ADR-0050.

Date: 2026-08-12 Status: accepted 2026-08-17 (the five executable proofs were run; record below)

Context

The product review of 2026-08-12 read the published surface against what the gate demonstrably does. Every location below was re-verified by grep on that day, on runward 0.33.5 with check --strict at exit 0 on this repository. The site says, in both languages:

• "aucune IA ne peut l'embobiner" / "no AI can fool it" (runward-site index.html:151,1768, js/i18n-en.js:24);
• "Sécurisé dès le premier jour" / "Secured from day one" (index.html:152,1769, js/i18n-en.js:25);
• "la rend impossible dès la conception" / "makes it impossible by design" (index.html:1166, js/i18n-en.js:105);
• "impossibles à embobiner par un prompt" / "impossible to talk into passing with a prompt" (content/docs/case-study.fr.md:37, docs/en/case-study.md:39, plus the generated copies under docs/);
• "que rien ne peut embobiner" and "aucun prompt ne l'embobine" (content/docs/compare.fr.md:36,108, plus generated copies).

What the gate actually does is narrower, and the narrowing is this project's own doctrine. 64 rules ship in templates/rules/; exactly one carries a machine-verifiable signature: (frontier-deterministic-boundary.md:6), and that signature's own nonScope says it proves a shape exists, not that it does its job. ADR-0045 closed fabrication and circularity after 22 proven false greens; its non-scope is explicit and unchanged: the gate verifies that evidence is present, resolves, and has not been tampered with, never that the evidence is about the rule it is attached to. Re-verified 2026-08-12: a real file, cited and sealed, whose content has nothing to do with its rule, passes check --strict. "No AI can fool it" is a claim about relevance, and relevance is the declared non-scope. A reader who tests the sentence wins against the site. That is the exact failure mode this project exists to refuse, printed in its own hero.

Three more defects on the same surface, same verification date:

1. A competitive claim nobody can replay. "un contrôle au niveau du code (juillet 2026) confirme qu'aucun autre scaffold ne les réunit" (content/docs/compare.fr.md:49, EN compare.en.md:49, plus generated copies). No script, no corpus, no dated capture exists in either repository. The amendment to ADR-0045 already names this class: an event nobody else can replay is an assertion, not evidence. "Confirms" over an unreplayable event is the same defect in marketing form.
2. Translations presented as verbatim. compare.fr.md:31 qualifies the BMAD citation "advisory … fournit des recommandations, ne bloque pas" as "mot pour mot"; the source sentence is English ("advisory … provides recommendations, not blocks"), so the French is a translation wearing verbatim's clothes. Same pattern on the Spec Kit and OpenSpec citations (compare.fr.md:30,32). The EN file's citation is a true verbatim; only the FR misstates its own nature.
3. Four compare pages predating the gate's own doctrine. The SPA pages #/apres-la-spec/{spec-kit,openspec,bmad,spec-kitty} (index.html:1534-1603) contain zero occurrences of "gate" across lines 1440-1605 (grep -c = 0, 2026-08-12); their last substantive edit is 2026-07-05 (git log), before ADR-0040 through ADR-0049 made the gate the central object. docs/compare/index.html, generated from current sources, says "gate" 8 times. The same site argues the comparison with and without its main argument.

The guard already names its own hole. test/unit/no-overclaim.test.js scans [README.md, docs, templates, runward, src] of this repository only (line 24), and its header (lines 3-7) records that the overclaim caught in the week of 2026-08-04 lived in the SITE repo, "which that test cannot read". The site repo has no runward/ directory and no test CI; its only workflow is sync-runward-docs.yml (daily bump of the pinned runward devDependency, resync, npm run build fail-loud). And the surface a visitor receives is built, not committed: build-i18n.mjs generates public/en.html; measured 2026-08-12, the stale root en.html carries 0 occurrences of "fool" while the built public/en.html carries 2. A guard that reads only the repo files can be green while the served page overclaims.

The worst case, named first. The provable sentence is commercially weaker than the forbidden ones. "Vérifie la présence, les pointeurs et l'intégrité, jamais la qualité" sells less than "no AI can fool it". This ADR chooses the weaker sentence anyway, because the alternative is a site that loses an argument with its own product the first time someone runs it.

Decision

1. The ceiling sentence. The strongest thing either repository may publish about the gate is the sentence already printed at content/docs/case-study.fr.md:51: runward check --strict "vérifie la présence, les pointeurs et l'intégrité de ce dossier, jamais sa qualité : la qualité, c'est ce que vous venez de lire, et la juger reste le travail d'un humain." Presence, pointers, integrity; never quality, never relevance. Every sentence stronger than this, in any phrasing and either language, is a violation.

2. One source of rules, shipped by the package. The RULES of no-overclaim.test.js move to a data file published with the npm package (dist/claims-rules.json or a runward/claims export); the CLI test keeps consuming them, and its three meta-guards survive unchanged: the safe-list tested in both directions, FROZEN_CITATIONS, and the scope assertion (> 100 files, named files). New rule classes, each word-bounded and each with its unless escape for negations and forbidden-list entries: absolutes about the gate (embobiner, fool it, talk into passing); "impossible by design" / "impossible dès la conception"; "secured from day one" / "sécurisé dès le premier jour"; a dated competitive claim with no replayable pointer beside it; "mot pour mot" or "verbatim" applied to a translated citation. Every new rule is proven in both directions before merge: it fires on the exact line it was written for, and stays silent on the corrected copy. The lesson is already written four times in that file: a guard that cries on the safe case gets switched off.

3. The site build fails loud on a claim. scripts/check-claims.mjs on runward-site consumes the rules from the pinned runward devDependency (the daily sync already carries new versions within 24 hours), and scans the BUILT surface public/ as well as the sources (index.html, js/i18n-en.js, content/, news/). It runs as the last step of npm run build, so a violation reds the Vercel deploy; a PR workflow runs the same check. Scanning is word-bounded on text, like the CLI guard, because HTML attributes and class names are noise. Before this guard is trusted, one violation is seeded on purpose and the red build demonstrated, then reverted: a guard that has never been seen failing is decoration.

4. The copy is corrected at the sources, FR and EN, then regenerated. The hero (index.html:151-152,1768-1769, keys hero.v1/hero.v2 of i18n-en.js:24-25) and the security section (index.html:1166, i18n-en.js:105) drop "impossible" for the actual mechanism: the trifecta kept incomplete (untrusted input, sensitive tool, exfiltration path, never all three) and external tools pinned to a known version. Case-study and compare reuse the ceiling sentence to the character, then gen-docs regenerates. The July claim either becomes replayable (dated corpus plus script, committed) or is rewritten as a dated observation against the sources [1]-[5] the page already cites, without "confirme" and without "aucun autre"; that choice belongs to the author and is not made here. Translated citations carry the EN original verbatim in quotes with the FR marked as a translation (the FROZEN_CITATIONS doctrine, no-overclaim.test.js:149-151). The site copy is a per-character contract (site CLAUDE.md): every rewording passes the author before it is written.

5. The four SPA compare pages name the gate. The word and the mechanism (deterministic verdict, no model in the path, never a judgment of quality) appear in each "what runward adds" cell of index.html:1534-1603, aligned with docs/compare which already names it 8 times. Text only; the design contract forbids a structural redesign and none is needed.

Alternatives considered

• Duplicate the rules list in the site repo. Two lists diverge in silence; the guard's own header documents that a guard whose reach is not asserted quietly shrinks to nothing. One source of truth, consumed on both sides, is the whole point of decision 2.
• Govern runward-site with runward (npx runward init on the site). The gate verifies the artifacts of a mission, not marketing copy: it would catch none of the claims listed above. It is also heavier than the problem, and deferrable with its trigger written down.
• A human checklist before publication. It existed: the adversarial fact-check doctrine predates every overclaim above, and they shipped anyway. The entire lesson of ADR-0045 is that vigilance is not a mechanism; a deterministic guard that breaks the build is.
• A new CLI command runward scan-claims <dir>. That grows the product surface to solve the maintainer's marketing problem. ADR-0039 keeps the operator layer outside the CLI; a data export consumed by a site script respects the boundary, a command does not.
• Correct the copy without installing the guard. The 2026-08-04 review already caught a site overclaim, and here is another one four weeks later. A correction without enforcement is the drift re-instructed, not a decision.

Consequences

• The worst case first: the headline gets weaker. The true sentence covers presence, pointers and integrity; never quality, never relevance. That is assumed and it is the advertised price of this decision: a reader who tests the gate finds exactly what the site announces, which is the only durable marketing this project can afford.
• The build becomes fail-loud on claims. A false positive now blocks a legitimate deploy; hence the obligation, inherited from the guard file's own lessons, that every rule ship with its unless escape and be tested in both directions before any merge.
• The rules file becomes a published, versioned surface of the npm package. A rule change travels with the release and reaches the site within 24 hours through the existing daily sync; no new update route to maintain.
• A reintroduced overclaim fails the day it is written, on the built surface, not at the next human review. The drift can no longer come back silently, which is what distinguishes this from the checklist it replaces.
• Cost: 4 to 5.5 maintainer-days. 0.5 (this ADR) + 1 (rules externalized, new classes, two-direction tests) + 1 (site guard, build and CI wiring, demonstrated red build) + 1.5 to 2 (copy FR/EN, author validation included) + 0.5 to 1 (the four compare pages). Making the July claim replayable would add 1 to 2 days of corpus and is not counted: it awaits the author's decision named in decision 4.

What this does not claim

This ADR does not make the gate unfoolable. Relevance remains unverified: 1 signed rule out of 64, and a real but unrelated proof still passes check --strict. That is precisely the claim being removed, not repaired; repairing it is a different piece of work with its own trigger, and nothing here advances it. It does not cover surfaces outside the two repositories: LinkedIn posts, HN comments and the PDFs on thibaultsouris.fr stay under the human verification doctrine, which this ADR neither strengthens nor replaces. It does not make the July competitive claim true: it makes it replayable or removes it. It does not replace human judgment of quality; neither does the gate, which is the ceiling sentence's entire content. And it crosses no phase of any mission: an ADR in proposed status is an argument, not evidence.

Ratification

Accepted 2026-08-17. The five proofs were executed, in both directions where the criterion says so. What was run, and what it returned:

1. The grep is clean. grep -rniE "embobin|fool it|impossible (dès la conception|by design)| sécurisé dès le premier jour|secured from day one" over the built public/ returns 0 matches outside the generated reference mirror (docs/(en/)reference/, exempt like vendor/adr: it is the product's own record, governed upstream by the CLI's no-overclaim, and its ADRs quote the forbidden phrases in order to forbid them).
2. The guard fires in both directions. node scripts/check-claims.mjs exits 0 on the corrected site (381 files, 11 package rules + 5 site classes) and exits 1 on a seeded violation, reverted immediately after. The site classes carry their own unless escapes rather than the package NEGATED screen — the gate absolutes ARE negative sentences, so NEGATED would have exempted the exact original violations.
3. npm test is green on the runward repository with the rules externalized and the scope assertion intact (> 100 files, named files).
4. The build fails loud: the guard runs as the last step of npm run build, so the seeded violation of proof 2 reds the Vercel deploy — demonstrated, then reverted.
5. The four SPA compare pages name the gate (8 occurrences of the gate and its mechanism across the four cells, FR and EN), and the July competitive claim is absent — rewritten as a dated observation with no "confirme" and no "aucun autre" (decision 4's second branch).

The copy now stands at the ceiling sentence of decision 1, in both languages, and cannot drift back: the guard is a build step, not a habit.

The criteria as originally set, kept as the record of the bar:

1. grep -rniE "embobin|fool it|impossible (dès la conception|by design)|sécurisé dès le premier jour|secured from day one" over runward-site public/ after a fresh build returns 0 matches outside negations.
2. node scripts/check-claims.mjs exits 0 on the corrected site, AND exits non-zero when a forbidden sentence is reinserted: both directions, exactly like the CLI guard's own tests.
3. npm test green on the runward repository with the rules externalized and the scope assertion (> 100 files, named files) intact.
4. One red Vercel build demonstrated on a seeded violation: proof commit, observed failure, then revert.
5. The four SPA compare pages contain the word "gate" (grep -c > 0 on index.html:1440-1605), and the July claim is either backed by a committed replayable script or absent.

Reevaluation trigger (mandatory, dated)

Trigger set on: 2026-11-05.

The decision is wrong and must be revisited if any holds: the claims guard cries on legitimate copy often enough that someone is tempted to switch it off (the failure mode this project has already paid for four times); the ceiling sentence turns out to lose readers who would have adopted on a truthful stronger one, measured rather than assumed; or a forbidden class is found to be under-specified by a claim that slipped through the guard and reached the built site.

Watched via: the site build (the guard is a build step), and the conformance-gate incident log.