Editorial revision — 2026-10-17: the original entry was published on 2026-08-20, with the addendum below on 2026-08-23. Its claim that visual coherence caused less checking has been withdrawn: the audits observed defects, not reader behavior. Original wording.
The finding class
This site presents itself as evidence-bound: claims carry markers, markers carry links, a ledger carries bindings. Two audits of the deployed site found the same defect class at two depths: the interface ran ahead of the verification architecture. Version 0.1 had stale project descriptions behind confident chips. Version 0.2 fixed the copy and built real machinery — and still asserted, in five places, mechanics it did not have. These observations establish a mismatch between the interface’s claims and its implementation; they do not establish an effect on readers.
- Said
- The ledger page claimed it was rendered from
claims.yaml. - True
- The ledger was hand-written to match the registry; CI only checked that each claim ID appeared somewhere in the HTML. Mirroring, not generation.
- Now
- The ledger is generated by generate_ledger.py; CI regenerates it and fails on any drift. Editing the ledger by hand now breaks the build.
- Said
- Claims were "bound" to commit SHAs.
- True
- The verifier checked that support URLs resolved (HTTP < 400). A resolving URL is liveness; it is not a binding — nothing compared the declared commit to the URL.
- Now
- The verifier requires the declared commit to appear inside the support URL, and the registry schema treats a commit without a matching URL as a failure.
- Said
- The hero promised event-responsive claim governance.
- True
- One global 120-day timer existed. Time-based review is real but it is not evidence-responsiveness; the named triggers (kernel change, manifest rewrite) were prose.
- Now
- Triggers are typed and tagged. Executable triggers run in CI: the bound file is fetched at its bound revision and at the default branch head, and divergence fails the build — the kernel module for CC-001, the manifest for CC-002, the thesis for GA-001, the README for GV-001, and this site's own DESIGN.md by local hash. Triggers CI cannot execute (certification expiry, a private artifact going public) are labeled manual, visibly, instead of borrowing the executable ones' credibility.
- Said
- The ledger correctly called "attested" a provenance state.
- True
- The schema then put
attestedin the same enum as untested/supported/contradicted — collapsing the exact distinction the site lectures about. - Now
- Schema v0.2 separates visibility, provenance, support role, evidential status, and maturity. An owner-attested claim with no public artifact now reads what it is: owner-attested · publicly untested. That is less flattering and more true, which is the trade this site exists to make.
- Said
- The design ledger claimed "CLS 0," "zero supply-chain surface," "no pipeline, nothing to break," and — after the first correction pass — still said the empty result card was "stamped inconclusive."
- True
- Layout shift was never instrumented; the supply-chain surface is small, not zero; a pipeline now deliberately exists; and the card had already been corrected to untested — the ledger describing it hadn't.
- Now
- All four corrected in DESIGN.md, and DESIGN.md itself is under a local-hash trigger: it can no longer change without the registry noticing.
Two audits found unsupported claims behind a coherent interface. We have not tested whether presentation discouraged checking. Following those audits, the site added generation, binding, trigger, hash, and kernel checks. Those checks cover named failure conditions; they do not measure reader comprehension or establish that every scientific inference is valid. Inspect current check outcomes and review dates.
What v0.3 still does not do
- The main page's prose and chips are hand-written; only the ledger is generated. A claim-component generator for every rendered surface remains future work.
- Executable triggers watch single bound files. Semantics can change outside a watched file; those triggers stay manual and say so.
- Continuous reproduction covers CC-001 — one claim, one kernel call. It is the template, not the coverage.
- No field performance data exists and none is claimed; the site runs no analytics by design.
- The central hypothesis — that claim-bound interfaces improve a reader's comprehension, calibration, or evidence-seeking — is untested. Untested, not inconclusive: no comparison against a conventional control version has been run. The null stands until it is.
The record improves by prosecution, not by polish. Findings against this site are welcome — bring the falsifier.
Addendum — 2026-08-23 (v0.4)
Residuals from the list above that have since moved, dated here so this log stays a log rather than being rewritten:
- The homepage's registry facts (claim count, last owner review) are now cross-checked by
verify_claims.py— a hand-written strip can no longer silently strand when the registry grows. - Continuous reproduction now covers CC-004's endpoint witnesses as well as CC-001's bounds, and reads its expected values from the registry instead of hard-coding them.
- The module pages and the claim observatory joined the ledger as generated, drift-checked surfaces.
- One more instance of the finding class surfaced during the v0.4 audit: DESIGN.md
asserted a build-time geometry check on the bounds figure that the repository did not
contain. The checker now exists (
scripts/verify_figures.py), runs in CI, and the gap is recorded as a dated correction in DESIGN.md itself. - CC-001's "immutable binding" was found pinned to a commit stranded by an upstream history rewrite — publicly resolvable only as a dangling GitHub object. It was re-bound to reachable history with the kernel bytes verified identical, and the rule is now enforced: verify_claims.py checks every bound commit reachable from its repository's default branch, one filtered clone per repo, in the same CI run.