The gate: scripts/check.sh

The gate is the shell script scripts/check.sh, which runs the repository’s six verification scripts in verify/ and then the wiki’s own consistency check, and exits with a failure if any check fails. The project calls it “the definition of done”: work is not finished until it passes, and a check may not be weakened to make it pass. Each verification script prints one line per check, a pass or a fail together with the measured value where there is one, so the record of a run shows what was tested as well as the verdict. What the gate tests is the arithmetic of the models the scripts encode and the form of the wiki; which claims it covers, and which it leaves to derivation, review or nothing, is listed below.

Status: Method, in force — untagged — root CLAUDE.md (‘The gate’); the results it checks carry [V] where the papers cite them. In the run of 2026-10-21 every check passed and the script ended “OK — all checks pass.”

The rule#

Root CLAUDE.md, ‘The gate’: “scripts/check.sh is the definition of done. It runs every script in verify/, then scripts/check-wiki.py on the wiki, and exits non-zero on any failure. It must pass before work is called finished. Do not weaken a check to make it pass. New verification scripts get wired into it.” The script’s own header: “The definition of done for this repo. Exits non-zero if any verification script fails. New arbitration and audit scripts in verify/ get added here.”

The gate is where the iron rule becomes enforceable. A number that a current document asserts gets a check in a script, and the script runs every time the gate runs, so a later edit to a script that changes a computed number, or a change of environment that changes the arithmetic, shows up as a failure. The scripts hold their own copies of the figures they compare against and do not read the documents, so an edit to a document’s figure does not trip them.

How it runs#

• Environment. The scripts run in a Python virtual environment, .venv, built from requirements.txt, which pins numpy==2.5.4 and sympy==1.14.0. Without it the gate prints check FAILED: no .venv. Create it with: and the command to create it, and exits with status 1.
• An explicit list. The script runs verify/battery.py, verify/soliton_obituary.py, verify/toy_coset.py, verify/toy_gw.py, verify/toy_s4.py and verify/toy_s5.py, in that order, named one by one rather than matched by a pattern. “Runs every script in verify/” therefore holds as long as the list matches the folder, as it does now; a new script counts only once it is added to the list. Root CLAUDE.md’s ‘Layout’ names only battery.py and soliton_obituary.py (flag 38 in review/wiki-build-flags.md); the list in the script is the fact.
• Every script runs. The shell options are set -uo pipefail, without -e. A failing script prints its name followed by FAILED and sets a status variable; the loop continues, so one run reports every failure rather than the first.
• Then the wiki. After the six scripts it runs scripts/check-wiki.py on the whole wiki, with the same failure handling.
• The verdict. If nothing failed it prints “OK — all checks pass.” and exits 0; otherwise it exits 1.
• Deterministic. Every script that draws random numbers fixes its seeds, so a rerun in the pinned environment prints the same values. verify/toy_s4.py says so in its docstring: “Fixed seeds, so the run is deterministic.”

The check line#

All six verification scripts define the same helper, check(name, ok, detail=""). It prints PASS or FAIL, the check’s name and, when a detail is given, the measured value in square brackets, and it records the failure. At the end a script with failures prints FAILED, the count and the names, and exits with status 1; a clean script prints a closing “OK” line of its own. Lines indented without a verdict print measured values for the reader (the T4 limits, the Monte Carlo rows, the truncation sweep) and gate nothing by themselves. verify/battery.py states the convention: “Every verdict is a boolean or a number against a tolerance, never a judgement.”

Two kinds of check need distinguishing:

• a check against something known independently of the documents, an independent anchor, or a property the derivation predicts (a symmetry, a sign, a scaling);
• a check labelled “parity” or “quoted:”, which compares the output with the figure a document prints. It confirms that the document prints what the code computes and that the environment still reproduces it. It is a reproduction, and by itself it tests nothing about whether the figure is right.

What each script gates#

#11 tags the toy results [D][V] (§4, the P2′ refinement; §10, “Toy-model support”) and records them in §14 item 11. #11 §12 ends: “The [V] numbers here are gated in verify/soliton_obituary.py.”

What is not gated#

The gate covers less than the record contains. These are outside it:

• Round-3 numbers the adjudicator did not compute. Rulings 3 and 6–13 have no checks, and neither has ruling 4’s referee range: the adjudicator “did not re-run the referee’s Kibble–Zurek integration”, and rulings 6 and 9 were “Accepted, not re-checked” (review/round3-adjudication.md; Review round 3). The soliton script’s docstring nonetheless claims every number of that record (flag 33 in review/wiki-build-flags.md). The withdrawn manuscript’s benchmark arithmetic (“flux, densities, σ_T/M”), which ‘What survives v3’ keeps, lost its checks when they were retired with the paper; of that list only the W′ overclosure is gated (benchmark arithmetic).
• Round-4 numbers outside the toy note. The adjudicator’s F1 sequence and F2’s eigenvalues at λ = 10 and 100 come from review/round4/adjudicator/arb.py, and its control re-runs from review/round4/adjudicator/ctrl.py; the gate checks the boost at t = 4 and the eigenvalues at λ = μ = 1.
• Analytic arguments. The toy note’s argument that the complex model’s breaking does not survive quantization is analytic (§4). The gate checks its premise, the exact continuous symmetry, not its conclusion (quantum restoration).
• A check mark without a script. HANDOFF.md §2 marks the Debt-3 figure “criterion fails by 10²⁸–10³¹” ✓, but no script in verify/ computes it, and #11 §9 tags it [D] (flag 16 in review/wiki-build-flags.md; the mean-field criterion). A wiki-build recomputation, review/wiki-build-checks/debt3a_check.py, reproduces it and “could be moved into verify/” (flag 16).
• Whole documents. verify/ holds no script for the numbers of the Debt-3 calc notes or of #13 (the Debt-3 calc notes; conjugate special relativity).
• Review and wiki-build scripts. The referees’ and adjudicators’ scripts stay in review/ and are not run: review/round3/adjudicator/round3_arb.py, review/round4/adjudicator/arb.py and review/round4/adjudicator/ctrl.py, with the referees’ scripts beside them. Root CLAUDE.md’s layout says of verify/ that “arbitration and audit scripts go here”, and the header of scripts/check.sh says the same; the sources do not settle the difference, so it stands as a gap (flag 38 in review/wiki-build-flags.md). The gate carries their corrected results instead, in verify/soliton_obituary.py and in checks labelled with round-4 findings (F3, F4, F6) in the toy scripts. The executed checks of the wiki build are likewise outside: “none is wired into scripts/check.sh” (review/wiki-build-checks/README.md).

The wiki check#

scripts/check-wiki.py applies the conventions of wiki/CLAUDE.md to every page. Its docstring lists the checks in order of severity:

• link — an internal link points at a page that does not exist, or is not an absolute /wiki/ link;
• anchor — a link points at a heading that does not exist, with each heading’s anchor computed the way Hugo computes it;
• heading — a heading holds a link, an image, a bare web address, emphasis or an attribute block such as {#id}, for which Hugo builds the anchor from only part of the heading’s text or takes it from the block, so the script cannot compute it;
• map — a page is not linked from its parent section’s index;
• source — a frontmatter source, or a repository path in inline code, names a missing file;
• frontmatter, weight, section — missing or invalid fields, siblings sharing a weight, a folder without an index, a file named claude.md, or two files that serve at one address;
• gist, statusline — the page’s opening: the italic passage and the status line;
• star — an unescaped starred symbol such as v*, whose asterisk can end an italic passage early;
• h1, fence — no top-level heading, and a language on every code fence;
• acronym — an acronym that is unregistered, or used without its expansion nearby or a link (registry: scripts/wiki-acronyms.txt).

Its reason for existing, from the docstring: “Hugo does not validate internal links: it renders /wiki/does-not-exist without complaint. This script catches a dead link in the Markdown, before any build.” Two later checks belong to the gate of the site that publishes the wiki, npm run check in site/ (site/CLAUDE.md): the build fails on a link to a missing page, and scripts/check-site.py checks the built pages, among other things that every heading got the anchor this script computes. Given file names, it reports only on those files while resolving links across the whole wiki; with --stats it also prints status counts and the pages with fewer than three inbound links from outside their parent’s map.

It checks form, not content. It does not compare a page’s numbers, quotations, tags or statuses with the source documents; that comparison was the work of the wiki build’s verifiers, whose suspected source errors are in review/wiki-build-flags.md with executed checks in review/wiki-build-checks/.

How the script list grew#

The commit history records each change:

• 2026-10-19, commit 008c3ee. The first gate ran verify/battery.py alone (“add battery gate”).
• 2026-10-19, commit 678e48a. knot_numbers.py was added (“Recompute every number in #14 v2”); it later checked every number of the soliton manuscript’s third version.
• 2026-10-19, commits d15cd4a and 5dbaa82. dark_sector.py joined the list for the worklog’s items O1 and O2 (“Script O1/O2: minimal dark sector overcloses; needs a light doublet”); when v3 was cut it was “merged into” knot_numbers.py and left the list again.
• 2026-10-19, commit 96f58da. After round 3, knot_numbers.py was given the corrected inputs and independent anchors, and the commit message records the consequence: “scripts/check.sh fails on v3’s affected claims, by design, until v3 is revised or withdrawn.”
• 2026-10-19, commit 5b6efaa. The manuscript was withdrawn and knot_numbers.py replaced by verify/soliton_obituary.py. The adjudication’s ‘Decisions’ explains why this is not a weakened check: “This retires the checks of a withdrawn paper. It does not loosen the gate: the gate verifies what current documents claim, and v3 no longer claims anything.” The commit message: “the gate is green again because the withdrawn paper no longer claims anything.”
• 2026-10-19, commits 0e0f51e to 56485d3. The four toy-model scripts were added chunk by chunk as the models were built.
• 2026-10-20, commit 17e1338. scripts/check-wiki.py was added (“Add the wiki gate: links, anchors, section maps, sources, page shape”); on 2026-10-21 it began checking starred symbols (commit b970e9d; timeline). On 2026-10-22 it began computing heading anchors as Hugo does (commit b41f199), then rejecting headings whose anchors it cannot compute (commit 503d34d) and two files that serve at one address (commit 5540adc).

A recorded run#

The closing lines of the run of 2026-10-21, after the six scripts’ own closing lines:

OK — S5 (weak form) and the section-condition structure check out.

wiki OK — 278 pages, no problems

OK — all checks pass.

The output of each script is quoted on its page, in full except for parts A and B of verify/toy_s5.py: the battery, the soliton-obituary script and the toy-model scripts.

A name shared#

“The gate” in the physics pages is a different thing: the velocity gate that switches conjugate matter’s self-interaction off at high relative velocity.

Sources#

• scripts/check.sh — the script list, the environment check, the failure handling, the final line
• scripts/check-wiki.py — the wiki checks, their order and their stated purpose
• Root CLAUDE.md ‘The gate’, ‘Layout’ — the definition of done; the scripts’ folder
• requirements.txt — the pinned environment
• verify/battery.py, verify/soliton_obituary.py, verify/toy_coset.py, verify/toy_gw.py, verify/toy_s4.py, verify/toy_s5.py — the check helper, the parity convention, the seeds
• papers/broken-born-checkpoint-v2.md §12, §14 item 11, and papers/reciprocity-toy-models.md §2–§5 — the claims the scripts gate
• review/round3-adjudication.md ‘Rulings’, ‘What survives v3’, ‘Decisions’ — the unchecked rulings; the retired benchmark checks; the script replacement
• review/round4-adjudication.md header — the arbitration scripts in review/
• review/wiki-build-checks/README.md and review/wiki-build-flags.md items 16, 33, 38 — the wiki build’s checks, outside the gate; the Debt-3 check mark; the soliton script’s claimed coverage; the Layout’s two scripts
• HANDOFF.md §2 — the ✓ on the Debt-3 figure
• wiki/CLAUDE.md ‘The gate’ — how the wiki check is run
• site/CLAUDE.md ‘The gate’ and scripts/check-site.py — the site’s later checks
• Git history: commits 008c3ee, 678e48a, d15cd4a, 5dbaa82, 96f58da, 5b6efaa, 0e0f51e, 56485d3, 17e1338, b970e9d, b41f199, 503d34d and 5540adc — the growth of the script list and of the wiki check