@bounded-systems/conformance-kit

A standalone, site-agnostic web-conformance toolkit: build-integrity tooling, fail-closed conformance gates, and provenance generators — extracted from bdelanghe/site and bounded-systems/site and generalized so a site vendors one kit instead of duplicating scripts.

Every site value (paths, thresholds, site URL, account/repo id, issuer/DID, SHACL shapes, the markdown renderer, the prose corpus, the build itself) is an INPUT, injected by the consumer via CLI args, env vars, or a passed config. Nothing here hardcodes robertdelanghe.dev, bounded.tools, an account, or an email.

integrity/    verify-site · verify (sigstore) · gen-sitemanifest · gen-provenance · structure-audit · http-probe
gates/        sbom (gen + completeness) · shacl-runner · seo-gate · axe-gate (axe-core a11y) · vuln-gate (npm audit) · html-validator-gate (vnu) · baseline-gate (web-features) · jargon-gate (plain-language) · readability-gate · commonmark-runner · semantic (lone)
gates/        Token Accessibility suite (static token-level a11y → TOKEN-A11Y.md): palette · pairing-extractor · typography · target-size · opacity-contrast · likeness · token-a11y (unified runner)
gates/conformance/  conformance-report — lone's conformance() projection (Node port of jsr:@bounded-systems/[email protected]) + a generic HTML renderer
generators/   gen-cid (IPFS UnixFS) · gen-identity (did:web + VC) · gen-snapshots (reader/markdown) · gen-print-snapshots (PDF) · openapi (static-API helper core)
emitters/     reprDigest (RFC 9530) · securityTxt (RFC 9116) · webManifest · markdown-sibling headers
lib/          schema-validate (zero-dep JSON Schema) · config (env/arg helpers)
fixtures/ test/  isolated verification of the generic logic

Design rules: zero-dep where the source was zero-dep; pure/offline gates read only the built output; deterministic generators are a function of their inputs (no wall clock); fail-closed (exit 1) on any violation.

Install / vendor

Three consumption models:

1. Vendor (recommended, matches the existing vendor/integrity/ pattern). Copy the kit at a pinned commit into vendor/conformance-kit/, write a hash-pin manifest (see vendor.example.json — mirrors bdelanghe/site vendor/integrity/provenance.json: source, commit, fetched, files{path: sha256}), and verify against it before every use. The site then imports / invokes the vendored copies. The kit's own provenance.json records which source repo + commit each tool was generalized from.
2. npm dep. npm i @bounded-systems/conformance-kit and use the ck-* bins (see package.json) or import the library modules.
3. Nix flake (reproducible, runtime-bundled). nix run github:bounded-systems/conformance-kit#ck-axe-gate -- dist, or add the flake to a home-manager / nix profile. Each ck-* bin is a hermetic, pinned closure; the gates that shell out get their runtime bundled in — ck-html-validator-gate carries a JRE for vnu, ck-vuln-gate carries npm — so no JRE/Node on $PATH is needed. (ck-axe-gate still needs a browser the consumer supplies via $AXE_RUNNER: tezcatl or Playwright.)

Runtime deps are declared in package.json (only the gates that need them pull them: linkedom/@mozilla/readability for structure-audit; jsonld/n3/ @zazuko/env-node/rdf-validate-shacl for the SHACL runner; sigstore for the in-process verifier). The Deno semantic runner pins its imports in gates/semantic/deno.json.

Tools — what each does + how a site consumes it (the input it must supply)

integrity/

| Tool | Invoke | Consumer supplies | |---|---|---| | gen-sitemanifest.mjs | DIST=dist node …/gen-sitemanifest.mjs | $DIST (build dir). Optional $MANIFEST_EXCLUDE (extra platform control files). Emits $DIST/site.sha256. | | gen-provenance.mjs | run at deploy after signing | GitHub Actions env (GITHUB_*), $OCI_REF/$OCI_DIGEST, optional $PROVENANCE_DOC_URL, $DIST. The emitted builder.repository becomes the identity the verifiers enforce. | | verify-site.mjs | node …/verify-site.mjs <https://site \| ./dist> | A deployed site (or local dir) carrying provenance.json + site.sha256 + its .sigstore.json bundle. Identity is read from provenance.builder.repository — nothing hardcoded. Shells to cosign if present, else SKIPs with a recipe. | | verify/verify.mjs | node …/verify/verify.mjs <url\|dir> | Same inputs; verifies the Sigstore bundle in-process (offline) via sigstore-js. | | structure-audit/audit.mjs | node …/audit.mjs <distDir> [--check] | <distDir>. Optional $STRUCTURE_ARTICLE_PREFIX (default blog/), $STRUCTURE_ERROR_PAGE (default 404.html), $STRUCTURE_AUDIT_SIDECARS (deploy-time live paths, e.g. /resume.pdf), $STRUCTURE_BASELINE (where the committed structure.json lives — keep it in the consumer, not the vendored kit). | | http-probe.mjs | node …/http-probe.mjs <https://site> [config.json] | A live URL and a probe config: $PROBE_CONFIG/2nd arg JSON {htmlRoutes,typed,missing}, or $PROBE_HTML_ROUTES+$PROBE_MISSING. Routes are NOT hardcoded. |

gates/

| Tool | Invoke | Consumer supplies | |---|---|---| | sbom/gen-sbom.mjs | ROOT=. DIST=dist node …/gen-sbom.mjs | $ROOT (lockfiles live here), $SBOM_LOCKFILES (comma list, default package-lock.json), $SBOM_NAME, $SBOM_NAMESPACE_BASE, $SBOM_CREATORS. Reads flake.lock if present. Emits $DIST/sbom.spdx.json. | | sbom/check-sbom.mjs | ROOT=. DIST=dist node …/check-sbom.mjs | Same $ROOT/$DIST. Fails closed unless pinned-set ⊆ SBOM ⊆ pinned-set and (optionally) the in-toto attestation reconciles. | | shacl-runner.mjs | node …/shacl-runner.mjs <shapes.ttl> <htmlDir> | The SHACL shapes file stays in the site (its structured-data contract) + the built-HTML dir. Optional $SHACL_CONTEXT (custom offline JSON-LD context; default schema.org). Fails unless every JSON-LD block conforms: true. | | seo-gate.mjs | node …/seo-gate.mjs [distDir] | $DIST. Optional $SEO_ERROR_PAGE, $SEO_DEPLOY_SIDECARS. Enforces canonical/title/description uniqueness + self-consistency, robots.txt (RFC 9309), sitemap, internal links. | | axe-gate.mjs | node …/axe-gate.mjs [distDir] | $DIST. Optional $AXE_PAGES (comma list, default: every *.html in dist), $AXE_TAGS (default wcag2a,wcag2aa,wcag21a,wcag21aa,wcag22aa), $AXE_IMPACT_THRESHOLD (minor/moderate/serious/critical, default serious), $AXE_RUNNER (playwright (CI, needs playwright + @axe-core/playwright + npx playwright install chromium) | tezcatl (macOS WebKit, local)), $AXE_REPORT (write the JSON report). Serves dist over an ephemeral origin (so assets resolve), runs axe-core per page, and fails closed on any violation at/above the threshold. The emitted report's axe: { serious, critical } envelope is exactly what conformance-report's a11y.axe-serious-critical criterion consumes — a clean run is what lets a site honestly assert it. | | vuln-gate.mjs | node …/vuln-gate.mjs [projectDir] | $VULN_ROOT (lockfile lives here, default .). Optional $VULN_OMIT_DEV (true→production deps only, default true), $VULN_THRESHOLD (highest tolerated known critical/high, default 0), $VULN_REPORT (write the JSON report). Runs npm audit and fails closed when the known critical/high count exceeds the threshold. The report's vulns: { knownCriticalOrHighVulns } envelope is what conformance-report's security.no-critical-vulns criterion consumes. | | html-validator-gate.mjs | node …/html-validator-gate.mjs [distDir] | $HTML_DIST. Optional $HTML_PAGES (comma list, default: every *.html), $HTML_THRESHOLD (default 0), $HTML_REPORT. Runs vnu (the Nu Html Checker, a self-contained Java jar — needs a JRE) --errors-only over the built pages and fails closed above the threshold. The report's htmlValidator: { errors } envelope is what conformance-report's html.validator-clean criterion consumes. | | baseline-gate.mjs | node …/baseline-gate.mjs [cssGlob] | $BASELINE_CSS (default dist/**/*.css). Optional $BASELINE_TARGET (widely/newly, default widely), $BASELINE_REPORT. Maps the shipped CSS to web-features Baseline data (via stylelint-plugin-use-baseline — headless, no browser) and fails closed when the site-wide status is below target. A feature behind an @supports query is a tested fallback and doesn't count against it. The report's baseline: { status, fallbackTested } envelope is what conformance-report's compatibility.baseline criterion consumes. | | palette-gate.mjs | node …/palette-gate.mjs <tokens.(json\|css)> <pairings.json> | Two inputs the consumer supplies: a token map (a DTCG tokens.json — primitive→semantic aliases resolved — or a tokens.css of --name: #hex custom properties) and a pairings.json declaring the fg/bg pairs that actually co-occur ({ "pairings":[{fg,bg,kind,size?,weight?,name?}], "categorical":[…], "thresholds":{…} }; kind ∈ text|large-text|ui, fg/bg are token names or literal #hex). Runs static colour-palette analysis — zero-dep, every primitive computed by hand: (1) CVD-safe contrast — simulates each colour under deuteranopia/protanopia/tritanopia (Machado-2009 matrices), recomputes the WCAG ratio per pair under each, and flags any pair dropping below AA, plus categorical collapse (CIEDE2000 ΔE below $PALETTE_COLLAPSE_DELTAE, default 10) post-transform; (2) APCA — implements APCA-W3 ~0.1.9, reports Lc per text pair against a font-size/weight-aware (or baseline $PALETTE_MIN_LC_TEXT 60 / $PALETTE_MIN_LC_LARGE 45) minimum, alongside the WCAG-2 ratio (complement, not replacement); (3) non-text contrast — kind:'ui' pairs require ≥3:1 (WCAG 2.2 SC 1.4.11). Thresholds are config-driven (pairings.json thresholds ⊕ $PALETTE_MIN_RATIO_{TEXT,LARGE,UI}) and it fails closed on any failure. $PALETTE_REPORT writes the per-pair JSON (WCAG ratio · APCA Lc · per-CVD ratios · pass/fail per check). The report's palette: { cvdSafe, apcaBaseline, nonTextContrast } envelope is what a future palette.* criterion consumes. | | jargon-gate.mjs | node …/jargon-gate.mjs [distDir] [--strict] | $JARGON_DIST. Optional $JARGON_ALLOWLIST (comma list of accepted terms), $JARGON_MIN_LENGTH (default 3), $JARGON_THRESHOLD (default 0, for --strict), $JARGON_REPORT. Flags undefined jargon in the prose: words not in a 275k-word English dictionary (compounds/possessives atomized first) that the page does not define via <abbr title>, <dfn>, or a <dl> glossary — for W3C COGA / WCAG 3.1.3 Unusual Words and for AI readers. WARN-only by default; --strict fails closed. Report carries a plainLanguage: { undefinedJargon, glossaryPresent } envelope (for a future cognitive.plain-language criterion). | | typography-gate.mjs | node …/typography-gate.mjs <type-tokens.(json\|css)> [config.json] | Token Accessibility suite. Type tokens (DTCG $type:"typography" recipes or .bs-text-* CSS) + a config.json declaring which styles are body ({ "body":["body"], "thresholds":{…} }). Static checks, each mapped to a SC: body line-height ≥ 1.5 (1.4.12); text-spacing achievability — spacing/line-height in overridable relative units, never px-pinned (1.4.12); min font-size — body ≥ ~16px (warn) / ≥ ~12px hard floor (error) + modular-scale sanity (1.4.4); weight×size legibility — thin weight (≤200) at small size → error, plus a requiredApcaLc cross-link to the palette gate (1.4.3/1.4.8). Fails closed on any error; $TYPO_REPORT writes the JSON. | | target-size-gate.mjs | node …/target-size-gate.mjs <config.json> | Token Accessibility suite. A config.json where the consumer declares which tokens are interactive targets ({ "targets":[{name,width,height\|size,exception?,reason?}], "tokens":{…}, "thresholds":{minPx,aaaPx} }). Enforces target ≥ 24×24px (2.5.8 AA → error below) and reports ≥ 44×44px (2.5.5 AAA) status; honours the 2.5.8 inline/essential/user-agent/spacing exceptions with an audit reason. No target tokens → coverage:"none" (vacuous pass + gap note). $TARGET_REPORT writes the JSON. | | opacity-contrast-gate.mjs | node …/opacity-contrast-gate.mjs <tokens.(json\|css)> <usages.json> | Token Accessibility suite — the cross-cutting guard. Token map + a usages.json declaring "opacity applied to a foreground" usages ({ "usages":[{fg,bg,opacity,kind,name?}], "opacityTokens":{…}, "thresholds":{…} }; opacity is 0..1 or a {token} ref). Composites fg over bg (Porter-Duff source-over) at the stated alpha and requires the effective WCAG contrast ≥ floor (4.5 text / 3 large/ui — 1.4.3/1.4.11), reporting both nominal and effective ratio so the drop is visible. Translucent-over-unknown-backdrop usages are flagged for review, not passed. Catches the bounded.tools opacity regression class. $OPACITY_REPORT writes the JSON. | | likeness-gate.mjs | node …/likeness-gate.mjs <tokens.(json\|css)> [config.json] | Token Accessibility suite. Two CIEDE2000 checks over the colour tokens: near-duplicate tokens (ΔE < ~2 ⇒ perceptually identical ⇒ consolidate candidate — warning, escalatable) and confusable categoricals (consumer-declared distinct sets that collapse under normal vision or deuteranopia/protanopia/tritanopia — error; supports 1.4.1). Config: { "categorical":[{name,members}], "ignore":[…], "thresholds":{dupDeltaE,collapseDeltaE,dupSeverity} }. $LIKENESS_REPORT writes the JSON. | | pairing-extractor.mjs | node …/pairing-extractor.mjs <tokens.(json\|css)> <style1.css> [style2.css …] | Token Accessibility suite — coverage engine. Derives the real fg×bg pairings from actual stylesheet usage (resolves var(--token)/literal colours; pairs by same-rule co-occurrence → ancestor-selector containment → root surface, tagged rule/surface/root confidence), unions any declared $PAIRING_DECLARED pairings in, scores every pair through the palette check, and emits a pairing matrix (WCAG · APCA Lc · per-CVD ratios) to $PAIRING_MATRIX (Markdown) / $PAIRING_REPORT (JSON). No DOM ⇒ a reviewed superset (over-generates safely); report-only unless $PAIRING_GATE=1. Removes the hand-maintained pairings list that let the opacity bug slip. | | token-a11y.mjs | node …/token-a11y.mjs <token-a11y.json> | Token Accessibility suite — unified runner (ck-token-a11y). One token-a11y.json drives every member (palette · pairing · typography · targetSize · opacity · likeness) over one token map and fails closed if any fails. See TOKEN-A11Y.md for the standard. $TOKEN_A11Y_REPORT writes the aggregate JSON. | | readability-gate.mjs | node …/readability-gate.mjs <corpus.json> [--strict] | The corpus is an input the site assembles from its copy: a JSON array of {id,text} or an {id:text} map. Optional $READABILITY_THRESHOLDS, $READABILITY_MIN_WORDS, $READABILITY_KNOWN_ACRONYMS. WARN-only unless --strict. | | ai-readability-gate.mjs | node …/ai-readability-gate.mjs [distDir] | Re-proves lone's semantic.ai-readability at build time: emits {llmsTxtPresent, linksResolve, markdownSiblings} — checks llms.txt exists, its internal links resolve (and none hit $AIR_PRIVATE paths), and every content page has a Markdown sibling ($AIR_SIBLING_SUFFIX, default .md; $AIR_SIBLING_IGNORE defaults to 404). Fail-closed ($AIR_STRICT=0 to report only); $AIR_REPORT writes the evidence JSON. Static only — the Accept: text/markdown content-negotiation half is served-edge behaviour, probe it with ck-http-probe. | | css-purity-gate.mjs | node …/css-purity-gate.mjs <a.css> [b.css …] | "No inline values — always tokens." A declaration-aware static scanner (no browser) that fails closed on a raw dimension (320px, 28px, 999px…) in any layout property (width/margin/padding/gap/border-radius/grid-template-*/font-size/…). Force every spacing/sizing/layout value through a coherent var(--bs-*) scale so the composition is predictable + verifiable instead of an overflow/overlap at render time — the static, shift-left counterpart to a runtime layout gate. $PURITY_PREFIX (default --bs-); $PURITY_DIMENSIONS=0 off; $PURITY_COLORS=1 also forbids literal colours (hex/rgb/named); $PURITY_DIM_PROPS overrides the property set; $PURITY_ALLOW / $PURITY_HAIRLINE=1 permit specific raw values. 0/%/fr/ch/auto/var()/calc(of tokens) always pass. | | commonmark-runner.mjs | node …/commonmark-runner.mjs <renderer.mjs> [fixtures.json] | The site's markdown renderer module (export renderMarkdown, or set $COMMONMARK_RENDER_EXPORT). Default fixtures pin a safe CommonMark subset + 4 hostile-HTML escapes; a site with a different renderer supplies its own fixtures.json. | | semantic/gate.ts | deno run --allow-read --allow-net …/gate.ts | Built HTML in $SEMANTIC_DIR (default dist/blog); $SEMANTIC_SELECTOR (subject node, default article). Imports jsr:@bounded-systems/lone; any error-severity finding fails CI. | | conformance-report.mjs | import { buildConformanceReport, renderConformanceReport } from "…/gates/conformance-report.mjs" | The site's evidence — loneFindings (the semantic gate's DOM findings, or null when no DOM was blessed → those criteria report not-assessed) + an external-evidence envelope whose fields it gathers from its own gates (jsonLdShacl, sbom, contentDigests, slsaProvenance, …). renderConformanceReport(report, { evidenceHref }) → a class-based HTML fragment; the consumer wraps it in its template and supplies per-criterion evidence URLs. Zero-dep; the conformance MODEL is a Node port of jsr:@bounded-systems/[email protected]'s conformance() in gates/conformance/. |

The conformance projection makes overclaim impossible by construction: the strong compact claim (COMPACT_CLAIM) is emitted only when every tier-1 required criterion has passing evidence; unsupplied criteria (manual WCAG audit, OWASP ASVS, field Core Web Vitals, Baseline) are not-assessed, never met — so automation can never print "WCAG 2.2 AA" or "ASVS conformant" on its own. tier-2/tier-3/cognitive criteria are reported + summarised per area but never widen the headline claim.

generators/

| Tool | Invoke | Consumer supplies | |---|---|---| | gen-cid.mjs | DIST=dist node …/gen-cid.mjs | $DIST. Walks the site.sha256 file set (or dist), computes the IPFS UnixFS dir CIDv1 with no daemon, records it into $DIST/provenance.json. | | gen-identity.mjs | IDENTITY_DOMAIN=… IDENTITY_REPO=owner/repo node …/gen-identity.mjs | $IDENTITY_DOMAIN, $IDENTITY_REPO (cert-identity regexp), $IDENTITY_SUBJECT (the credentialSubject JSON, default $DIST/resume.json), optional $IDENTITY_SUBJECT_SCHEMA, $IDENTITY_VC_NAME/DESCRIPTION, $IDENTITY_VALID_FROM_PATH. Emits did.json + a W3C VC 2.0. | | gen-snapshots.mjs | node …/gen-snapshots.mjs [distDir] | $SNAPSHOT_DIST (default dist). Optional $SNAPSHOT_PAGES, $SNAPSHOT_BASE_URL (recorded as source in the front-matter), $SNAPSHOT_SUFFIX (default .reader). For every built page, runs @mozilla/readability (the Firefox/Safari Reader engine, via linkedom — headless, no browser) and writes a clean reader <page>.reader.html + an analysis-friendly <page>.reader.md (YAML front-matter + Markdown via turndown). The Markdown is the durable, diffable twin of the page — far easier to run NLP/LLM analysis over than scraping live HTML — and doubles as the AI-readable Markdown sibling. (The printed/PDF view needs a print-CSS renderer and is a separate generator.) | | gen-print-snapshots.mjs | node …/gen-print-snapshots.mjs [distDir] | $PRINT_DIST (default dist). Optional $PRINT_PAGES, $PRINT_RENDERER (default tezcatl, or a "cmd {url} {out}" template), $PRINT_WAIT (default 600), $PRINT_SUFFIX (default .print). The print/PDF twin of gen-snapshots: serves dist over an ephemeral origin (so assets resolve) and renders each page's @media print view to <page>.print.pdf via tezcatl (macOS-native WebKit — no Chromium). A LOCAL / macOS-deploy artifact: on a host without the renderer (e.g. a Linux CI runner) it SKIPS with a note. | | openapi.mjs | import { sortKeys, writeApiFile, embedSchema, jsonResponse, validateOpenapi } | The generic core of a static-API generator. The per-endpoint projection of a site's contracts (profile/posts/corpus/VC, etc.) stays in the site's build; this module provides deterministic JSON output, schema embedding, and OpenAPI 3.1/3.2 well-formedness validation. Pair with lib/schema-validate.mjs to self-check emitted docs. |

emitters/

import { reprDigest, securityTxt, securityTxtExpires, webManifest, markdownSiblingHeaders } from "…/emitters/index.mjs" — pure helpers a site's own build.mjs calls to emit standards-compliant artifacts (RFC 9530 Repr-Digest, RFC 9116 security.txt, the W3C web app manifest, the _headers Content-Type rules for .md siblings). All values injected; the page content stays in the site.

@bounded-systems/verify (vendored here; published elsewhere)

The in-process Sigstore verifier (integrity/verify/verify.mjs) is vendored in this kit so sites can pull it into a hermetic build. It is no longer published from here: the canonical home of the @bounded-systems/verify JSR package is now its own repo, bounded-systems/verify. That repo owns the package manifest (deno.json) and the keyless-OIDC release workflow; cut releases there. The copy here is kept byte-for-byte in sync with the published source.

Consumers run it straight from JSR:

deno run -A jsr:@bounded-systems/verify https://your-site

Releasing

Versioning is owned by @bounded-systems/mint — intent files in, signed release out — replacing the old hand-edited package.json version + publish-branch fast-forward.

1. Per PR — drop one intent in .release/:

   ---
   bump: minor   # patch | minor | major
   ---
   short summary of the change (becomes the changelog line)

   The version CI job (mint plan, pinned to a mint SHA) validates every intent and previews the next version on each PR; it fails closed on a malformed one.

2. Cut a release — on main:

   mint version   # bump package.json (+ lockfile), prepend CHANGELOG.md, consume intents
   mint release   # cut the signed v<version> tag (CI keyless-signs the provenance)

The v<version> tag then drives two independent jobs:

• publish.yml — publishes @bounded-systems/conformance-kit to npm via OIDC trusted publishing (unchanged; mint owns version + tag, not the registry push).
• release.yml — calls mint's reusable release-provenance.yml: emits the deterministic in-toto release Statement (tag → version plan → commit), keyless-signs it (cosign/OIDC), and attaches it to the GitHub release. Verify with cosign verify-blob — the same bundle @bounded-systems/verify consumes.

Migration: the publish branch / manual package.json bump is retired. The tag mint release cuts is now the single release trigger.

Test

npm install && npm test    # cases against fixtures/, in isolation

The suite verifies the generic logic end-to-end: gen-sbom against a sample lockfile; shacl-runner against sample shapes+HTML → conforms: true; structure-audit / seo / readability / commonmark against sample inputs; gen-sitemanifest + gen-cid + verify-site round-trip on a sample build; gen-identity; the emitter/openapi/schema helpers; the conformance projection; and the axe-gate (its classification/threshold/report logic deterministically, plus a real end-to-end pass on the known-bad + known-good fixtures/axe/ snippets when a browser engine — tezcatl or Playwright/Chromium — is on PATH; skipped, like the cosign step, when none is). (The Deno semantic runner is exercised by the consuming site, as it needs Deno + JSR.)

Provenance / determinism

The gates are pure functions of the built output; the generators are deterministic functions of their inputs (the SBOM creation date is derived from flake.lock, never a wall clock; the CID re-derives from the served bytes with any IPFS implementation). Site-specific artifacts — SHACL shapes, the prose corpus, the markdown renderer, thresholds, copy, and build.mjs itself — are inputs, never part of the kit.