Phase-E audit — scientific math workbench and repository state

Audited tree: feature/scientific-math-workbench at c23f686, worktree UNI-FLAGELLUM-math-workbench. Audit date: 2026-07-21. Method: six parallel dimension audits (truth contract, numerical rederivation, test adversarial, documentation drift, product/runtime, data provenance), each finding then put to an independent adversarial verifier instructed to refute it by default; plus a completeness critic tasked with finding what the six dimensions missed; plus direct execution by the orchestrator.

Raw findings 74 · confirmed after adversarial verification 55 · refuted 19. Confirmed severities: 2 blocking, 9 major, 18 minor, 26 informational.

The two blocking findings were not produced by the six dimensions. They were found by the completeness critic and then confirmed by direct execution. This is itself the audit's most important methodological result: a dimension set aimed at the served page audited the shop window and missed the instrument behind it.

1. Observed outcomes

Every gate in the CLAUDE.md required-validation block was executed in this worktree on 2026-07-21. Literal results:

Production and development dependency risk are reported separately, as required. cross-study:verify-raw is BLOCKED, not passed, and must never be reported as a pass from this clone.

Executable-surface sweep (orchestrator, headless). 81 corner trajectories × 200 steps (16,200 agent steps) across all four declared slider ranges at min/default/max; 320 duration-calculator cells over eight states × ten durations spanning 1e-3–1e8 s; 72 DLT cells. Results: no non-finite value; posterior and policy posterior normalise to 1 within 1e-9; surprise never negative; stators within [0,11]; speed never negative; two identical 100-step runs bit-identical; P₊ + P₋ = −dS/dt to <1e-4 relative; mean dwell equals ∫S dt to <1e-3. No defect found in this sweep.

2. Blocking findings

E-B01 — A gate renders PASS with zero evidence on disk

• Severity: blocking. Affected claim: X01_SOURCE_INTEGRITY: PASS, and by conjunction every cross-study parity claim that rests on source integrity.
• File and line: experiments/data/cross-study-motor-evidence.json (sourceIntegrity.localArtifacts), surfaced at app/math-workbench/scientific-math-workbench.tsx:377 via statusClass; guard at tests/cross-study-parity.test.mjs:37,45.
• Reproduction and evidence:

  X01_SOURCE_INTEGRITY -> PASS
  declared artifacts: 12
  present on disk  : 0
  all verified==True? True
  experiments/upstream-cache: No such file or directory
  

  The gate's own criterion begins "Every cached artifact matches its frozen SHA-256 and byte size; each ZIP passes CRC testing…". No cached artifact exists. tests/red/c2-source-byte-gate.test.mjs already fails at HEAD with exactly this diagnosis and is excluded from npm test by design.
• Root cause: the gate result is derived from frozen verified: true claims inside the evidence file rather than from current verification. tests/cross-study-parity.test.mjs:37 asserts artifact.verified === true — reading the claim — and :45 short-circuits on checkedArtifacts === 0, making the block vacuous precisely when the evidence is absent.
• Minimal correction: compute X01's status at report-generation time from artifacts actually present. When experiments/upstream-cache/ is absent, X01 must emit BLOCKED_EXTERNAL, not PASS. Remove the checkedArtifacts === 0 short-circuit so the test fails rather than passes vacuously.
• Required failing test: a test asserting that for every sourceIntegrity.localArtifacts[] entry, either the file exists and its SHA-256 recomputes, or the gate status is not PASS.
• Validation: npm run cross-study:run && npm run cross-study:verify && npm test.
• Rollback: revert the status-computation commit; the report regenerates.
• Scientific impact: this is a direct violation of the CLAUDE.md required-validation clause "If a required dataset, archive, instrument, credential, or service is absent, mark its gate BLOCKED, NOT RUN, or EXTERNAL VALIDATION REQUIRED; never report a pass." A green badge on the only served page currently asserts verified source integrity that has not been performed in this clone.

E-B02 — A green test enforces a truth-contract violation

• Severity: blocking. Affected claim: the OBSERVED truth class itself.
• File and line: lib/walkthrough.js:356-359; enforced at tests/walkthrough.test.mjs:41.
• Evidence:

  export function truthClassForMode(mode) {
    if (mode === "OBSERVED_REPLAY" || mode === "LIVE_INSTRUMENT") return "OBSERVED";
  

  assert.equal(truthClassForMode("LIVE_INSTRUMENT"), "OBSERVED");
  

• Root cause: an unpinned live serial stream is classified OBSERVED and written into every exported lesson record via createObserverRecord (lib/walkthrough.js:387). CLAUDE.md permits OBSERVED only for source-pinned recorded measurements; a live instrument frame has no source pin, no protocol id and no hash. audits/phase-b/b1-mutation-ledger.json already records this as M03b … SURVIVED — "the derived-labelled-OBSERVED route is live and unguarded" — an open, recorded, undischarged hole.
• Minimal correction: introduce a distinct class for live instrument data (LIVE_UNPINNED or equivalent) and update TRUTH_CLASSES; or require a source pin before OBSERVED is returned. Update tests/walkthrough.test.mjs:41 to assert the new class — the test currently pins the violation and must be changed with the code.
• Required failing test: a semantic test asserting that no runtime mode without a sourceBinding/pin can yield OBSERVED.
• Validation: npm test.
• Mitigating fact, recorded not used as an excuse: the walkthrough is not served at HEAD (/ redirects to /math-workbench), so no user currently sees this label. The code path and the test that enforces it are both live in the tree.
• Scientific impact: this is the exact truth-laundering the contract exists to prevent, and it is currently protected by a passing test.

3. Major findings

E-M01 — The shipped expected free energy double-counts ambiguity

• Severity: major. Affected claim: the EFE equation card (G(π) = risk + ambiguity − information gain + effort), docs/SCIENCE.md:83, and worksheet 5, which instructs the reader to compute a value the tool does not produce.

• File and line: lib/uni-motor.js:328.

• Evidence. informationGain = entropy(qOutcome) − ambiguity (:326), and risk (:318) is a cross-entropy −Σ q(o) ln C(o), not a KL divergence. Substituting:

  G_code = crossEntropy + ambiguity − (H[q(o)] − ambiguity) + effort
         = (crossEntropy − H[q(o)]) + 2·ambiguity + effort
         = KL[q(o) ‖ C]        + 2·ambiguity + effort
  

  Independent numerical confirmation (crossEntropy − H == KL verified to 1e-12; G_code == KL + 2·ambiguity verified to 1e-12). On the real model after 40 steps:

  policy	G shipped	ambiguity	G canonical	inflation
  RUN	2.087727	0.802929	1.284799	+62.5 %
  TUMBLE	2.057086	0.776965	1.280121	+60.7 %

  The error does not cancel between policies, because the duplicated ambiguity differs per policy. G(RUN) − G(TUMBLE) is 0.030642 shipped versus 0.004678 canonical — a 6.5× inflation of the decision-relevant difference, moving the policy posterior from [0.495322, 0.504678] to [0.469397, 0.530603]. This changes action selection.

• Root cause: risk is a cross-entropy while the assembly at :328 is written as if it were a KL divergence, so the epistemic term is both added (as ambiguity) and subtracted-then-re-added (inside −informationGain).

• Minimal correction: delete the + ambiguity addend at :328, giving efe = risk − informationGain + effort, which is identically KL + ambiguity + effort. Land this in one commit together with docs/SCIENCE.md:83 and scientific-math-workbench.tsx:125, so declared equation, implementation and UI never disagree mid-sequence.

• Required failing test: an independent-oracle test that computes KL[q(o)‖C] + ambiguity + effort from the posterior without importing policyTerms, and asserts equality with agent.efe to 1e-12. It must fail on the current tree.

• Validation: npm test; the closed-loop trajectory changes, so any pinned trajectory fixture must be regenerated and the regeneration recorded.

• Rollback: revert the single commit.

• Scientific impact: a shipped scientific equation is wrong, and the repository's own record predicted this gap: audits/phase-a/phase-a-findings.md:227-228 states that independent rederivation of the EFE/VFE algebra "is Phase B (B2) work and has not been done", and docs/audit/PHASE-A-FINDINGS.md:336 lists "EFE algebra verdict" under Explicitly NOT done. This audit supplies that verdict, and it is adverse.

E-M02 — Four CSS custom properties are used and never declared

• Severity: major. Affected claim: the entire visual truth-status channel on the only served page.
• File and line: app/globals.css — 21 var() references to --cyan, --gold, --void, --failure; zero declarations anywhere in the repository. :root declares sixteen different tokens (--signal, --prediction, --ground, --danger, …).
• Evidence: grep -c "var(--cyan\|var(--gold\|var(--void\|var(--failure)" → 21; grep "\-\-cyan:\|--gold:\|--void:\|--failure:" → no match.
• Root cause: the workbench stylesheet was written against a token vocabulary that differs from the one :root defines.
• Consequence: every one of the 21 declarations is invalid-at-computed-value-time. The ProbabilityRow bars render fully transparent (invisible), and the .workbench-status pass / fail / unresolved badges render identically — the PASS/FAIL/UNRESOLVED colour channel is dead.
• Minimal correction: add four aliases to :root (--cyan: var(--signal); --gold: var(--prediction); --void: var(--ground); --failure: var(--danger);) or rewrite the 21 call sites. Four lines, fully reversible.
• Required failing test: parse app/globals.css, extract every var(--name) reference and every --name: declaration, assert reference ⊆ declaration. Currently red for four names.
• Scientific impact: the badge colour is one of the seven declared resonance axes ("UI truth badge and species label"). It is currently non-functional, and this was not detected by the reported live browser testing.

E-M03 — The strongest truth claim in the repository is guarded by a vacuous regex

• Severity: major. File and line: tests/rendered-html.test.mjs:37.
• Evidence: assert.match(html, /full biological parity[\s\S]*?FALSE/i). The lazy any-character bridge plus /i lets the nav buttons' aria-pressed="false" satisfy the match. The assertion cannot fail even if the rendered parity value were TRUE.
• Minimal correction: assert the exact data-bound substring (/full biological parity <!-- -->FALSE<\/code>/), or parse crossStudy.summary.fullBiologicalParityAchieved from the flight payload and assert === false. Deleting the /i flag alone is the one-character stopgap.
• Companion: :36 assert.match(html, /B3[\s\S]*?PENDING/) is satisfied by a hardcoded literal PENDING span in the flow tab, so it passes even if the data-bound B3 prospectivity string were replaced by a fabricated result (E-m10).

E-M04 — Truth-laundering on the product's badge surface is undetectable

• Severity: major. File and line: app/math-workbench/scientific-math-workbench.tsx:80 (modelCatalog).
• Evidence: no test in the suite reads modelCatalog. Relabelling all eleven truth strings to "OBSERVED" leaves npm test (103/103), the build, tsc and lint green.
• Root cause: modelCatalog is a module-private literal inside a client component with no exported surface and no schema.
• Minimal correction: move modelCatalog to a plain data module, export it, constrain truth to a frozen allow-list, and assert that only entries whose source resolves into a source-pinned module may carry an OBSERVED-class label.
• Scientific impact: CLAUDE.md names this exact mutation class as one the suite must detect: "Mutation tests must demonstrate that truth laundering … [is] detected." It is not.

E-M05 — Five of six workbench views have zero rendered-HTML coverage

• Severity: major. File and line: tests/rendered-html.test.mjs:29.
• Evidence: the workbench is a client component whose default view is "flow"; only the flow tab is server-rendered. The equation library, gate ledgers, both calculators, the planned/not-run panel and the twelve printable worksheets are never rendered by any test.
• Minimal correction: either add a non-DOM structural test importing modelCatalog and the gate-status classifier, or drive the initial view from a search parameter so /math-workbench?view=evidence is server-renderable and assertable. The latter also makes each view deep-linkable and citable.

E-M06 — The analysis-code provenance pin is stale

• Severity: major. File and line: experiments/results/observed-experiment-report.json:9 (identities.analysisCodeSha256).
• Evidence: lib/observed-experiment.js was refactored in 09ec601 and now hashes to 85a4a2e9…; the frozen report, experiments/results/audit-manifest.json and public/observed-experiment-report.json all still declare b757971e…. No test in npm test re-verifies the pin.
• Minimal correction: re-run npm run experiment:run so all three carry the current digest, and add the manifest-verification loop (mirroring tests/science-gates.test.mjs:62-66) to tests/observed-experiment.test.mjs. The numeric report must be unchanged; the existing replay test proves it.
• Do not rewrite the preMutationSha256: b757971e… entries under audits/phase-c and audits/phase-d — those are append-only history.

E-M07 — The shipped worker bundle leaks build-machine absolute paths

• Severity: major. File and line: dist/server/index.js:14546.
• Evidence: the Cloudflare worker bundle hard-codes absolute Windows build-machine filesystem paths — including the developer's username — as the @font-face src URLs. Every deployed page emits them in its <head>, so all Geist/Geist Mono font requests fail and the local path is publicly disclosed. Eleven font files ship unreferenced.
• Root cause: a path-separator bug inside vinext's font plugin (node_modules/vinext/dist/plugins/fonts.js), not repository code.
• Minimal correction: drop next/font/google from app/layout.tsx and reference the already-emitted woff2 files with hand-written @font-face rules using /assets/… URLs. This additionally removes a build-time network fetch to fonts.gstatic.com.
• Note: established by reading the plugin source on a Windows build host. Linux/CI build behaviour is NOT VERIFIED.

E-M08 — "Prediction residual" is not a residual

• Severity: major. File and line: app/math-workbench/scientific-math-workbench.tsx:325.
• Evidence: the tile renders observation.ligandUm − agent.predictedLigandUm, but predictedLigandUm is computed from that same observation inside the same stepAgent call (lib/uni-motor.js:375). The displayed number is algebraically −signedRate·0.25 and can never be a prediction-versus-later-observation residual.
• Root cause: the prediction and the observation it is differenced against are from the same tick.
• Minimal correction: either relabel to "model self-term (one-step extrapolation offset)" and state that no prospective residual is computed, or carry the previous step's predictedLigandUm in SystemState and difference the current observation against the previous prediction — then the label is earned.
• Scientific impact: docs/VERUM.md defines RESIDUAL as "observed minus predicted with units" and PREDICTED as "committed forecast for comparison with a later observation." The shipped tile satisfies neither.

E-M09 — The documented reproduction path does not work

• Severity: major. File and line: README.md:9-14; docs/LIVING-SCIENCE-WALKTHROUGH.md:132.
• Evidence: both instruct the reader to complete the thirteen-step walkthrough at the root URL. / returns HTTP 307 to /math-workbench; the walkthrough is unreachable in the committed deploy.
• Minimal correction: add a STATUS: NOT SERVED banner to the walkthrough document and correct the README's surface inventory. See §6.

4. What the six dimensions missed

The completeness critic's findings, beyond the two blocking items already given. Each is grounded in a file read at HEAD; the first four were confirmed by the orchestrator.

1. docs/VERUM.md contradicts CLAUDE.md on the definition of OBSERVED. VERUM:6 — "crossed the declared sensor boundary with a timestamp". CLAUDE.md — "Only source-pinned recorded measurements may be labelled OBSERVED." These are different contracts, and VERUM's is what licenses both E-B02 and the "Observed ligand" label on synthetic data. Two governing documents disagree; the resolution decides whether several findings are defects at all. This must be settled before the builder's truth lattice is frozen.
2. VERUM declares a closed 11-class vocabulary and says "The UI must never upgrade one class into another." The workbench invents roughly seventeen badge strings outside it (RUNNABLE MODEL, ADVERSE RETAINED, BLIND COVERAGE FAILURE, FALSE TODAY, …). The rule is unenforceable against an open vocabulary.
3. A fabricated timestamp in a provenance artifact. scripts/verify-ito-raw-archive.py:71 writes "verifiedAt": "2026-07-17T21:00:00Z" as a hardcoded literal on every run, into experiments/results/ito-raw-archive-verification.json and its public mirror, and tests/cross-study-parity.test.mjs:47-51 reads that file as evidence. Whatever the byte-reproducibility motive, the field asserts a verification time that is not the verification time.
4. Two SHA-256 identities for one logical artifact. experiments/data/wadhwa-2022-events.json (CRLF, 32ec7ebf, 508 103 bytes) and public/wadhwa-2022-derived-events.json (LF, d119ca60, 487 890 bytes) are the same evidence pinned twice, in observed-experiment-report.json and in lib/walkthrough.js respectively.
5. Circular oracles in the science suite, not just the workbench tests. tests/cross-study-parity.test.mjs:68-73 asserts on numbers the Python script under test computed about itself; nothing rebuilds the 175×175 GMC generator or recomputes its stationary eigenvector. scripts/independent-science-check.mjs:41 contains assert.ok(Math.abs(x + (1 − x) − 1) < 1e-14), true for every finite double, and :48 and :52 emit status:"PASS" and publicArtifactMismatchDetected:true as hardcoded literals. science:verify is a required gate.
6. Four of eight green cross-study gates have no independent rederivation. scripts/independent-cross-study-check.mjs re-derives only X03, X06, X08, X09. X02, X04, X05 and X07 are not independently checked, contradicting "Numerically rederive important results through an independent implementation."
7. The Python side has no runtime identity. requirements-experiments.txt pins numpy==2.3.5 scipy==1.16.3, but nothing binds any report to them, while the gates depend on least_squares, brentq and linalg.eig. The JS side already concluded "artifact identity includes runtime identity" (tests/red/README.md) and never applied it. This audit's environment matched by luck: Python 3.12.10, numpy 2.3.5, scipy 1.16.3.
8. Open, undischarged obligations in the ledgers. SEMANTIC-COVERAGE-LEDGER.md Entry 3: AC4 FAILS, headline is 20 credited not 22; AF1 is BLOCKING and its "whether to re-run is Codex's call" decision was never made; the blind battery remains 0/10. phase-b-roadmap.v2.json lists B1/B3/B4/B5 as NOT STARTED while b1-mutation-ledger.json holds 16 results and B3/B4 results exist on a sibling branch (§5). b1-mutation-ledger.json records that the headline adverse lognormal result is defended by exactly one assertion (tests/observed-experiment.test.mjs:47-48), and that removing it leaves all three verify gates green.
9. The red tests are frozen to 9c3a644e…, four commits behind HEAD, so even when re-run they cannot measure the current tree. C2 was re-run during this audit and still fails.
10. Mathematical claims with no test at all: lib/cad.js gear geometry (144 lines, only three regex string matches); the GMC generator's stoichiometry and Arrhenius forms; the lattice J-fit's profile likelihood; logGamma's Lanczos approximation (erf was tested, logGamma was not); and lib/walkthrough.js:361-374, the only place in the repository where joules and nats appear together — tested circularly.
11. lib/observed-experiment.d.ts omits runObservedExperiment, which the module exports. Nothing tests that declarations match runtime exports, so tsc --noEmit passes over drifted types.

5. Cross-branch state divergence

Not a defect in this branch, but a repository-information fact that changes what this branch's surface means.

phase-2/b3-model-competition contains executed, committed results that this branch renders as non-existent:

The workbench is honest but stale: by the sibling branch's own rule — "Prospectivity is decided by the commit graph, not by prose" — flipping the badge on this branch would be wrong. But if this branch is what deploys, the public surface understates the completed science by two audit phases.

The B3 result also refines the headline adverse finding in a way this branch's prose does not capture: the ~0.0369-nat lognormal advantage is an event-pooled quantity, and the result file explicitly warns that motor-equal primary numbers "are NOT directly comparable to the repository's published event-pooled figures." The workbench presents the event-pooled numbers without that distinction.

Consequence for CLAUDE.md: this branch's copy is behind. Any edit here must extend the sibling branch's version, or the merge will drop the earned execution-rigor lessons.

6. Minor and informational findings

Confirmed, with file and line. Corrections are given in the plan of record (§7).

7. Refuted candidates, kept visible

Nineteen candidate findings did not survive adversarial verification. The ones worth recording because a future reader will re-derive the suspicion:

• Weibull scale and lognormal mu inconsistency. logWeibull derives its scale from weibullScale(shape) while survival() reads models.weibull.scale; the same split exists for lognormal mu. Refuted by execution: for the frozen fitted parameters the derived and stored values are identical to relative error 0 (0.6996038164387606; mu = −1.245527443530609). Density and survival are the same distribution. The structural fragility remains — a future refit that writes only one of the two would silently diverge — but there is no live defect.
• redirect() emits a cacheable 308. Refuted: it emits 307.
• Non-determinism from Date.now()/Math.random() in rendered values. Refuted by sweep: two identical 100-step runs are bit-identical.
• Pseudoreplication in the 95 % intervals. Refuted: motor-cluster bootstrap confirmed in lib/observed-experiment.js:251.
• Train/holdout leakage. Refuted: the split is at motor granularity and parameters are training-only; proven by existing gates G1–G3.
• Mobile horizontal page overflow. Refuted: no page scroll occurs. The real symptom is narrower — the gate id overprints the title and badge at 320 px. Re-filed as minor.

8. Coverage limitations

Stated so that no reader mistakes this audit's scope for the repository's.

1. No live browser testing was performed. The sandboxed browser in this environment cannot reach the host loopback (vinext dev binds [::1] only and ignores --host), and the proxy workaround was denied. All product findings are static-analysis or build-output based. Rendering, interaction and accessibility were not observed in a real browser. E-M02 in particular predicts a visual failure that has not been seen.
2. No automated accessibility run (no axe, no Lighthouse, no assistive technology).
3. Linux/CI build behaviour is NOT VERIFIED. E-M07's confinement to Windows build hosts rests on reading the vinext plugin source, not on producing a POSIX build.
4. No systematic mutation battery was run against app/math-workbench/. E-M04's hole was demonstrated by hand-mutating scratch copies, not by executing the suite against a mutated repository.
5. The cross-study Python gates were not rederived (scripts/run-cross-study-parity.py), nor the lattice/GMC/RFT modules beyond their catalogue citations, nor the CAD export.
6. No deployed artifact outside this worktree was checked. git show main:app/page.tsx returns <UniFlagellumLab />, so published main differs from this branch; whether any live deployment serves either shape is NOT VERIFIED.
7. Sibling branches were enumerated only where they bore on B3 prospectivity (§5).
8. Several findings were established by reconstructing code paths in a scratchpad. The repository was never modified; git status was clean before and after.

9. Adverse results

Kept visible, per contract, and not softened.

• A shipped scientific equation is wrong (E-M01), and the repository's own record shows it was known to be unverified.
• A gate reports PASS with zero evidence on disk (E-B01).
• A passing test enforces a truth-contract violation (E-B02).
• The strongest truth claim in the repository is guarded by an assertion that cannot fail (E-M03), and truth laundering on the badge surface is undetectable (E-M04).
• A published provenance pin is false (E-M06).
• The shipped bundle discloses the build machine's filesystem path including the OS username (E-M07).
• The truth-status colour channel is dead on the only served page (E-M02).
• Two secondary held-out comparisons cross zero and the served page says so nowhere; the M3−M1 interval is not rendered at all (PROV-06).
• Four hypothesis records carrying explicit overclaim fences are unreachable in the deployed product (TC-05).
• Two governing documents disagree on the definition of OBSERVED (§4.1).
• AC4 still FAILS and AF1 remains BLOCKING and undecided (§4.8).
• The Phase-C blind mutation battery remains 0/10.
• M2 lognormal still out-scores M3 on held-out point score.
• cross-study:verify-raw is BLOCKED; 8 high-severity dev-dependency advisories are open.

10. Release-claim status

Not established by this audit. One required gate (cross-study:verify-raw) is BLOCKED_EXTERNAL from this clone, the raw-source byte chain is EXTERNAL VALIDATION REQUIRED, and two blocking findings are open. Passing the software tests is necessary but is not proof of biological parity. The repository's own reports continue to record fullBiologicalParityAchieved: false, four of seven computational gates passing, and B3 PENDING on this branch.

Correction order and the implementation plan that depends on it are in docs/UNI-STACK-BUILDER-PLAN.md §11 (Phase 0.5 onward) and in the working notes accompanying this audit.