the wheel (package-data is stan_models/*.stan, not **), so no packaging change is needed — but they should fold into archive/submission-2026-04/ in Phase B rather than leaving three archives in two places.

Phase B — the archive split

Decided 2026-08-14: repo-root archive/, not installed. The final package ships production models only; the archive is a readable folder rather than a git tag, and reproduction goes through an explicit model_dir=.

archive/
  submission-2026-04/
    README.md              <- what this reproduces (v0.1.10), and the exact commands
    stan_models/           <- 9 files
    notebooks/             <- SI_code2, SI_code3
    figures/               <- the additive-fit parents

B1 — Stan models: 17 → 9 shipped, 8 archived

DONE 2026-09-03 (a2153c0 on main). The counts in the original heading were transposed — the "stays" table lists 9 files (one row covers two) and the "moves" table 8. Corrected above. Three notes for whoever reads this next:

The "verified, not assumed" claim below is wrong for SI_code02a. Its cell 30 has a live get_posterior(..., stan_file=...) on the additive model, as do scripts/run_param_sensitivity.py and run_manuscript_refits.py. Handled by a fallback in StanCompiler.resolve_stan_path(): an archived model still resolves by plain stem, and announces that it is superseded. A path substitution would not have worked — stan_model_name is written verbatim from that argument and must stay a bare name for case-id resolution.

The model_dir= reproduction route does not exist through get_posterior(), which builds StanCompiler() with no arguments. The archive README documents the absolute-path route instead, which is verified.

The Streamlit blocker was not one. culmesocore appears in streamlit_app/config.py:45 as a posterior variable name (t0_culmesocore), not a model load. pages/{prediction,computation}.py never referenced it.

Also done in the same pass, beyond the plan: _select_invT_stan_file() narrowed to the two constraints that ship (hard_constraint, reparameterized, soft, and model_type="ensemble" now raise ValueError naming the archive), utils/naming.py::CONSTRAINT_CODES left intact with a comment saying why, docs updated, and tests/test_stan_model_archive.py added. tests/test_streamlit_params.py scans the new archive so the additive beta_* labels stay valid. 351 passed, 11 skipped.

Done 2026-09-07 (e0c729b5): B1's last bullet — folding src/TEXAS/stan_models/archive/ (15 .stan + README) and archive_pre_annotated/ (4) — moved to archive/pre-submission/stan_models/ (with the latter as its pre_annotated/ subdirectory), not into archive/submission-2026-04/. Four archive locations now exist under archive/, indexed by archive/README.md.

Stays in src/TEXAS/stan_models/ (8):

file	why it stays
`gen_logi_fixed_culmeso.stan`	stage-1 culture+mesocosm hyperpriors
`gen_logi_fixed_hier_crtp_univ_priorApprox.stan`	thermal-only fit; also produces `R2_thermal`
`gen_logi_fixed_hier_crtp_multiv_priorApprox_eiv_boundedT.stan`	the production calibration
`invT_gen_logi_fixed_multiv_marginal_unconstrained_boundedT.stan`	production inverse
`invT_gen_logi_fixed_univ_marginal_unconstrained.stan`	univariate inverse (quickstart, SI03)
`invT_gen_logi_fixed_multiv_marginal_unconstrained.stan`	multivariate inverse comparator in SI03
`invT_gen_logi_fixed_*_marginal_truncated_prior.stan` (2)	live docs — `docs/why_plugin_p50_differs.md` is in `_toc.yml` and explains the plug-in/marginal gap through them
`linear_model.stan`	used by `SI_code1` and every SI notebook

Moves to archive/submission-2026-04/stan_models/ (9):

file	why it moves
`gen_logi_fixed_hier_crtp_multiv_priorApprox_eiv.stan`	the initial submission's production model (`GHEA`)
`gen_logi_fixed_hier_crtp_multiv_priorApprox.stan`	superseded non-EIV intermediate
`gen_logi_fixed_hier_crtp_multiv.stan`	superseded full-hierarchical intermediate
`gen_logi_fixed_culmesocore.stan`	not used by any manuscript notebook (Streamlit only)
`invT_gen_logi_fixed_univ_unconstrained.stan`	non-marginal inverse; `SI_code3` only
`invT_gen_logi_fixed_multiv_unconstrained.stan`	non-marginal inverse; `SI_code3` only
`invT_gen_logi_fixed_*_marginal_hard_constraint.stan` (2)	zero references anywhere

Verified, not assumed: SI_code02's only mention of the additive _eiv model is in its header markdown table (documenting the difference) and in commented-out code. The bounded-T arm never compiles it.

Before moving, these break and must be handled:

src/TEXAS/utils/naming.py — the compset decoder maps A (additive) and the culmesocore training set. Keep the decoder intact; it must still parse tx.GHEA.* case ids or every archived posterior becomes unreadable.
streamlit_app/pages/{prediction,computation}.py reference culmesocore.
tests/test_naming.py, tests/test_run_tokens.py assert on archived names — they are testing the name grammar, not the models, so they stay.
scripts/prepare_review_archive.sh is written entirely against the legacy flat names of the initial submission. It belongs in archive/, not scripts/.
src/TEXAS/stan_models/archive/ (15 .stan + README) and archive_pre_annotated/ (4) — done 2026-09-07 (e0c729b5): folded into archive/pre-submission/stan_models/ (the latter as its pre_annotated/ subdirectory), not archive/submission-2026-04/. Neither ships in the wheel.

B2 — Notebooks

Move SI_code2_TEXAS_analysis.ipynb and SI_code3_paleo_showcases.ipynb to archive/submission-2026-04/notebooks/. Superseded by SI_code02_*_boundedT and SI03_*_modelswitch respectively.

SI_code3 is the one with a trap. It requests date-stamped legacy posterior names (..._scaledRI_cren3_050126_eiv). Those resolve only because the files on disk still carry the old name in their filename attr — a re-run would break them. Archive the notebook and the posteriors it needs, together, or it stops being reproducible.

B3 — Figures and posteriors

Additive-fit parents (fig7–fig14, the three Appendix prior plots) → archive. fig6's parent stays where it is — its _boundedT counterpart contains the additive-vs-bounded comparison, which must never appear in the paper.
data/cache/.../superseded/ and the flat legacy .nc names → the Zenodo archive record, not the repo (the cache is gitignored either way).
review_archive_v0.1.8/ at the repo root is untracked staging from April. Delete it, or fold it into archive/.

B4 — Root cleanup

site/, dist/, html_figures/, outputs/, logs/ are all untracked and gitignored, but Zenodo archives the working tree as a tarball. Clear them before tagging.

Phase C — v1.0.0

download.py still ships the additive posteriors as the public default. _ZENODO_FILES has five entries — two of them the additive multivariate (GHEA), the rest univariate/culmeso — and no bounded-T entry at all. A user who follows the README today downloads the superseded calibration. This is the single most important v1.0.0 fix — it is a correctness problem, not tidiness.
Same repointing in README.md:153, docs/index.md:119,136,137,186,203,265.
Upload the bounded-T posteriors to Zenodo; keep the GHEA ones in the record as the submission archive, clearly labeled.
pyproject.toml package-data stays stan_models/*.stan — correct, since the archive lives outside src/. Confirm the wheel drops from 17 to 8 models.
CITATION.cff / .zenodo.json → 1.0.0; drop the "in prep" / "prepared to submit" language once accepted.
Upload the 5 TEXAS-derived NetCDFs (16 MB) to the Zenodo data record 10.5281/zenodo.22131367 and add them to TRAINING_DATA_REGISTRY: Tierney22_PNAS_PETMDA/PETMDA_OCN_annual_regridded.nc and the four Zhu19_Science_EoceneSim/fullDepth_regridded_iPETM/*.nc. They were untracked on 2026-09-07 (repo-hygiene plan, Task 7) and SI_code00 and SI_code03 read them, so until this lands a fresh clone cannot run those two notebooks end to end.

Bootstrap on a different machine (Linux or Windows)

This file is tracked, so it arrives with the clone. Everything below assumes you have just sat down at a machine that has never seen this work.

git clone https://github.com/PaleoLipidRR/TEXAS.git && cd TEXAS
git checkout feat/revision1-validation-groupA

# 1. Hydrate Git LFS. THIS IS NOT OPTIONAL — without it, ~97 data files are
#    133-byte pointer stubs and every dataframe silently comes back empty.
git lfs install
git lfs pull

# 2. Verify hydration BEFORE running anything. Must print 0.
git lfs ls-files -n | while IFS= read -r f; do \
  [ -f "$f" ] && [ "$(stat -c%s "$f")" -lt 300 ] && \
  head -c 40 "$f" | grep -q 'git-lfs.github.com/spec' && echo "$f"; done | wc -l

# 3. Environment (conda is canonical; uv/venv also works)
conda env create -f environment.yml && conda activate texas-env
pip install -e .            # editable, or Stan recompiles from scratch every time

# 4. CmdStan must be discoverable, then confirm the whole toolchain
texas-doctor

# 5. Green baseline
pytest -q                   # expect 175 passed

What does NOT travel between machines, and what to do about it:

Local-only	Consequence	Fix
`stash@{0}` (Phase 0 snapshot)	No safety stash elsewhere	Irrelevant once the branch is pushed — the commits are the backup
`backup/pre-merge-20260809`	No rollback point elsewhere	Pushed to origin; `git fetch` brings it down
Compiled Stan binaries	First sample recompiles (slow, once)	Nothing — expected, and platform-specific anyway
`data/cache/**` posteriors	Reconstructions cannot be loaded	`git lfs pull`, `TEXAS.download_posteriors()`, or re-run
Kriged grids loose in the `data/cache/` root (pre-2026-09-07 layout)	Read with a printed fallback note; ~189 MB of superseded `1.0deg`/`2.5deg`/`kriged_halo_*` files never get cleaned up	`python scripts/migrate_kriged_cache.py` (dry run), then `--apply --delete-superseded`. Repeat on the Windows box — `data/cache/**` is gitignored
CmdStan install	Nothing samples	`texas-install-cmdstan`

CHECK THIS FIRST ON EVERY MACHINE: the LFS smudge filter

On 2026-08-10 this repo's .git/config was found overriding the global LFS filter with --skip:
filter.lfs.smudge  = git-lfs smudge --skip -- %f
filter.lfs.process = git-lfs filter-process --skip
--skip means LFS content is never materialized on checkout, so every git checkout, branch switch, stash apply, and merge silently writes 133-byte pointer stubs instead of data. It is almost certainly a leftover from the July LFS-over-budget period. This was the root cause of every LFS symptom in that session — the stash de-hydration, files turning back into stubs after a git checkout --, and the historical 88-of-97.

Diagnose and fix (repo-local, so it must be done on each clone):
git config --show-origin --get filter.lfs.smudge   # want the GLOBAL one, no --skip
git config --unset filter.lfs.smudge               # only if it shows --skip
git config --unset filter.lfs.process
git lfs pull
Verify with the stub count above (want 0), then confirm a checkout no longer stubs: delete an LFS file, git checkout -- it, and check its size.

Where am I?

Run this first, always. The working tree shrinks as you commit, so git status alone tells you which step you are on.

cd /c/Users/ratta/Documents/GitHub/TEXAS
git status -sb
git log --oneline -8
git stash list          # should be empty; if not, see "Recovery" at the bottom

Map the output to a phase:

`git status` shows	You are at
26 status lines, branch `feat/revision1-validation-groupA`	Phase 1, nothing committed
fewer status lines, new commits on the branch	Phase 1, mid-way — find the first unticked step below
only the two 1.8 leftovers, branch unchanged	Phase 1 done → go to Phase 2
clean tree, on `main`, `TEXAS-revision/` present	Phase 2 done → go to Phase 3

Notes on that count:

docs/_scripts/ is one status line but three files (28 files total; __pycache__ is gitignored).
RESUME.md never appears — .gitignore:108 already lists it. It is a local working file; delete it when done.
The two PhanTEX_*.csv files dropped out of the count during Phase 0: they were stat-dirty only (identical LFS oids c1658da/8fcee58 on both sides) and the stash cycle cleaned them. No content was lost.

Phase 1's file lists were checked against the working tree: every dirty file is assigned to exactly one step, none double-assigned, none missed.

HANDOFF — 2026-08-12, refit DONE and READY (read this first)

The refit finished at 16:20 and the audit passed. The main-text figures were re-run against it and are committed. Everything below this section predates the refit; this section is the current state.

python scripts/run_manuscript_refits.py audit        # re-run the audit any time
cat data/revision1/groupA/manuscript_refit/comparability_audit.json
cat data/revision1/groupA/manuscript_refit/case_ids.json

What ran, and what it produced

scripts/run_manuscript_refits.py all — started 14:07, done in 1:54:52, 71 runs = 7 forward + 64 reconstructions. It refit every manuscript case at one budget so the parent additive-EIV and bounded-T arms differ in the model and nothing else.

Forward 400/1000: not the cheapest cell for any single model, but the cheapest clearing all four gates for all three. A per-model budget would sample the two arms differently, which is a confound in exactly the comparison being made.
Inverse 500/1000, M=300.
Seed 42, 4 chains, proxy scaledRI_cren3, NO3 cutoff 1.0, both SST and thermoT.

The comparability audit reports READY — all 15 checks ok. The ones that carry the argument:

check	result
one forward budget / one inverse budget	400/1000 · 500/1000, M=300
both arms fitted, SST and thermoT	`['bnd', 'eiv']` for each
identical training rows across arms	n_obs = 1513, both targets
identical `R2_thermal` across arms	0.74558 (SST), 0.75711 (thermoT)
every reconstruction paired across arms	0 unpaired
reconstructions used this run's calibrations	0 used a legacy name
no date stamps in filenames	0 stamped
strict R-hat gate on forward posteriors	no failures

The seven case ids it wrote (case_ids.json) are already the short form — the version token is gone, so that decision has landed in what is on disk:

culmeso|cultureT  tx.GCDU.cul.sri03.p0
univ|SST          tx.GHPU.sst.sri03.p0        univ|thermoT  tx.GHPU.thm.sri03.p0
eiv|SST           tx.GHEA.sst.sri03.G23-N10   eiv|thermoT   tx.GHEA.thm.sri03.G23-N10
bnd|SST           tx.GHEB.sst.sri03.G23-N10   bnd|thermoT   tx.GHEB.thm.sri03.G23-N10

Rerunning is resumable and safe to interrupt: every completed run is already in manifest.csv and is skipped. kill -TERM <pid> finishes the run in flight, writes it, and exits; a second signal aborts. The lockfile is data/revision1/groupA/manuscript_refit/.run.lock, and the script refuses to start while the sensitivity sweep holds its own lock — two Stan jobs on this box share one binary cache and one set of cores.

Figures re-run against the refit — committed 2026-08-12

SI03 now runs both temperature targets, not SST alone: the GIG run plan goes 28 → 56 Stan runs and all 56 invT posteriors load. Four bounded-T panels were written from it and committed beside the additive-EIV originals rather than over them (figures(boundedT): main-text panels from the post-refit posteriors):

fig7 calibration curves · fig11 Tasman Sea · fig12 GIG · fig13 PETM

The SI sweep panels under figures/manuscript/revision1/ are committed too.

SI03 stops at the extreme-RI load cell (cell 78, In[63]): data_list_extreme_example.pkl is not in this machine's posterior cache, so 13 cells below it never ran. It is a missing input, not a broken cell — fig14 dates from 07-06 and was made elsewhere.

What this does NOT disturb

SI_code3_paleo_showcases.ipynb is the original submission's paleo analysis. Nothing it reads is overwritten: forward refits take the next free member (.002 beside .001) and the reconstructions carry new scenario tags, so the date-stamped files it loads stay exactly where they are.

SI_code03_paleo_showcases.ipynb is the revision notebook and is what the refit feeds. It has been switched to the case ids and re-run; its load cells now find all 56 reconstructions.

Naming changed: no more dates in filenames

The run date is now recorded in the run_timestamp attr instead. It was not recorded anywhere before — the date lived only in the filename, so this had to be added first or the date would have been lost outright.

The date was also doing collision-avoidance, which the run/member token now does properly: save_posterior(run="auto") takes the next free member, so a refit lands beside the run it repeats rather than on top of it. Because a legacy name pins no member, resolve_posterior_path now returns the newest member — an ascending scan would have served the first fit of a configuration forever, silently, since nothing downstream reports which member it loaded.

SI03's NO3 scenarios are named rather than dated (no3_modern, no3_01, no3_001, no3_10). Stripping the date exposed that it was load-bearing: the modern-NO3 scenario's tag was the bare date.

Post-refit cleanup — FLATTEN the cache, then dedupe (decided 2026-08-12)

Two decisions, both to run once the refit lands. Neither is safe mid-run: the job writes into this cache, and a restart under changed layout code would give one refit two layouts.

0. Drop the version token, in the same change. Decided 2026-08-12. v026 is TEXAS.__version__ with separators stripped, and the pip version is the wrong signal for what that position does — wrong in both directions. A docs-only release bumps it and orphans every existing case directory, because resolution matches the token as an exact string (verified, not hypothetical). A .stan prior change without a release does not bump it, so two genuinely incompatible posteriors share one identity — and CLAUDE.md logs several such changes (Q removal 2026-03-24, the sigma_proxyObs_crtp prior 2026-04-08).

The position existed for collision avoidance. The run/member token now does that job properly: two fits of one configuration get .001 and .002 whatever the reason they differ. So the version is removed from the name and recorded as a texas_version attr beside run_timestamp.

tx.GHEA.sst.sri03.G23-N10.001.fwd.nc          <- 29 chars, from 41

Accepted tradeoff: two Zenodo deposits from different paper versions could each carry ...001.fwd.nc for one configuration from different model code. Zenodo's DOI versioning covers it, since a reader downloads one deposit.

CaseName.version removed from the dotted form; default_version() retired or kept only to populate the attr
parse_case() accepts BOTH forms — every case id on disk and in the notebooks has v026 in it, so parsing must stay backward compatible
texas_version written in extract_and_update_metadata alongside run_timestamp
the case_id attr and case_ids.json regenerate to the short form

1. Flatten both caches. Today they hold two layouts at once — forward 17 flat + 18 in case directories, invT 35 flat + 46 in 6 directories — which is the confusion this fixes. The case directory earns nothing: the leaf already carries the whole case id, so it is self-identifying either way, and the directory only repeats it. Flat also matches Zenodo, halves the path (72 -> 41 chars), and still groups a calibration with its reconstructions, because ...001.fwd.nc and ...001.inv.<site>.nc sort adjacent.

Target:

data/cache/TEXAS_posterior_cache/tx.v026.GHEA.sst.sri03.G23-N10.001.fwd.nc
data/cache/TEXAS_invT_posterior_cache/tx.v026.GHEA.sst.sri03.G23-N10.001.inv.U1482.ud.nc

Order of work:

naming.fwd_relpath() / inv_relpath() return a bare leaf, not <case>/<leaf>
resolve_posterior_path() gains the flat-leaf candidate and KEEPS the two directory forms, so nothing on disk has to move for reads to work
next_free_run() scans flat files as well as directories — it currently only looks at directories, so after flattening it would restart at .001 and collide
io._generate_filename_base() drops its <case>/ prefix
download.py::_local_dest() becomes the identity function
migrate_cache_layout.py --flatten to move what exists, dry-run first
tests: a flat leaf resolves, a directory leaf still resolves, members still increment

2. Dedupe. At least three known duplicate pairs, all from the old date-stamping:

..._eiv_SST_gdgt23ratio_no3_1.0_scaledRI_cren3.nc and its _041626_eiv twin — identical statistics, so almost certainly identical draws
tx.v026.GCDU.cul.ri3.none.001/fwd.nc vs ...cul.sri03.p0.001/ — verified byte-identical draws
tx.v026.GHPU.sst.ri3.none.001/fwd.nc vs ...sst.sri03.p0.001/ — draws DIFFER; the new one is today's 400/1000 refit, so keep that and retire the old, but decide deliberately rather than by script

Compare draws, not file size, before deleting anything: two of the three pairs differ in bytes while agreeing in content.

Done since the refit landed (2026-08-12)

Ran audit — READY, 15/15. case_ids.json written.
Pointed SI03 at the case ids (decided 2026-08-12: case ids are the canonical identity from the resubmission on). Its fwd_name() built legacy names, and a legacy name cannot reach the refit posteriors at all: the cache holds 17 flat files with exactly those names, and an exact flat hit is the first thing resolve_posterior_path tries — so SI03 was silently loading the pre-refit fits. Do not delete the flat files — SI_code3_paleo_showcases.ipynb, the original submission, reads them. They are the compatibility layer; new work names case ids explicitly.
Re-ran SI03's figure cells against the new posteriors; four bounded-T panels committed.
Filled the bounded-T grid. run_manuscript_refits.py fits each arm with the full predictor set only, so bounded-T had 2 of the 6 cells the additive arm has, and SI_code02's five-layer prior figures were two layers short. scripts/fit_boundedT_single_predictors.py fitted tx.GHEB.{sst,thm}.sri03.{G23,N10} at the refit's budget, loading culmeso and the univariate baseline from the refit manifest rather than resampling. All 0 divergences, max R-hat 1.0096. It writes its own manifest, so audit still reports READY 15/15.
Put SI_code02 on case ids. 21 name sites across 10 cells. This was not tidiness: legacy names hit the surviving flat files first, so those figures were mixing a pre-refit culmeso and univariate baseline with a post-refit bounded-T layer, silently (culmeso t0 35.80 vs 35.64).
SI03 runs clean end to end (cells 0-88, 0 failures). fig11-14 all regenerated in one pass from the audited cache. The fig14 cell held two more dead names: one raised, and one -- the draws_tag -- failed SILENTLY into a Gaussian fallback, so the KDE branch had been dead. Both sites now plot real posterior KDEs (4000 draws).
Rebuilt the extreme-RI section of SI03. Its generator was pinned to a date-stamped posterior that no longer exists. Now case ids, both arms, per-variant pickles that do not collide with SI_code3's. bounded-T moves ODP1259 by 3.33 degC and Co1010 by 0.38 degC.

Decided, do not relitigate

tx.GHPU.sst.sri05.p0 fails the strict R-hat gate (1.01407) and that is accepted. Checked 2026-08-13. The failing fit is the SRI05 univariate stage, which enters the analysis only through R2_thermal, i.e. as a prior scale. The fits that actually produce the coefficients both pass — SRI05 eiv 1.00887, SRI05 bnd 1.00827 — and MCSE on gamma_G23 is 0.00123 on a median of 0.7641 (0.16%), so the 20.5% G23 spread statistic is precise to a fraction of a point. SRI05 is a comparator, not a production calibration; no refit. Note it does set the upper end of that spread, so "unused" is the wrong reason to dismiss it — "its coefficient fit converged" is the right one.
The univariate model is the SLOWEST of the three to converge, not the fastest, despite having no predictors: only 9 of 27 budget cells clear R-hat <= 1.01, and the worst parameter rotates among k/v/t0 trading off against each other. The production 400/1000 budget clears it comfortably (1.00465, ESS 760); the cheapest clearing cell is 400/600.

Both thermoT uncertainty maps are STALE — regenerate first (2026-08-13)

Each was written before the reconstructions it draws finished. The maps glob the global_coretop_* invT posteriors, so an early run silently renders a subset of the 1513 sites rather than failing.

figure	inputs written	figure written	verdict
`figXX_..._thermoT_..._boundedT`	13:53 → 16:12	13:23	built before any input existed
`figXX_..._thermoT_...` (additive-EIV)	12:34 → 13:47	13:13	missing the last batches
`fig10_..._SST_..._boundedT`	09:53 → 11:02	13:23	OK
`fig9`, `figS13`	(SST)	—	OK, SST only

Both are committed — the bounded-T one in 16bfb3e, the additive-EIV one in 04a5b24 (committed 2026-08-13 believing it was a clean recache; it was not). SST maps are fine; only thermoT is affected, in both arms.

Now safe to regenerate: the bounded-T arm reached 28/28 at 16:12 and the additive-EIV arm is 28/28, so both sets are complete for the first time. Re-run the map cells in SI_code2 (untagged) and SI_code02 (_boundedT).

regenerate figXX_..._thermoT_... from SI_code2
regenerate figXX_..._thermoT_..._boundedT from SI_code02
confirm each covers all 1513 sites, not a subset

Still open when you get back

AppendixA_culmesoT_prior_distributions_boundedT.pdf is stale — written 17:29 on 2026-08-12, before the four single-predictor fits landed and from the pre-rename cell sources. Regenerate from SI_code02.

Under-coverage explained 2026-08-12 — it is interval WIDTH, not bias. One number accounts for both figures: the mean 68% half-width is 0.863-0.867x the residual SD (3.83 degC against 4.43 degC). Feed that ratio through a Gaussian and it predicts cov68 = 0.612 against 0.60 observed, and cov90 = 0.846 against 0.84 observed. Both, to within a couple of points, from one quantity.

The +0.93 degC bias is **not** the cause: removing it changes coverage by
-0.005, i.e. it slightly *lowers* it. And the ratio is stable to 0.004
across every budget cell, which is why this was never sampler noise --
more draws cannot widen an interval the model does not think is wide.

So the honest statement for the SI is that the predictive intervals
understate the true error by about 14%: the residual spread contains
site-level variability (oceanographic, depth-habitat, bioturbation) that
the calibration's noise model does not carry. Reproduce with
`data/revision1/groupA/param_sensitivity/invt_budget_sites.csv`.

> ### CORRECTED 2026-08-14 — this is a STRESS-SET result, not a population one
>
> **Do not quote the 14% figure as the model's interval calibration.**
> `invt_budget_sites.csv` is the 200-site sampler-tuning subset built by
> `invt_subset()` in `scripts/run_param_sensitivity.py`, which fills
> equal-width bins over the proxy *range* with an equal quota each. Its own
> docstring is explicit: *"a stress set, not a validation set: error
> statistics over it are not population statistics for the compilation."*
> It deliberately over-weights the asymptote regions, where intervals are
> worst. The 200 sites are also **not held out** — they are drawn from the
> same coretops the calibration was fitted on.
>
> Recomputed over **all 1513 sites** from the `global_coretop_b01..b07`
> invT posteriors (`tx.GHEB.sst.sri03.G23-N10`, alignment checked per site
> against `scaledRI_cren3`), the production calibration is close to
> calibrated, and the bias even **flips sign**:
>
> | | stress set (n=200) | population (n=1513) |
> |---|---|---|
> | bias | **+0.93** degC | **-0.99** degC |
> | RMSE | 4.52 | 4.35 |
> | R2 | — | 0.824 |
> | half-width / residual SD | 0.865 | **1.05** |
> | cov68 (nominal 0.68) | 0.60 | **0.664** |
> | cov90 (nominal 0.90) | 0.84 | 0.853 |
>
> The population RMSE/R2 reproduce the manuscript's reported inverse
> in-sample figures (4.4 degC, 0.82), which is the check that the
> alignment is right. Univariate for comparison: RMSE 3.87, R2 0.860,
> cov68 0.770 — **inverse skill is better without the predictors**, which
> is the evidence behind R2C4 and R1C2.
>
> The stress-set numbers remain valid *as a statement about the asymptote
> regime* — coverage really does degrade there — and R1C2 in the R2R now
> says exactly that, sourced to the population figures.

The invT drift floor rests on one seed replicate (0.271 degC). Two or three more would make the "budget does not matter" claim rigorous.
Phase 5A done 2026-08-13 (58ddb86). _generate_filename_base now calls inv_relpath instead of respelling the leaf; scenario widened to take a sequence, which was the only real difference between them. Names unchanged byte for byte, checked over 216 combinations before committing. The drift this closes had already happened and was invisible: only the unreachable copy was wrong (it still documented <case>/<leaf> and a run number), so nothing broke and nothing could have caught it. The test now asserts the two builders agree rather than asserting a literal. The other two Phase 5A items were already fixed — save_invT_posterior no longer drops proxy_name (it delegates to the shared builder), and the cache flattening has landed (fwd_relpath/inv_relpath both return a bare leaf).
The branch is 0 behind main as of 2026-08-13 (110 ahead). The "6 behind" note was stale.
DOI reconciled 2026-08-12. data/README.md cited 19666745 while README.md, CITATION.cff and download.py used 20032542. Aligned on 20032542, which download.py documents as the currently published record and actually fetches from. Check this if 19666745 was the concept DOI (all-versions) rather than a superseded version DOI — in that case the right move is the opposite one, and citing the concept DOI is better practice. Could not verify from here without network access.
~~streamlit_app/pages/calibration_data.py reads post["Q_crtp"]~~ — already fixed (checked 2026-08-13). Q_crtp survives only in two comments explaining the repair; there is no live read of it anywhere in streamlit_app/ or src/. The page also lost its private copy of the curve in the same pass, which is what let it drift out of step.
fig6's four slopes RESOLVED 2026-08-14. All four reproduce, including the one that did not on 08-13. Thermal R² 0.747, G2/3 OLS −0.0058, G2/3 ODR −0.0059, NO₃ OLS −0.0296, and NO₃ ODR −0.072 — the draft value is correct. The 08-13 figure of −0.052 was the degenerate standalone re-fit, exactly as suspected; running the hierarchical culmeso→coretop path reproduces −0.072. §6.3's bracket claim stands on 1.7–4.3. Settled by notebooks/manuscripts/SI_code01_boundedT_variance_partitioning.ipynb (35a256d), which reimplements the two-step protocol with a switch for the parameterization. ρ for NO₃ is still unconfirmed (−0.328 computed against the draft's −0.38) — that is a Spearman on the plotted subset, not an output of the fitter, so it needs checking in SI_code1 cell 123 itself.
fig6_ODR_regression_dilution.pdf — superseded, not regenerated (2026-08-14). The manuscript moved to the bounded-T formulation, so the relevant figure is now fig6_ODR_regression_dilution_boundedT.pdf (35a256d), built from the current spreadsheets by SI_code01. The parent PDF stays as the record of the original submission and is deliberately not rebuilt. The prior-elicitation worry is closed by the slopes above: they did not move, so the β priors on disk were not elicited from superseded data.

Uncommitted, deliberately

Four SI notebooks only (2026-08-14): SI_code00_PreProcessing, SI_code2_TEXAS_analysis, SI_code02_boundedT_TEXAS_analysis, SI_code02a_model_param_sensitivity_test. SI_code2 is the dangerous one — its run cells are uncommented, so executing it overwrites the audited 400/1000 posteriors. Left alone; commit when you are happy with them.

The figures and the two LFS training spreadsheets that used to sit here are now committed (cca55ca), so a fresh clone gets the same inputs and figures rather than silently different ones.

STATUS: Phases 0, 1, and 2 are DONE — resume at Phase 3 (2026-08-10)

Handoff — Linux box → Windows, 2026-08-11

Written at the end of a Linux session so the Windows machine can pick up cold.

Do this first on Windows, before trusting any dataframe:
git checkout feat/revision1-validation-groupA && git pull
git config --show-origin --get filter.lfs.smudge   # must be the GLOBAL one, NO --skip
The Linux clone's .git/config had the --skip smudge override described below; it was unset there on 2026-08-11 and LFS is fully hydrated (0/99 stubs). That fix is per-clone and does not travel — assume Windows still has it until you have checked. The GitHub LFS budget is restored, so git lfs pull works again; the Zenodo fallback is no longer required.

What changed in this session — documentation only, no code touched:

Phase 5 was rescoped. It used to say "rename the .stan model files". That was a mistake: the target is the posterior .nc filenames (up to 118 chars). Phase 5 is rewritten around that, with four verified defects (5A–5D) found by auditing the code and the live cache rather than by reading docs. The .stan rename is out of scope.

A "fact" in the do-not-re-derive list was wrong and is struck through: the forward cache has 2 case-id collisions, not 0. Anything that migrates the cache must be blocked until Phase 5C lands.

The invT-cache "known gap" is machine-dependent and was stated as universal; both machines' contents are now recorded side by side.

Verified on Linux, 2026-08-11: texas-doctor → Stan sampling READY; pytest -q → 173 passed, 2 skipped (the "175" below counts the 2 Windows-only skips as passes — expect the same total, split differently, on Windows). Note the editable install's metadata had gone stale at 0.2.1, which is why texas-doctor was missing as a command; pip install -e . --no-deps fixed it. If a console script is missing on Windows, that is the first thing to try.

Stale local branches were deleted on the Linux box only (working-branch, archive/laptop-before-merge, restructure-repo, tutorial, local gh-pages) — all verified superseded. Windows may still list them.

Everything below is pushed to origin. Nothing important lives only on one machine any more; you can pick this up from any clone.

feat/revision1-validation-groupA @ 3e72cd8 — 9 commits, 175 tests passing
main @ d70405a — gridT merged (Phase 2), 85 tests passing (85 not 175 is correct: test_naming.py + test_stan_ascii.py are 90 tests that live on the feature branch, not yet on main)
backup/pre-merge-20260809 @ a0b3887 — rollback point, on origin
The Phase 0 stash was verified redundant against the pushed commits, then dropped

The feature branch is now 6 commits behind main. Merge or rebase before opening PRs in Phase 4.

Two commits beyond the original 7, both for cross-machine portability:

chore: normalize line endings deterministically across platforms — adds * text=auto so a Windows clone (autocrlf=true) and a Linux clone (autocrlf=false) cannot commit the same file with different endings. Third-party published_data/**/*.txt excluded to stay byte-identical. Verified zero churn via git add --renormalize ..
docs: track RESUME.md as the cross-machine handoff note — this file used to be gitignored as scratch, so the plan could not follow the work to the Linux boxes. Now tracked, with the bootstrap section above.

ab10ce0 notebook(SI03): model-switchable paleo showcases (additive EIV vs bounded-T)
a9a8d9f data(revision1): revised training spreadsheet, SI notebook updates, regenerated AppendixA
ac4ce3d docs: generated call map, regenerated on every deploy
fc68ae4 feat(predict): fwd_cache_dir to resolve posteriors outside the default cache
06c2e3a feat(naming): CESM-style case names for posteriors, dual-read
30ccdff fix(windows): compile from an ASCII-sanitized copy of the .stan source
c1a75bb feat(boundedT): bounded-T model support across the package

Resume at Phase 3. Working tree is clean. Phase 1.8's two leftovers are resolved: the bounded-T inverse .stan is now committed (the branch was not self-contained without it — a fresh clone could not run MODEL_VARIANT="boundedT" at all), and .claude/settings.json is gitignored as a machine-specific allowlist.

SI03 is ready to run

Verified by executing cells 0–31 headlessly on 2026-08-10:

LFS hydrated (0 stubs / 97), all deps present, texas-doctor → READY
Paths resolve on any machine (repo root by pyproject.toml; OneDrive by searching ~/OneDrive* for Postdoc/WOA23)
With TEMP_PARAMS = ["SST"]: the run plan reports 28 Stan runs and the load cell looks for exactly 28 files — run and load agree
With thermoT added: 44 runs, 12 skipped, reported up front, because the bounded-T thermoT (g23+no3) forward posterior does not exist

Nothing is cached for the paleo sites, so every column is NaN until you set RUN_INVT = True. data/cache/ is gitignored — those posteriors were never in git and cannot come from LFS or Zenodo. They must be generated, on whichever machine you choose.

Sample counts: MD98-2152 200, U1482 259, DSDP591 46, U1510 43, ODP959 371, South Dover Bridge 53.

Three things that happened, for the record

git stash de-hydrated an LFS file. The Phase 0 stash cycle turned PhanTEX_v001_modified_121025.csv (3.6 MB) into a 133-byte pointer stub, and my Phase 0 note wrongly called that "stat-dirtiness cleaned". 75 of 97 LFS files were stubs (most pre-existing). Fixed with git lfs pull — 296 MB, 0 stubs remain. Before stashing in this repo, count stubs; after restoring, count again.
data/spreadsheets/ is gitignored as a directory — its files are tracked individually, so step 1.7 needs git add -u <file>, not git add <dir>.
git commit --amend hit the wrong commit and merged SI03 into the data commit. Recovered with git reset --soft ac4ce3d and re-committing the two separately. Every commit message now matches its contents (verified).

Phase 0 — safety net (2 min) — DONE

Nothing here is committed yet, so this is the only irreversible state in the repo.

0.1 Make a backup branch pointing at the current commit

git branch backup/pre-merge-20260809
git branch --list 'backup/*'

0.2 Snapshot the uncommitted work as a stash that stays on the stack

git stash push --include-untracked -m "pre-merge snapshot 20260809"
git stash apply          # put it all back; the stash stays as a copy
git stash list           # must show: stash@{0}: On ...: pre-merge snapshot 20260809
git status -sb           # must show the same ~28 dirty paths as before

Why both: the branch protects committed history, the stash protects the uncommitted tree. Drop the stash (git stash drop) only after Phase 1 is fully committed and verified.

Phase 1 — commit the working tree in themed commits

All on the current branch. Do not switch branches until this phase is done — a dirty tree of this size will not survive a checkout cleanly.

Run the test suite once before starting, so you know the baseline is green:

.venv/Scripts/python.exe -m pytest -q      # expect: 175 passed

1.1 Bounded-T support across the package

Covers gamma-vs-beta detection, bounded-T invT model selection, gamma_* parsing in metadata, and gamma_* groups in the prior plots. Also carries the fwd_cache_dir and fwd_case changes — those are interleaved with the bounded-T hunks in the same files and cannot be split without git add -p, which this environment cannot run.

git add src/TEXAS/data/builder.py src/TEXAS/stan/invT.py \
        src/TEXAS/stan/metadata.py src/TEXAS/plotting/prior_plot.py
git commit -m "feat(boundedT): bounded-T model support across the package

Detect gamma_G23/gamma_NO3 in a forward posterior and route to the bounded-T
inverse model automatically, renaming the Stan data keys beta_* -> gamma_*.
Parse bounded parameter declarations in metadata; add gamma_* groups to the
prior plots. Also attaches fwd_case/fwd_posterior_name provenance to invT
posteriors so a reconstruction can be traced to the calibration it used."

Verify: git show --stat HEAD lists exactly 4 files.

1.2 ASCII-safe Stan build copies

git add src/TEXAS/stan/compiler.py tests/test_stan_ascii.py
git commit -m "fix(windows): compile from an ASCII-sanitized copy of the .stan source

cmdstanpy opens .stan with the platform locale codec (cp1252 on Windows), so
any non-ASCII byte in a comment raises UnicodeDecodeError at compile time.
Sanitize the disposable build copy only; model sources keep their Unicode.
Tests assert the invariant for every shipped model."

Verify: .venv/Scripts/python.exe -m pytest tests/test_stan_ascii.py -q

1.3 CESM-style case naming

git add src/TEXAS/utils/naming.py src/TEXAS/stan/io.py tests/test_naming.py CLAUDE.md
git commit -m "feat(naming): CESM-style case names for posteriors, dual-read

Replace concatenated-description filenames (95-122 chars, growing with every
new axis) with fixed dot-delimited positions: tx.v025.GHEB.sst.ri3.G23-N10.001
as a case directory holding fwd.nc and its inv.*.nc reconstructions.

Nothing on disk is renamed. load_posterior() accepts either a case id or a
legacy long name and finds the file under either layout, so existing caches,
Zenodo downloads, and old notebooks keep working.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>"

Verify: .venv/Scripts/python.exe -m pytest tests/test_naming.py -q → 53 passed.

1.4 Forward-cache directory plumbing

git add src/TEXAS/predict.py
git commit -m "feat(predict): fwd_cache_dir to resolve posteriors outside the default cache"

1.5 Docs call map — MUST be one commit

docs.yml invokes docs/_scripts/build_callmap.py. Committing the workflow without the script breaks the docs deploy on main. __pycache__ is already gitignored, so git add docs/_scripts/ is safe.

git add .github/workflows/docs.yml docs/README.md docs/_config.yml docs/_toc.yml \
        docs/_scripts/ docs/callmap.md docs/_static/callmap.html
git status --porcelain -- docs/ .github/     # must be empty
git commit -m "docs: generated call map, regenerated on every deploy

build_callmap.py runs before the book build so the call graph, the
reachability report, and the API cannot drift apart; it fails loudly if
callmap_content.py names a function that no longer exists."

Verify: git show --stat HEAD | grep -c _scripts → at least 3 (the three script files).

1.6 SI03 model-switch notebook

git add notebooks/manuscripts/SI_code03_paleo_showcases.ipynb
git commit -m "notebook(SI03): model-switchable paleo showcases (additive EIV vs bounded-T)

One MODEL_VARIANT flag drives the whole notebook. The active variant fills the
canonical column names so all four manuscript figure cells are unchanged; only
the saved figure filename gains a _boundedT tag. LOAD_BOTH loads the other
variant into *_alt columns to feed the reviewer-comparison section (per-site
delta-T table, 1:1 + residual figure, provenance CSV).

Run and load iterate one NO3_SCENARIOS registry, so a run's filename and a
load's filename cannot drift apart.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>"

1.7 Revised data and regenerated figures

The two PhanTEX_*.csv files are stat-dirty only — their LFS oids are unchanged (c1658da -> c1658da, 8fcee58 -> 8fcee58). git add cleans them without recording a content change. Only ds_gridded_...csv really changed (2 lines).

git add data/spreadsheets/ notebooks/manuscripts/SI_code00_PreProcessing.ipynb \
        notebooks/manuscripts/SI_code2_TEXAS_analysis.ipynb \
        notebooks/manuscripts/SI_code3_paleo_showcases.ipynb \
        figures/manuscript/finalized/main-text/AppendixA_culmesoT_prior_distributions.pdf
git commit -m "data(revision1): revised training spreadsheet, SI notebook updates, regenerated AppendixA"

1.8 Decide the two leftovers

git status --porcelain      # should show only these two

.claude/settings.json — Claude Code project settings, currently untracked and not gitignored. Either commit it (shares hooks/permissions with collaborators) or add .claude/ to .gitignore. Your call.
src/TEXAS/stan_models/invT_gen_logi_fixed_multiv_marginal_unconstrained_boundedT.stan — do not commit. It is already on revision/boundedT-si-figures, byte-identical apart from CRLF vs LF (177-line file, 177-byte size delta). Phase 3 brings it in from that branch. Leave it untracked for now.

1.9 Phase 1 gate

.venv/Scripts/python.exe -m pytest -q     # expect: 175 passed
git log --oneline -7                      # your 7 new commits
git status -sb                            # only the two leftovers from 1.8
git push origin feat/revision1-validation-groupA

Once this is green and pushed, drop the safety stash: git stash drop.

Phase 2 — land the free win: the gridT branch — DONE

origin/claude/gridt-inversion-characterization-15i183 is 5 commits, all new files under TEXAS-revision/, with zero overlap against any other branch or against your working tree. It is the newest work in the repo (2026-08-01), unreviewed, and easy to lose track of. Nothing else depends on it.

2.1 Merge it

git checkout main && git pull
git merge --no-ff origin/claude/gridt-inversion-characterization-15i183 \
    -m "Merge gridT inversion characterization + explainer"
ls TEXAS-revision/          # 9 files: 6 assets + 3 docs
git push origin main

Expected: clean merge, no conflicts. If git reports any conflict, stop — something changed since 2026-08-09 and this plan's assumptions need rechecking.

Phase 3 — reconcile `revision/boundedT-si-figures`

4 commits: the two bounded-T .stan files, ~15 regenerated figures, and SI_code2. After Phase 1 it collides on three files. Handle each deliberately — do not let git auto-resolve binaries or notebooks.

3.1 Start the merge and see the damage

git checkout main && git pull
git merge --no-ff revision/boundedT-si-figures
git status --short --diff-filter=U       # the conflicted set

3.2 Resolve, file by file

File	How to resolve
`src/TEXAS/stan_models/invT_..._boundedT.stan`	Now committed on the feature branch too, so this is a same-content conflict (CRLF only). Take either: `git checkout --theirs <path>`
`figures/.../AppendixA_culmesoT_prior_distributions.pdf`	Binary — you must decide. Your Phase-1 commit and the branch each regenerated it. Open both and pick the one from the newer `prior_plot.py`. Likely yours (`--ours`), since Phase 1 carries the rewritten plotting code.
`notebooks/manuscripts/SI_code2_TEXAS_analysis.ipynb`	Modified on both sides. Do not take either blindly — diff the cell sources first (see command below) and merge by hand.

# inspect the notebook conflict without drowning in output/base64
git show :2:notebooks/manuscripts/SI_code2_TEXAS_analysis.ipynb > /tmp/ours.ipynb
git show :3:notebooks/manuscripts/SI_code2_TEXAS_analysis.ipynb > /tmp/theirs.ipynb
.venv/Scripts/python.exe - <<'EOF'
import json
for tag, p in (("OURS", "/tmp/ours.ipynb"), ("THEIRS", "/tmp/theirs.ipynb")):
    nb = json.load(open(p, encoding="utf-8"))
    print(tag, len(nb["cells"]), "cells")
EOF

3.3 Finish

.venv/Scripts/python.exe -m pytest -q      # expect 175 passed
git add -A && git commit
git push origin main

pyproject.toml is not a real conflict: this branch and groupA both bump 0.2.5 → 0.2.6 identically.

Phase 4 — PR #15

4.1 Read the recommendation, then decide

PR #15 ("Revision-1 analysis workflow: Group A + hand-off plan") has been open since 2026-07-15. The branch has moved 6 commits past where it opened, and Phase 1 adds 7 more → 22 commits, against a 2026-09-08 deadline.

Recommendation: split. Themes 1.3 (naming), 1.5 (docs call map) and 1.6 (SI03 notebook) are independent of the Group-A validation work and touch disjoint files. Cherry-pick each onto its own branch off the updated main, open three small PRs, and leave #15 as what it claims to be.

# after Phases 2 and 3, main is current
git checkout main && git pull

git checkout -b feat/case-naming main
git cherry-pick <sha of 1.3>
git push -u origin feat/case-naming

git checkout -b docs/callmap main
git cherry-pick <sha of 1.5>
git push -u origin docs/callmap

git checkout -b notebook/si03-modelswitch main
git cherry-pick <sha of 1.6>
git push -u origin notebook/si03-modelswitch

Get the SHAs with git log --oneline -7 feat/revision1-validation-groupA. These three commits touch files nothing else touches, so the cherry-picks should be clean.

4.2 Rebase what's left of #15 onto the updated main and re-request review

Phase 5 — shorten and systematize the posterior `.nc` filenames

Scope correction (2026-08-11). An earlier draft of this phase described renaming the 17 .stan model sources. That was a mistake — the .stan files are not the problem and are not in scope. The artifacts that need shortening are the posterior .nc output files, whose names run to 118 characters and grow with every new axis. The .stan rename is a separate, optional idea; if it is ever revisited it must be its own phase, because it carries a full-recompile cost and six silent string-parsing hazards that the .nc work does not.

The problem, measured

Today's cache on this Linux box: 17 forward + 35 inverse .nc, and zero case directories. Worst offender at 118 characters:

MD98-2152_invT_gen_logi_fixed_multiv_unconstrained_thermoT_gdgt23ratio_no3_1.0_scaledRI_cren3_050126_no3_001_direct.nc

The CESM-style scheme already exists in src/TEXAS/utils/naming.py (605 lines) and compresses hard — forward names measured across all 17 files:

	legacy	case id
shortest	49	27
longest	104	32

tx.v026.GHEA.sst.ri3.G23-N10.001/       <- the case = one calibration identity
    fwd.nc                              <- the forward posterior
    inv.U1482.ud-050126.nc              <- a reconstruction derived from it

So the scheme is not what is missing. The wiring is. The forward half works; the inverse half is written but not connected, and nothing on disk has moved.

Decided and implemented 2026-08-11 (commit follows this file)

Four naming decisions, all landed with tests, none of them touching the published Zenodo record:

axis	was	now	why
leaf name	`fwd.nc`	`<case>.fwd.nc`	a bare leaf loses its identity the moment it is copied out, and Zenodo's namespace is flat
proxy code	`ri3`	`sri03`	`ri3` read as "ring index variant 3"; it means scaled RI with crenarchaeol counted as 3 rings
no predictors	`none`	`p0`	reads as a value next to `.001`; position kept, CESM-style, because fixed positions are what make the id parseable
`TEXRI_cren3`	`ri3`	`tri03`	it shared a code with `scaledRI_cren3`, collapsing two distinct proxies onto one case id

Old spellings still parse, so the case directories already on disk resolve; they are simply no longer written. download_posteriors() now unpacks a flat Zenodo file into its case directory, so the local cache is one uniform layout whether a posterior was sampled here or downloaded.

Honest accounting on the leaf change: it is not free. Full path goes 39 → 72 chars against a bare fwd.nc. The leaf — the part you publish and read — goes ~100 → ~41. The genuinely free option was dropping the case directory entirely and going flat; the directory was kept because it groups a calibration with its reconstructions and gives one local layout.

scripts/migrate_cache_layout.py does the eventual move. Dry-run by default, refuses on any collision, copies-and-verifies before pruning, and skips inverse posteriors entirely (see 5D).

What is actually broken — verified 2026-08-11, not from reading docs

5A inv_relpath() is dead code in production. It is the documented canonical inverse-name builder, exported in naming.__all__, and nothing outside tests/test_naming.py calls it. The real save path is io._generate_filename_base() (stan/io.py:316), which reimplements a different leaf format inline:

| | produced |
|---|---|
| `inv_relpath()` | `inv.<site>.<cc><k>[-<scenario>]-<NNN>.nc` |
| `_generate_filename_base()` | `<case>/inv.<site>.<cc><k>[-<tag>]` |

The production path has **no run number** and folds scenario and run into
one undifferentiated tag list. Two spellings of one format is how a naming
scheme rots. Fix: delete the inline branch and call `inv_relpath()`.
Verify: `tests/test_naming.py:301-323` currently asserts the *inline*
behavior, so those two tests must be updated in the same commit.

5B save_invT_posterior() is entirely case-unaware. The public, __all__-exported entry point (stan/io.py:269) builds f"{site}_{name}_{ttype}.nc" by hand. It never consults fwd_case, never calls _generate_filename_base, and silently drops proxy_name — so a scaledRI and a TEX86 reconstruction of the same site overwrite each other. Two invT save paths disagreeing is worse than either alone. Route both through one function.

5C Forward case ids collide — 2 of 17 today. case_from_attrs() defaults the run/member token to .001, and filename_suffix is not recoverable from the attrs, so a refit and its original land on the same id:

```
..._SST_gdgt23ratio_no3_1.0_scaledRI_cren3.nc              -> tx.v026.GHEA.sst.ri3.G23-N10.001
..._SST_gdgt23ratio_no3_1.0_scaledRI_cren3_041626_eiv.nc   -> tx.v026.GHEA.sst.ri3.G23-N10.001
```

(Same collision on the `thermoT` pair.) Migrating in this state would
**destroy one posterior of each pair.** Fix: persist the run token as an
attr (`case_run`) at save time so it survives a round-trip, and have the
migration script derive it from the legacy date stamp for existing files.
A migration must refuse to run while any collision remains.

5D The 35 cached invT posteriors have no recoverable parent. Checked every one: 0 of 35 carry a fwd_case attr. Worse, it cannot be reconstructed from the filename — an invT model name records curve, structure and constraint (gen_logi_fixed_multiv_unconstrained) but not the training set or estimator, which is exactly what the compset encodes. build_invT_inputData attaches fwd_case now (stan/invT.py:311), so anything run from today forward is fine; these 35 predate it.
```
**Recommendation: do not migrate them.** Leave them under legacy
dual-read, which already works, and let them age out as sites are re-run.
Guessing a parent case by matching temptype + predictors would be a guess
recorded as provenance — the one thing a naming scheme must never do.
```

5E Zenodo is the freeze point. Unchanged and still the schedule driver. utils/download.py:79 hardcodes five posterior filenames exactly as published on 10.5281/zenodo.20032542:

```
"filename": "gen_logi_fixed_hier_crtp_multiv_priorApprox_eiv_SST_gdgt23ratio_no3_1.0_scaledRI_cren3.nc"
```

Those are what every reader of the paper downloads. A published DOI's
files cannot be renamed in place; changing them means a new deposit
version, and the accepted paper's data-availability statement points at
whichever version it cites. **The `.nc` naming must be final before the
deposit the paper cites, and is frozen forever after.**

5F Migration script + doc sweep. Only after 5A–5C. A scripts/migrate_cache_layout.py that is dry-run by default, refuses to proceed on any collision, copies rather than moves until verified, and reports every source → destination. Then update the hardcoded long names in docs/index.md, docs/stan_models_explanation_v2.md (4 names), CLAUDE.md, and SI_code2 / SI_code3 (7 names between them).

Also unresolved: data/README.md cites DOI 10.5281/zenodo.19666745 while README.md, CITATION.cff and download.py use 20032542. Reconcile before submission regardless of the naming decision.

Recommended schedule against the 2026-09-08 deadline

Now → submission: 5A + 5B + 5C only, and rename nothing. All three are bug fixes wearing a refactor's clothes — a dead canonical function, a public API that silently overwrites on proxy_name, and an id collision that would eat data the moment anyone migrates. They are invisible to reviewers, touch no file on disk, cost no recompile, and have zero Zenodo impact. They also make the revision reruns safer, because every new posterior written from here on gets a correct, collision-free identity.
After submission, before the final archive: 5F + the Zenodo re-deposit in one dedicated session. Migrating mid-revision buys nothing a reviewer sees and risks the final figures.
5D stays "do nothing" permanently unless those 35 sites get re-run anyway.

Cached posteriors keep their old stan_model_name and their old flat filenames. That is intended — dual-read already handles both layouts, and the attr records what was actually run.

Known gaps — not blockers, but do not forget

Bounded-T forward grid is incomplete. boundedT_thermoT_gdgt23ratio_no3_1.0 does not exist; the additive model has it. The thermoT variant comparison is incomplete until you fit it. The preflight cell in SI03 prints this.
invT cache contents differ per machine — data/cache/ is gitignored, so it does not travel. Check before trusting either statement below. - Windows box, when this was written: 72 files, all global_coretop_b* CV blocks, no paleo sites. SI03's load cells report everything missing until you run download_posteriors() or set RUN_INVT = True. - Linux box, 2026-08-11: the opposite — 35 files, all paleo sites (Co1010, DSDP591, MD98-2152, ODP1172, ODP1259, ODP959, SDB, U1482, U1510, WL), and no global_coretop_b* CV blocks at all. So the CV outputs are the ones missing here.
```
Neither machine has both sets. Run `ls data/cache/TEXAS_invT_posterior_cache/`
first and believe that, not this file.
```

Facts already established (do not re-derive)

All three branches are 0 commits behind main — no rebasing needed to merge.
backup/pre-pull-20260731 is 2 behind main and fully superseded; safe to delete.
Branch file-overlap: gridT ∩ groupA = ∅; boundedT-si-figures ∩ groupA = pyproject.toml only (identical version bump).
Working tree ∩ boundedT-si-figures = the 3 files listed in Phase 3.
~~Naming scheme verified against all 17 cached forward posteriors: 0 case-id collisions, 0 round-trip failures.~~ Wrong — corrected 2026-08-11. Re-measured with case_from_attrs() over all 17: 15 unique ids, 2 collisions (tx.v026.GHEA.sst.ri3.G23-N10.001 and tx.v026.GHEA.thm.ri3.G23-N10.001 each claimed by two files). The earlier check must have tested round-tripping a supplied filename_suffix rather than recovering it from attrs — which is precisely the gap. Dual-read itself is fine. See Phase 5C; do not migrate the cache until this is fixed.
Baseline test count: 175 passed (122 before this work + 53 new).

Recovery

Symptom	Fix
Wrong files in a commit	`git reset --soft HEAD~1`, re-stage, re-commit
Working tree lost	`git stash list` → `git stash apply stash@{0}` (Phase 0.2)
Commits lost	`git reflog`, or `git reset --hard backup/pre-merge-20260809`
Merge going badly	`git merge --abort` — always safe before you commit the merge
Cherry-pick conflict	`git cherry-pick --abort` and merge the branch instead
Tests fail after a merge	`git log --oneline main..HEAD` to see what came in; suspect `builder.py`/`invT.py` first — they carry the most interleaved changes

FilesExpand file tree

RESUME.md

Latest commit

History

RESUME.md

File metadata and controls

RESUME — merging completed work back to main

Latest session — 2026-09-02: v0.3.1 — patch bump for post-0.3.0 fixes/features

Previous session — 2026-08-27 (desktop): v0.3.0 resubmission prep — the Zenodo swap is staged

Done this session

The proplot -> ultraplot migration had not reached environment.yml

CI: actions/checkout does not fetch Git LFS

v0.3.0 RELEASE STATUS (2026-08-27)

Release sequence (order matters — the record id must be in the tag)

Previous session — 2026-08-23 (desktop): the default calibration now ships, and the nitrate token is readable

1. The full multivariate T₀-shift calibration ships inside the wheel

2. N10 → N1p0 (and temptype= is now genuinely optional)

3. One fact, one attr

Cache audit (this machine, 2026-08-23)

Also done

⚠️ Still open

HANDOFF — 2026-09-08 (laptop): all five plans merged; repo is release-ready

Where things stand

The baseline is environment-dependent — 602/24 here, and that is correct

What changed beyond the plans

Eleven defects were found in the plan text, not in execution

A gate the plans did not have, and should have

Two decisions, resolved

Still open

Two deploys can race

HANDOFF — 2026-09-08 (desktop): three of five finalization plans merged

Where things stand

The plans do NOT arrive with a pull

Build an environment before anything else

Run the cache migration — it does not travel

Four search traps that cost real time here

Plan text goes stale — ten defects came from the plans, not from execution

Open decisions — mine, not the assistant's

Smaller things left standing

Session — 2026-09-07: repo finalization, and six branches archived as tags

Latest session — 2026-08-22 (desktop): the gridded CV finally reached Part 1

The scope was bigger than "re-run one cell"

⚠️ Two claims must be softened in the response to reviewers

What changed in the notebook

What is dirty and why

Resolved: the untracked fig8 pair

Still open

Session — 2026-08-21/22 (desktop): rename finished, Fig. S15 landed

What actually changed

⚠️ figS17 is stale — RESOLVED 2026-08-22, see the section above

Picking this up after the restart

Two edits waiting on a re-run

Reading the docs before they are published

Housekeeping done

Session — 2026-08-17 (desktop): everything committed and pushed

What this session did beyond committing

_draft has moved on, and has 7 unresolved refs

Still to do

Session — 2026-08-16 (laptop): figure renumbering + model comparison

1. Figure renumbering — the ODR figure moved to the SI

2. .gitignore boundary fixed (the one flagged "backwards, NOT yet fixed")

3. WAIC / cross-validation — found, made durable, and extended

Still to do — all closed 2026-08-17 except the refit

Latest session — 2026-08-15 (evening): rename finished, docs synced, all pushed

What is pushed where (everything below is on the remotes)

The rename, in one paragraph

Still to do (laptop)

Laptop environment caveats

Latest session — 2026-08-15: boundedT → t0shift rename

Latest session — 2026-08-14 (evening): the R2R is fully drafted

The PETM prior question is CLOSED — no run required

CORRECTION: the "intervals are 14% too narrow" result was wrong

Section 6.2 / Section 7 drafting (2026-08-14)

TWO manuscript repos — know which one you are in (2026-08-15)

THE SIDECAR LIVES ON THE LaCie DRIVE (2026-08-15)

Moving this work to another machine (the zip question)

Do NOT push this to Google Drive through the gvfs mount (tested 2026-08-14)

Still open

Latest session — 2026-08-14

Still open after this session

RESUME — merging completed work back to `main`

CI: `actions/checkout` does not fetch Git LFS

2. `N10` → `N1p0` (and `temptype=` is now genuinely optional)

`_draft` has moved on, and has 7 unresolved refs

2. `.gitignore` boundary fixed (the one flagged "backwards, NOT yet fixed")

Latest session — 2026-08-15: `boundedT` → `t0shift` rename

Phase 3 — reconcile `revision/boundedT-si-figures`

Phase 5 — shorten and systematize the posterior `.nc` filenames