@nebulacomponents/citable 0.1.0

package/dist/universal/.agents/skills/citable/commands/optimize-page.md

@@ -0,0 +1,53 @@
+---
+command: /citable optimize-page
+purpose: Improve one page's retrieval eligibility, answer clarity, entity relationships, and internal links while preserving its conversion role and claim boundaries.
+inputs: [page reference (URL path or source file), site target]
+optional_inputs: [specific findings to address, ref-date]
+preconditions:
+  - .citable/ initialized; page present in page registry (else run inspect + register first)
+  - repository build and test commands known (project.yaml)
+  - working tree clean or user-acknowledged
+registries_read: [pages, queries, prompts, entities, claims, evidence]
+registries_written: [pages (metadata only), claims (surface updates)]
+detectors_run: [all page-scoped detectors before and after]
+artifacts_created: [before/after inspect output, diff, post-change audit run]
+refusal_conditions:
+  - page carries regulated claims and no factual_reviewer is assigned
+  - requested change matches a ⛔ anti-pattern
+  - required facts (pricing, ICP, limitations) are unknown → blocked with required_input
+failure_behaviour: build or render failure → report command, exit code, error output; no success status
+---
+
+# Workflow
+
+1. `citable inspect <page> --target <built-output>` — capture the BEFORE state
+   (intent, entities, claims, findings, conversion action, ambiguity list).
+2. Read the page registry entry: intended query and prompt coverage, conversion
+   role, lifecycle class. If ambiguity list is non-empty, resolve registry
+   gaps before editing content.
+3. Order of work (never reversed):
+   a. **Technical defects first** — status, canonical, robots, sitemap, links
+      (TECH/LINK findings).
+   b. **Semantic ambiguity** — apply semantic-clarity and intent-alignment
+      rubrics; fix subject/category/terminology drift.
+   c. **Answer clarity** — direct answer block in first 50–100 words; atomic
+      claims; evidence adjacent; scope and exclusions in-passage
+      (answer-extractability rubric; ANS findings).
+   d. **Claim boundaries** — every edited claim keeps its registered scope and
+      exclusions verbatim or narrower; never broaden while "polishing".
+      Attach evidence references beside claims.
+   e. **Entity relationships & internal links** — upward hub link, lateral
+      related links, commercial next step; descriptive anchors.
+   f. **Schema** — only where justified by registry data (`citable schema`
+      proposals); never hand-add assertions.
+4. Style constraints: no keyword stuffing (PAGE-009 must not newly fire), no
+   generic AI prose (ANS-001 must not newly fire), match the site's voice.
+5. Preserve the conversion role — apply the conversion-alignment rubric; the
+   registered conversion_action must remain present and reachable.
+6. Build the project with the recorded build command. If it fails → failure
+   behaviour, stop.
+7. Re-audit: `citable audit --target <built-output>` and
+   `citable compare-snapshots` — the page's prior findings should resolve and
+   no new critical/high findings may appear.
+8. Report: findings resolved, findings remaining (with reasons), unresolved
+   conditions, registry diffs, and the run IDs of before/after evidence packages.

package/dist/universal/.agents/skills/citable/commands/page-work.md

@@ -0,0 +1,48 @@
+---
+command: /citable create-page, answer-block, architect, interlink, consolidate, metadata
+purpose: Page-construction and architecture workflows sharing optimize-page's guardrails.
+---
+
+# create-page
+1. Preconditions: a registered query/prompt with no existing target (else it's
+   ARCH-004 territory — consolidate instead); an assigned content owner; the
+   facts the page needs (from registries and operator), never invented.
+2. Compose per the generative-ready pattern: canonical answer (50–100 words) →
+   definition + exclusions → why it matters (operational, not marketing) →
+   mechanism → evidence → comparison → boundary conditions → implementation →
+   revision metadata → related entities.
+3. Claims on the page must pre-exist in the claim registry (or be added as
+   candidates and pass substantiate + evidence-strength review before launch).
+4. Register the page; run post-modification validation (build → audit →
+   compare-snapshots).
+
+# answer-block
+Insert/repair a direct answer block on an existing page. Constraints: the block
+answers the page's principal question, is self-contained (answer-extractability
+rubric), includes in-passage scope, and does not displace the conversion action
+(conversion-alignment rubric). Verify ANS-001/002/008 do not fire afterwards.
+
+# architect
+Propose site/topic architecture: one canonical URL per material intent,
+hub-and-spoke relationships, shallow depth for priority pages, separation of
+informational and transactional intents. Output is a written proposal +
+page-registry changes in `draft`; it never mass-moves URLs without a migration
+plan (redirect map, sitemap regeneration, canonical updates) recorded first.
+
+# interlink
+Add internal links only where a semantic relationship exists in the registries
+(same topic entities, claim→evidence, spoke→hub, implementation→commercial
+next step). Descriptive anchors; never bulk term-matched insertion (LINK-004).
+Verify LINK-001/002/003 and ARCH-001/002/006 after.
+
+# consolidate
+For ARCH-004/005 findings: choose the survivor URL (stronger inbound links +
+better intent fit), merge unique content preserving claim boundaries, 301 the
+losers, update sitemap + canonicals + registry (losers → status retired with
+successor recorded), rerun audit. Never delete content that is the only
+publication surface of an active claim without re-homing the claim.
+
+# metadata
+Title/description work: titles identify and differentiate (PAGE-001/002),
+descriptions summarize accurately and never promise absent content. Both are
+derived from page substance — metadata never asserts what the page doesn't say.

package/dist/universal/.agents/skills/citable/commands/validate-render.md

@@ -0,0 +1,24 @@
+---
+command: /citable validate-render
+purpose: Confirm answer-bearing content exists in served HTML, not only after client-side rendering.
+preconditions: [build command known, built output or deployed URL available]
+failure_behaviour: build/render failure → preserve command, exit code, stderr; status incomplete, never success
+---
+
+# Workflow
+
+1. Build with the project's recorded build command. Non-zero exit → stop,
+   report failure evidence.
+2. Run `citable audit technical --target <built-output>`; TECH-011
+   (render-dependent primary content) is the primary detector here.
+3. For each index-target page, compare extracted text of the initial HTML with
+   the intended primary content (title, H1, canonical answer, claims). Content
+   present only in JS payloads, accordions requiring interaction, images,
+   PDFs, or canvas → finding with the specific missing passages.
+4. If a headless browser is available in the environment, render and diff
+   initial vs rendered text to quantify the gap; if not, record
+   `incomplete_checks: rendered comparison unavailable in this environment` —
+   do not extrapolate.
+5. Verdict per page: `server_rendered | hybrid_acceptable | render_dependent |
+   not_determined`. Never report site-wide success while any check is
+   not_determined without listing it.

package/dist/universal/.agents/skills/citable/manifest.json

@@ -0,0 +1,54 @@
+{
+  "name": "citable",
+  "version": "0.1.0",
+  "managedBy": "citable-cli",
+  "provider": "codex",
+  "providerName": "OpenAI Codex",
+  "scope": "bundle",
+  "sourcePackage": "@nebulacomponents/citable@0.1.0",
+  "variantPath": ".agents/skills/citable",
+  "files": {
+    "commands/ingest.md": "sha256:cf35659aecd9d0e160909003efdbe056864fc5d824c71b4ef0ee801ce07ff598",
+    "commands/measurement.md": "sha256:3550b1e747c5a38cdb712016b9bc773d93596bc4d1b49a309f9cddd816c1b662",
+    "commands/optimize-page.md": "sha256:35866f3d7cdeef0376ea7a7d7036085a9b9cf95bbb300ecf4bd77610ebd74d85",
+    "commands/page-work.md": "sha256:965a66e59bdf7b0da482d26fade53b3d265885ece5d463fe83b0789ebefaef3e",
+    "commands/README.md": "sha256:13774e705356b9eafb42aa522cc10337d86dfb973d533976f041a21819fb6cb4",
+    "commands/validate-render.md": "sha256:a8e1d04d40b84d57c35039250d98c6a7302cd44a148a8c3b9d9cfb92678c4347",
+    "policies/crawler-policy-template.md": "sha256:15c88a4eeb21f53f71e5250dd3eb2d2791a814b304eaefce8bb30e2bdf2e2e59",
+    "references/anti-patterns.md": "sha256:de63f8e646af0fbf479b45eb1d050ae62a8637acc320d21b87e4a083fd4c47e7",
+    "rubrics/answer-extractability.md": "sha256:c8e741143c34e53afdb3dc7b6976ce7b4485dce13b12d14c193fccb17bacf1da",
+    "rubrics/claim-boundedness.md": "sha256:0bab1b85248c252b966175912ebfae13bf18aa7e41573f7466831ac78b5d3e27",
+    "rubrics/comparison-fairness.md": "sha256:7afc86903aa4a695060be7b6bca42ba505c106ea7cfa153344189afc5b78a5ba",
+    "rubrics/conversion-alignment.md": "sha256:85882736cd15320e5af5a872fd4263b5c40ec202c39023e8eb416848f942905e",
+    "rubrics/entity-clarity.md": "sha256:5b8003019fe81cf055757bee059eeb56f92753a683a0f94f7641143b88a29e74",
+    "rubrics/evidence-strength.md": "sha256:c9b9ea2454ba9e71e42909060821b24703bf6afe39b9a324e01fe234138af71c",
+    "rubrics/information-gain.md": "sha256:97bb02015efeca9b83fc503c7f2e4dadad1358cb47453dfa36d76bc0cac71ceb",
+    "rubrics/intent-alignment.md": "sha256:16ca19bfeb4e58ea4910a2e212a935bcfeeaaa06a66b34834f82f2e79439cc87",
+    "rubrics/narrative-accuracy.md": "sha256:0e0671175329c4da78855514ac1512d16a22217ca8694ca4a283d4d604b89028",
+    "rubrics/README.md": "sha256:d80b94d44b5a4e0fe9b6e1708a2f6d6a156a1aec5b4d4a632669803dd8b67d1e",
+    "rubrics/recommendation-eligibility.md": "sha256:38d2055fc5f6ab5e40fd1b48de4e0518fc4382da0b7062ad09c5afebcc673fc3",
+    "rubrics/semantic-clarity.md": "sha256:66d9f4843a4a620ea8c00625c5e17c02baeaa373f2b445ce13de6f81d78ec7ba",
+    "rubrics/source-authority.md": "sha256:8dbfe46bdcd0b5351f038be503601268c7d0abb9bed83be9ca58bf54dec5a73b",
+    "schemas/claim.schema.json": "sha256:70c521e51a24031a1e1f73eca04c2a2650e760f3fdc82e9553eb26c0b5c428e9",
+    "schemas/common.defs.json": "sha256:1ae1ae4a9b0babadbcba3288dd3d1adb3e152f87f1e51c08a397a0f29878e95f",
+    "schemas/competitor.schema.json": "sha256:7d34c1f131eba23d94404921c4defbdabb402f83bc5a748a1de40fbe89d65cca",
+    "schemas/config.schema.json": "sha256:a87f3bb0087de09035b9f2dafea66e32e4fcc2e3e329abcd0fd8a5ec33318f2a",
+    "schemas/crawler.schema.json": "sha256:0541de509d0b5de065e1dd3ff52212ba389eaf7f5a9c2309b5e5beb232b4b356",
+    "schemas/detector-quality.schema.json": "sha256:5e9e1ac80b1198a0ff8fa17b9f7a543d3ccf3ed3d60196152f6ff095cfdebccd",
+    "schemas/entity.schema.json": "sha256:12f9222122255c86059185bc66e528e2828b430f7abc5683e6759774eb555553",
+    "schemas/evidence.schema.json": "sha256:a78d39cd1d66c89198b0ecc9af48c8d3d8659c7b5cbb9f96c43d18fcf2b4aeb7",
+    "schemas/experiment.schema.json": "sha256:704ea5c14cd4409c3a1806b8b1c124bcd9f2fe9e313dfa93662889bed0cb6441",
+    "schemas/finding.schema.json": "sha256:2366bbbd0f6359ecc8d262979b159811ad10d2eaacc4816d5075bc6f7e9efc13",
+    "schemas/page.schema.json": "sha256:195d2f91d67550627c7047fd8e98d08dc386bc5858a02239d6c23b99193eaac2",
+    "schemas/project.schema.json": "sha256:ea48fdb6480f7047ba4594bd5908bc1f79c1ad6b9a2c204df900959b5f203dc6",
+    "schemas/prompt-result.schema.json": "sha256:b55be47d7d58d7b1dd8c3b1e849e24e55216212c592cde7eb836953379b85737",
+    "schemas/prompt.schema.json": "sha256:69782525f9de3a9f92649f2bc2fb4a0b2ed4288d6028663b18e39918c9f7fca1",
+    "schemas/query.schema.json": "sha256:59829f78461f115d73486e6350dcea5daf5a8e77576013b7a6d25913f240c55d",
+    "schemas/run.schema.json": "sha256:46c3178620f3a358d936d535aeadb83454b9b35b551e53e6615cc9862063dacb",
+    "scripts/README.md": "sha256:1c3e952950ef353849621b07c5f615bf7cd20cc5654da4635492a7bd6384e747",
+    "SKILL.md": "sha256:637bb755bca6a2225c5b69edab0b8be7dff294e784d7ce25cec23546b847324c",
+    "templates/finding-review.md": "sha256:b270a83d3f015150aac80249ae9314979c0befea39ff92ec455f9a5aaa73404c",
+    "VERSION": "sha256:e9dd8507f4bf0c6f42458e41aea833ad0bd3f6127272335eee9bf4d58541ed67"
+  },
+  "treeHash": "sha256:7b3f30bfd6e7c79f549a4ca9a2269a32afa899a995437ed9b1b8d7e86d562bf2"
+}

package/dist/universal/.agents/skills/citable/policies/crawler-policy-template.md

@@ -0,0 +1,47 @@
+# Crawler policy template
+
+Access is decided **per crawler and per purpose**. Search-crawler allowance is
+never permission for model training (premise 3.3). Every decision needs an
+owner, business rationale, legal rationale, and review date in
+`.citable/crawlers.yaml`; robots.txt is generated to match (CRAWL-001 verifies
+the two agree).
+
+| Crawler | Vendor | Purpose | Decision defaults to |
+| --- | --- | --- | --- |
+| Googlebot | Google | search_indexing (also gates AI features eligibility) | undecided → usually allow |
+| Bingbot | Microsoft | search_indexing (also Copilot surfaces) | undecided |
+| OAI-SearchBot | OpenAI | ai_search_discovery (ChatGPT search citation) | undecided |
+| GPTBot | OpenAI | model_training | undecided — separate decision, legal owner |
+| ChatGPT-User | OpenAI | user_initiated_retrieval | undecided — separate decision |
+| PerplexityBot | Perplexity | ai_search_discovery | undecided |
+| Perplexity-User | Perplexity | user_initiated_retrieval | undecided |
+| ClaudeBot | Anthropic | model_training (verify current vendor docs) | undecided |
+
+Rules:
+
+1. `undecided` is a temporary state; CRAWL-006 flags it. A decision requires
+   decision_owner + rationale.
+2. Allowing a vendor's search crawler without recording a training-crawler
+   decision for the same vendor is flagged (CRAWL-002).
+3. Vendor documentation URLs and IP-validation methods are recorded per entry;
+   user-agent strings are spoofable and must not be the sole trust basis.
+4. Robots semantics: robots.txt controls crawling, not indexing. Use `noindex`
+   when removal from search results is the objective (TECH-003 remediation).
+5. Blocked-crawler decisions must state whether the block covers licensed or
+   partner access paths too, or only public crawling.
+
+Example generated robots.txt for: search allowed, OpenAI search allowed,
+training blocked:
+
+```text
+User-agent: GPTBot
+Disallow: /
+
+User-agent: ClaudeBot
+Disallow: /
+
+User-agent: *
+Allow: /
+
+Sitemap: https://example.com/sitemap.xml
+```

package/dist/universal/.agents/skills/citable/references/anti-patterns.md

@@ -0,0 +1,61 @@
+# Citable anti-pattern library
+
+Canonical reference used by all commands. When a requested change matches an
+entry, name it, state the risk, and offer the defensible alternative. Entries
+marked ⛔ must be refused outright (premise 3.6), not merely discouraged.
+
+## Content anti-patterns
+
+| Anti-pattern | Risk | Defensible alternative |
+| --- | --- | --- |
+| Generic AI introductions ("In today's rapidly evolving landscape…") | Suppresses answer extraction; signals commodity content (ANS-001) | Direct answer in the first 50–100 words |
+| Excessive preamble before the answer | Passage extraction pulls filler instead of the claim | Lead with the canonical answer block |
+| Keyword stuffing | Spam-policy exposure; unreadable prose (PAGE-009) | Semantic coverage driven by explanation needs |
+| Synthetic information gain (rewriting top results) | No citation reason; no durable differentiation | Original data, experience, taxonomy, benchmarks |
+| Repeated conclusions / decorative FAQs | Dilutes extractable claims; FAQ schema risk (SCHEMA-007) | One canonical answer location per question |
+| ⛔ False precision / unsupported statistics | Legal + reputational; CLAIM-004/ANS-005 | Publish numbers only with methodology and period |
+| Circular definitions | Engine cannot anchor the concept (ANS-003) | Genus–differentia definition plus exclusions |
+| ⛔ Fake comparisons / unbounded superlatives | Deceptive-claim exposure (CLAIM-007) | Explicit comparison set, criteria, and evidence |
+| ⛔ Invented expertise / anonymous substantive content | E-E-A-T failure; regulatory exposure | Real named authors with verifiable profiles |
+| Hidden commercial bias | Undisclosed endorsement liability | Disclose commercial relationships |
+| Duplicated programmatic prose | Index bloat, quality dilution | Programmatic pages only with unique data + owner |
+
+## Technical anti-patterns
+
+| Anti-pattern | Risk | Defensible alternative |
+| --- | --- | --- |
+| "Index everything" | Crawl waste, quality dilution | Deliberate indexing_intent per page |
+| Schema saturation | Invisible-assertion violations | Schema only for visible, registry-backed facts |
+| Sitemap dumping (redirects, 404s, non-canonicals) | TECH-008/009/010 | Generate sitemaps from the canonical URL set |
+| Canonical misuse (variants → irrelevant page, pagination → page 1) | Signal consolidation failures | One canonical per distinct content object |
+| JavaScript-only primary content | Render-stage failure (TECH-011) | SSR/SSG for answer-bearing content |
+| ⛔ Crawler cloaking / bot-specific deceptive rendering | Spam-policy removal | Serve equivalent content to bots and users |
+| Unbounded faceted URLs | Index bloat, crawl waste | Controlled parameters + canonical rules |
+| Redirect-dependent architecture | Chain latency, signal loss (TECH-013) | Direct links to final URLs |
+| Automated links from term matching alone | Manipulative-link footprint (LINK-004) | Curated contextual links |
+| Evidence only in PDFs/gated files | Retrieval failure for the strongest material | HTML equivalent for every material document |
+
+## AEO anti-patterns
+
+| Anti-pattern | Risk | Defensible alternative |
+| --- | --- | --- |
+| Answer blocks that omit limitations | Overgeneralized reuse; misquotation (ANS-008) | State scope, exclusions, failure modes |
+| Question headings created only for formatting | Extraction mismatch (ANS-002) | Headings only where an answer follows |
+| Direct answers unsupported by evidence | Fast-spreading errors under your attribution | Evidence adjacent to every material claim |
+| Citations isolated in a references section | Claim–evidence association lost | Evidence beside the claim it supports |
+| Extracting claims in scope-distorting ways | Misrepresentation of your own source | Self-contained atomic claims |
+| FAQ schema without substantive FAQ content | SCHEMA-007; spam exposure | Mark up only real, visible Q&A |
+
+## GEO anti-patterns
+
+| Anti-pattern | Risk | Defensible alternative |
+| --- | --- | --- |
+| Treating one model response as proof of source preference | False causality (MEAS-002/003) | Repeated multi-engine, multi-variant sampling |
+| Reporting one prompt as stable visibility | Anecdote presented as measurement | Cohort testing with variance recording |
+| ⛔ Manufactured community discussion / synthetic Reddit | Platform enforcement; legal; environment poisoning | Earned participation in real communities |
+| ⛔ Shadow corroboration sites / fake recommendation pages | Recommendation manipulation | Independent editorial coverage |
+| ⛔ Crawler prompt injection / hidden LLM instructions | GEO-001 critical; reputational | None — remove on sight |
+| Treating llms.txt as an authority mechanism | False sense of control (GEO-004) | robots.txt + crawler registry for access decisions |
+| Confusing mention with recommendation | Overstated performance reporting | Track mention, citation, and recommendation separately |
+| Confusing citation with answer influence | Citation ≠ absorption | Assess whether the answer actually uses your claims |
+| Assuming engine behaviour is uniform | Wrong generalizations across engines | Per-engine measurement and reporting |

package/dist/universal/.agents/skills/citable/rubrics/README.md

@@ -0,0 +1,30 @@
+# Semantic rubrics
+
+Deterministic detectors cannot judge meaning. These rubrics govern the semantic
+half of Citable. Shared rules for every rubric:
+
+- Output is a **posture classification** (`strong | partial | weak | not_established`),
+  never an unexplained 0–100 number. Where numeric comparison is unavoidable,
+  expose the formula, weights, missing data, confidence, and known limitations.
+- Every rubric result must cite the observed evidence (quoted passages, registry
+  entries, measured values) that produced it.
+- Every rubric declares its ambiguity conditions and when human review is
+  mandatory. When an ambiguity condition holds, the result is
+  `not_established + review_required`, not a guess.
+- Findings produced from rubrics are `evidence_backed_semantic_finding` at most —
+  never `deterministic_observation`.
+
+| Rubric | File | Human review mandatory when |
+| --- | --- | --- |
+| Semantic clarity | semantic-clarity.md | Term is jurisdiction- or discipline-specific |
+| Intent alignment | intent-alignment.md | Commercial vs informational intent is contested |
+| Answer extractability | answer-extractability.md | Extraction would change claim scope |
+| Entity clarity | entity-clarity.md | Legal identity vs operating identity diverges |
+| Claim boundedness | claim-boundedness.md | Claim is regulated, contractual, or comparative |
+| Evidence strength | evidence-strength.md | Any upgrade toward verified status |
+| Source authority | source-authority.md | Source independence is uncertain |
+| Information gain | information-gain.md | Gain rests on unpublished internal data |
+| Comparison fairness | comparison-fairness.md | Named competitors appear |
+| Recommendation eligibility | recommendation-eligibility.md | Pricing/geography/ICP data incomplete |
+| Conversion alignment | conversion-alignment.md | Page has no declared conversion action |
+| Narrative accuracy | narrative-accuracy.md | Contradiction involves legal or security facts |

package/dist/universal/.agents/skills/citable/rubrics/answer-extractability.md

@@ -0,0 +1,44 @@
+# Rubric: answer extractability
+
+Judges whether a page's principal answer survives being lifted out of the page
+without distortion. Complements deterministic ANS-00x detectors.
+
+## Scoring dimensions
+
+1. **Position** — canonical answer within the first 50–100 words after the H1.
+2. **Self-containment** — passage names its subject explicitly; no deictic
+   references ("as shown above"), no pronoun whose referent is outside the passage.
+3. **Atomicity** — one claim per key sentence; each independently understandable.
+4. **Scope survival** — if the passage is quoted alone, its conditions and
+   exclusions travel with it (in-sentence, not three paragraphs later).
+5. **Definition quality** — "[Term] is [genus] that [differentia]" plus explicit
+   "It does not include…".
+6. **Structural support** — procedures as ordered steps; comparisons with stated
+   basis and prose equivalent of any table.
+
+## Posture
+
+- `strong`: dimensions 1–4 all hold for the page's principal question.
+- `partial`: answer present but scope/exclusions live outside the extractable span.
+- `weak`: answer buried after preamble, or dependent on visual context.
+- `not_established`: the page has no identifiable principal question (fix the
+  page registry first — this is an intent problem, not a formatting problem).
+
+## Evidence required
+
+Quote the exact passage judged, with its position (paragraph index / selector).
+
+## Example / counterexample
+
+- Strong: "AI execution governance is the set of enforceable controls that
+  determines whether a specific AI-initiated action is authorized at the moment
+  of execution. It does not include model training governance."
+- Weak: "This matters more than ever. As we saw above, there are many angles to
+  consider before diving into what this means for your organization."
+
+## Ambiguity conditions → human review
+
+- Extraction would change the claim's scope (dimension 4 fails in a way that
+  affects a registered claim).
+- The principal question is contested between informational and commercial intent.
+- Content is legally qualified language where tightening wording changes meaning.

package/dist/universal/.agents/skills/citable/rubrics/claim-boundedness.md

@@ -0,0 +1,36 @@
+# Rubric: claim boundedness
+
+Judges whether a claim's wording stays inside its evidence, scope, and
+exclusions — the semantic complement to CLAIM-00x detectors.
+
+## Scoring dimensions
+
+1. **Scope explicitness** — conditions under which the claim holds are stated
+   in the claim or immediately adjacent, not implied.
+2. **Exclusion honesty** — known non-covered cases stated; material limitations
+   not omitted (a limitation the buyer would consider material must appear).
+3. **Type fidelity** — opinion phrased as opinion, aspiration as aspiration,
+   comparison with its comparison set, performance with its conditions.
+4. **Quantifier discipline** — "all/every/always/never/only" only when evidence
+   covers the universal; otherwise scoped quantifiers.
+5. **Temporal honesty** — version/date-bound facts carry their version or date.
+
+## Posture
+
+`strong`, `partial` (scope stated, exclusions thin), `weak` (wording exceeds
+evidence or universal quantifiers uncovered), `not_established` (claim not in
+registry — run `/citable map-claims`).
+
+## Example / counterexample
+
+- Bounded: "In configured enforcement deployments, Gatekeeper validates each
+  action before execution. Advisory-only deployments are excluded."
+- Unbounded: "Gatekeeper stops every unauthorized AI action." (universal
+  quantifier, no scope, no exclusions → `weak`; substantiate outcome at best
+  `verified_narrowed` after rewording).
+
+## Ambiguity → human review
+
+- Regulated, contractual, security, financial, medical, certification, or
+  customer claims (always — mirrors CLAIM-006).
+- Narrowing the wording would change a published commercial commitment.

package/dist/universal/.agents/skills/citable/rubrics/comparison-fairness.md

@@ -0,0 +1,30 @@
+# Rubric: comparison fairness
+
+Judges whether a comparison page is legitimate analysis rather than a stacked
+deck. Complements ANS-007 (basis exists) with a judgment of whether the basis
+is honest.
+
+## Scoring dimensions
+
+1. **Criteria legitimacy** — comparison dimensions are ones a buyer would use,
+   not dimensions selected because only the author wins them.
+2. **Alternative completeness** — legitimate alternatives a buyer would
+   shortlist are included; excluding the strongest competitor is a fairness
+   failure, not an editorial choice.
+3. **Evidence per cell** — comparative assertions about *other* vendors carry
+   sources and retrieval dates (their docs, public pricing), not assumptions.
+4. **Loss disclosure** — the author's product loses at least where it actually
+   loses; a comparison with zero conceded dimensions is presumptively unfair.
+5. **Currency** — competitor capabilities dated; stale claims about others are
+   both unfair and legally risky.
+
+## Posture
+
+`strong` (all five), `partial` (fair criteria, thin sourcing), `weak`
+(criteria stacked or no conceded dimensions), `not_established` (competitor
+data unavailable → mark incomplete; never guess competitor capabilities).
+
+## Ambiguity → human review
+
+- Any named competitor (legal review of comparative claims — always).
+- Competitor capability changed recently and sources conflict.

package/dist/universal/.agents/skills/citable/rubrics/conversion-alignment.md

@@ -0,0 +1,37 @@
+# Rubric: conversion alignment
+
+Judges whether optimization preserved or improved the page's commercial
+function — the guard against "optimized for machines, ruined for buyers"
+(premise 3.7).
+
+## Scoring dimensions
+
+1. **Action preservation** — the registered conversion_action is still present,
+   visible, and functional after changes.
+2. **Stage fit** — the action matches the funnel stage of the mapped
+   queries/prompts.
+3. **Friction honesty** — added answer blocks, definitions, or schema did not
+   push the conversion element below the fold on the page's dominant viewport
+   class (verify, don't assume).
+4. **Trust continuity** — limitations and scope statements added for AEO/GEO
+   are framed accurately but not self-sabotagingly (state the limitation; don't
+   editorialize against your own product).
+5. **Accessibility preserved** — interactive conversion elements keep their
+   labels, focus order, and keyboard operability after edits.
+
+## Posture
+
+`strong`, `partial` (action present, stage fit doubtful), `weak` (action lost,
+demoted, or broken), `not_established` (no conversion_action registered —
+that's an inspect/registry finding first).
+
+## Evidence required
+
+Before/after presence of the conversion element (selector or line reference)
+and the registry conversion_action row.
+
+## Ambiguity → human review
+
+- Page has no declared conversion action but obviously sells something.
+- Optimization requires choosing between answer position and conversion
+  position — commercial owner decides, not the optimizer.

package/dist/universal/.agents/skills/citable/rubrics/entity-clarity.md

@@ -0,0 +1,41 @@
+# Rubric: entity clarity
+
+Judges whether an external system can resolve who/what an entity is, what it
+owns, and what category it belongs to — across owned and observed surfaces.
+
+## Scoring dimensions
+
+1. **Identity completeness** — canonical name, type, canonical URL, and (for
+   organizations) legal vs operating identity explained.
+2. **Relationship graph** — ownership and product-family relationships recorded
+   and mirrored in structured data (@id references, publisher links).
+3. **Cross-surface consistency** — same category and description across website,
+   profiles, directories, documentation (small wording drift is normal;
+   *category contradiction* is the failure).
+4. **Alias governance** — every name variant in the wild is either registered
+   as an alias or scheduled for correction.
+5. **External anchoring** — sameAs/profiles limited to verified authoritative
+   accounts.
+
+## Posture
+
+`strong`, `partial` (identity complete, cross-surface unverified),
+`weak` (category contradiction or ownership ambiguity), `not_established`
+(identity fields missing — mark the entity `incomplete`; never invent legal
+names, founders, addresses, or registration status).
+
+## Evidence required
+
+Registry entity row, schema blocks, and the exact conflicting descriptions with
+their sources and retrieval dates.
+
+## Counterexample (category contradiction)
+
+Website: "AI execution governance platform" / LinkedIn: "ethical AI consulting"
+/ directory: "AI marketing automation" → `weak`, with each conflicting surface
+listed as a contradiction finding.
+
+## Ambiguity → human review
+
+- Legal identity vs operating identity diverge (renames, acquisitions).
+- Entity shared between organizations (joint ventures, resold products).

package/dist/universal/.agents/skills/citable/rubrics/evidence-strength.md

@@ -0,0 +1,53 @@
+# Rubric: evidence strength
+
+Judges whether attached evidence actually supports a claim *as worded*. This is
+the mandatory semantic gate before any claim moves toward `verified`
+(`citable substantiate` handles only the deterministic preconditions).
+
+## Scoring dimensions
+
+1. **Directness** — does the evidence address the claim's exact proposition, or
+   an adjacent one? ("handles 10k rps" is not supported by "architecture is
+   horizontally scalable".)
+2. **Primacy** — primary artifact (test result, specification, certification,
+   registration) vs secondary report of one.
+3. **Methodological completeness** — methodology, test conditions, measurement
+   period, population, and environment recorded and appropriate.
+4. **Currency** — evidence dated within the claim's temporal class (see
+   lifecycle table in the AEO source doc §8).
+5. **Independence** (for comparative/market claims) — at least one source not
+   controlled by the claiming organization.
+6. **Coverage of scope** — evidence covers every deployment/scenario inside the
+   claim's declared scope; anything uncovered must appear in exclusions.
+
+## Posture
+
+- `strong`: all six dimensions satisfied; claim wording does not exceed evidence.
+- `partial`: directness + currency satisfied, but scope coverage or primacy
+  incomplete → outcome is `verified_narrowed` with explicit narrowed scope.
+- `weak`: evidence is adjacent, stale, or secondary-only → `insufficient_evidence`.
+- `not_established`: evidence inaccessible or methodology absent.
+
+## Confidence
+
+`confirmed` only when the evidence artifact itself was read (not its registry
+row). `high` when methodology and directness verified. Never above `medium`
+when working from registry metadata alone.
+
+## Example / counterexample
+
+- Supported: "Gatekeeper validates actions before execution *in enforcement
+  deployments*" ← architecture spec + live validation report covering
+  enforcement mode. Scope stated, exclusions stated.
+- Not supported: "reduces compliance costs by 70%" ← a customer quote saying
+  "significantly reduced our costs". Quote is directional, not quantitative;
+  outcome is `insufficient_evidence` with required_input: baseline, period,
+  population, methodology, source.
+
+## Ambiguity conditions → human review
+
+- Any upgrade toward `verified`/`verified_narrowed` (always).
+- Evidence in a language or jurisdiction the reviewer cannot assess.
+- Contradictory evidence exists of comparable primacy.
+- Claim is legal/regulatory, security, financial, medical, certification, or
+  names a customer — legal/SME review regardless of evidence quality.

package/dist/universal/.agents/skills/citable/rubrics/information-gain.md

@@ -0,0 +1,33 @@
+# Rubric: information gain
+
+Judges whether a page contributes information not available elsewhere — the
+difference between a citable source and a summary of other sources.
+
+## Scoring dimensions
+
+1. **Original artifact presence** — at least one of: original data, benchmark,
+   controlled experiment, incident analysis, reference architecture,
+   proprietary taxonomy (defined, not just named), survey, reproducible
+   methodology, first-party operational observation, reusable tool.
+2. **Non-derivability** — could a competent writer produce this page from the
+   current top results alone? If yes, gain is synthetic.
+3. **Specificity** — concrete versions, numbers-with-methodology, named failure
+   modes; not abstract advice.
+4. **Reproducibility** — methods described well enough to re-run or falsify.
+
+## Posture
+
+`strong` (dimension 1 + 2 hold), `partial` (original framing, no original
+artifact), `weak` (restatement of existing sources), `not_established`
+(gain claimed from internal data that is not published — publish it or drop
+the claim of originality).
+
+## Counterexample
+
+"10 best practices for AI governance" synthesized from existing articles, with
+an LLM paraphrase pass — scores `weak` regardless of length or formatting.
+
+## Ambiguity → human review
+
+- Gain rests on unpublished internal data (publication/licensing decision).
+- Original data involves customers or regulated metrics.

package/dist/universal/.agents/skills/citable/rubrics/intent-alignment.md

@@ -0,0 +1,41 @@
+# Rubric: intent alignment
+
+Judges whether page content actually serves the intent declared in the page
+registry and the queries/prompts mapped to it.
+
+## Scoring dimensions
+
+1. **Declared vs delivered** — a page targeting "vendor evaluation" must contain
+   evaluation material (capabilities, limitations, pricing qualification,
+   comparison), not a re-titled awareness article.
+2. **Intent purity** — informational and transactional intents separated unless
+   the registry deliberately combines them.
+3. **Query/prompt fit** — for each mapped query_id/prompt_id, the page contains
+   the expected answer components recorded in the registry.
+4. **Next-step coherence** — the conversion action matches the funnel stage
+   (awareness pages don't push "buy now"; selection pages don't end at a blog
+   subscription).
+
+## Posture
+
+`strong` (all mapped queries served, next step coherent), `partial` (primary
+query served, variants uncovered), `weak` (declared intent contradicted by
+content), `not_established` (no registry mapping — run `/citable map-queries`
+first).
+
+## Evidence required
+
+The registry mapping rows plus quoted page passages that do (or fail to) serve
+each expected answer component.
+
+## Counterexample
+
+A page registered for "how to implement X" (implementation intent) whose body
+is a category pitch with no steps — flagged by ANS-006 deterministically; this
+rubric additionally judges whether the *substance* matches even when an ordered
+list exists.
+
+## Ambiguity → human review
+
+- Commercial vs informational intent contested for a revenue-relevant page.
+- Two registries map conflicting intents to the same URL (also ARCH-004/005).