lazycodex-ai 5.1.44 → 5.1.46

package/packages/omo-codex/plugin/skills/ulw-research/scripts/fact-pairs.mjs

@@ -0,0 +1,187 @@
+// fact-pairs.mjs - (source, report sentence) pairs for the meaning verifier.
+//
+// extractPairs walks the prose sentences of a filled HTML report and emits, in document order:
+// a quote pair per distinct ledger quote cited by a findings sentence (facts sharing one quote
+// share one pair), a claim pair per sentence bound only by data-claim (the statement comes from
+// the claim index), and a translation pair per data-translates span. force_strong routes a pair
+// to the strong reader when the lexicon raised name_outside_quote or scope_widened_warn, when the
+// hedge-position heuristic fired, or when the source hedges and the sentence's last six tokens do
+// not. No model, no IO: the readers run as separate lanes. Plain Node ESM; no Hangul here.
+
+import { isExemptKind, proseSentences } from "./fidelity-binding.mjs"
+import { isKoreanTerm, lexicalPattern, lexiconChecks } from "./fidelity-lexicon.mjs"
+import { sentenceKey } from "./fidelity-state.mjs"
+import { buildContext } from "./gates-static.mjs"
+import { parseHtml, proseText } from "./html-lite.mjs"
+import { HEDGES } from "./lexicon.mjs"
+import { normalizeText } from "./text-normalize.mjs"
+
+/** Every pair kind the producers below emit; consumers count and schema pairs by this list. */
+export const PAIR_KINDS = Object.freeze(["quote", "claim", "translation"])
+const TAIL = 6
+
+const lang = (text) => (isKoreanTerm(text) ? "ko" : "en")
+/** Sentence text with its fact spans cut out, so a Korean value does not make an English sentence Korean. */
+const outsideFacts = (sentence) => sentence.facts.reduceRight((text, fact) => text.slice(0, fact.start - sentence.start) + text.slice(fact.end - sentence.start), sentence.text)
+
+/** Latin hedges match whole words ("exact"); Korean hedges match as substrings, one syllable as a token prefix ("stem"). */
+const HEDGE_RES = HEDGES.map((term) => lexicalPattern(term, isKoreanTerm(term) ? "stem" : "exact"))
+
+/** Hedge terms that are also month names; capitalized and followed by a date number, they are months. */
+const MONTH_WORDS = new Set(["may"])
+
+/** A capitalized Latin match after the first word is a name or a month ("in May"), not a hedge. */
+const isHedgeAt = (text, match) => {
+	if (!/^\p{Lu}/u.test(match[0]) || isKoreanTerm(match[0])) return true
+	if (match.index !== 0) return false
+	// At the start of a text a capital is only sentence case, so the word still hedges - unless it is a
+	// month name followed by a date number ("May 2024 sales rose.").
+	return !(MONTH_WORDS.has(match[0].toLowerCase()) && /^\s+\p{Nd}/u.test(text.slice(match.index + match[0].length)))
+}
+
+export const hasHedge = (text) => {
+	const normal = normalizeText(text)
+	return HEDGE_RES.some((re) => [...normal.matchAll(re)].some((match) => isHedgeAt(normal, match)))
+}
+
+/** The source hedges and none of the sentence's last six tokens carries a hedge. */
+function tailDropsHedge(source, sentence) {
+	if (!hasHedge(source)) return false
+	return !hasHedge(normalizeText(sentence).split(/\s+/u).filter(Boolean).slice(-TAIL).join(" "))
+}
+
+function slugOf(scope) {
+	const heading = scope.section?.heading
+	if (!heading) return ""
+	return (heading.attrs.id ?? proseText(heading))
+		.toLowerCase()
+		.replace(/[^\p{L}\p{N}]+/gu, "-")
+		.replace(/^-+|-+$/gu, "")
+}
+
+function claimMap(claims) {
+	const list = Array.isArray(claims) ? claims : (claims?.claims ?? [])
+	return new Map(list.map((claim) => [claim.claim_id, claim]))
+}
+
+const idsOf = (attr) => String(attr ?? "").split(/[\s,]+/u).filter(Boolean)
+
+/** Claim ids that bind a sentence: on the block or an ancestor, or on an inner element whose text the sentence holds. */
+function claimIdsFor(ctx, block, sentence) {
+	const { closest, descendants } = ctx.util
+	const up = block.owner.attrs["data-claim"] !== undefined ? block.owner : closest(block.owner, (node) => node.attrs?.["data-claim"] !== undefined)
+	const ids = up ? idsOf(up.attrs["data-claim"]) : []
+	const text = normalizeText(sentence.text)
+	for (const node of descendants(block.owner)) {
+		if (node.attrs?.["data-claim"] === undefined) continue
+		const inner = normalizeText(proseText(node))
+		if (inner !== "" && text.includes(inner)) ids.push(...idsOf(node.attrs["data-claim"]))
+	}
+	return [...new Set(ids)]
+}
+
+function exempt(ctx, node) {
+	return isExemptKind(node.attrs["data-kind"]) || Boolean(ctx.util.closest(node, (cur) => isExemptKind(cur.attrs?.["data-kind"])))
+}
+
+/** Lexicon routing for every fact span of one quote group in one sentence. */
+function lexiconRouted(group, sentence) {
+	const probes = group.facts.map((fact, index) => ({
+		pair_id: String(index),
+		quote: group.quote,
+		sentence: sentence.text,
+		span: [fact.start - sentence.start, fact.end - sentence.start],
+		qualifier: group.row.kind === "quote" ? null : (group.rows.get(fact.id)?.qualifier ?? null),
+	}))
+	return lexiconChecks(probes).forceStrong.length > 0
+}
+
+function quotePairs(sentence, rows, filled, slug, reportLang) {
+	const groups = new Map()
+	for (const fact of sentence.facts) {
+		const row = rows.get(fact.id)
+		if (!row?.quote || (filled && !filled.has(fact.id))) continue
+		const group = groups.get(row.quote) ?? { quote: row.quote, row, rows, facts: [], ids: [] }
+		group.facts.push(fact)
+		if (!group.ids.includes(fact.id)) group.ids.push(fact.id)
+		groups.set(row.quote, group)
+	}
+	const allIds = [...new Set(sentence.facts.map((fact) => fact.id))]
+	return [...groups.values()].map((group) => ({
+		kind: "quote",
+		sentence_key: sentenceKey(slug, allIds),
+		fact_ids: [...group.ids].sort(),
+		quote: group.quote,
+		sentence: sentence.text,
+		source_lang: lang(group.quote),
+		report_lang: reportLang ?? lang(outsideFacts(sentence)),
+		force_strong: lexiconRouted(group, sentence) || tailDropsHedge(group.quote, sentence.text),
+	}))
+}
+
+function claimPairs(ctx, block, sentence, byClaim, slug, reportLang) {
+	if (sentence.facts.length > 0 || exempt(ctx, block.owner)) return []
+	return claimIdsFor(ctx, block, sentence)
+		.map((id) => byClaim.get(id))
+		.filter(Boolean)
+		.map((claim) => ({
+			kind: "claim",
+			sentence_key: sentenceKey(slug, [claim.claim_id]),
+			claim_id: claim.claim_id,
+			statement: claim.statement,
+			sentence: sentence.text,
+			source_lang: lang(claim.statement),
+			report_lang: reportLang ?? lang(sentence.text),
+			force_strong: lexiconChecks([{ pair_id: "c", quote: claim.statement, sentence: sentence.text }]).forceStrong.length > 0 || tailDropsHedge(claim.statement, sentence.text),
+		}))
+}
+
+function translationPair(node, rows, filled, slug, reportLang) {
+	const id = node.attrs["data-translates"]
+	const row = rows.get(id)
+	const text = proseText(node)
+	if (!row?.quote || text === "" || (filled && !filled.has(id))) return null
+	return {
+		kind: "translation",
+		sentence_key: sentenceKey(slug, [id]),
+		fact_ids: [id],
+		quote: row.quote,
+		sentence: text,
+		source_lang: lang(row.quote),
+		report_lang: reportLang ?? lang(text),
+		force_strong: tailDropsHedge(row.quote, text),
+	}
+}
+
+/**
+ * @param {string} html the filled report
+ * @param {{ rows: object[] }} ledger
+ * @param {{ claims: object[] } | object[] | null} claims indexClaims(...) output or its claims array
+ * @param {{ fact_id: string }[]} [fillMap] fact fill's map; when given, only facts it filled are paired
+ * @returns {object[]} pairs with pair_id P1, P2, ... in document order
+ */
+export function extractPairs(html, ledger, claims, fillMap) {
+	const ctx = buildContext(parseHtml(html), {})
+	const rows = new Map((ledger?.rows ?? []).map((row) => [row.fact_id, row]))
+	const filled = Array.isArray(fillMap) ? new Set(fillMap.map((entry) => entry.fact_id)) : null
+	const byClaim = claimMap(claims)
+	const htmlLang = ctx.all.find((node) => node.tag === "html")?.attrs.lang
+	const reportLang = htmlLang ? htmlLang.toLowerCase().split("-")[0] : null
+	const out = []
+	const seenSpans = new Set()
+	for (const block of proseSentences(ctx)) {
+		const findings = Boolean(block.scope.section) && block.kind === null
+		if (!findings) continue
+		const slug = slugOf(block.scope)
+		for (const sentence of block.sentences) {
+			out.push(...quotePairs(sentence, rows, filled, slug, reportLang), ...claimPairs(ctx, block, sentence, byClaim, slug, reportLang))
+		}
+		for (const node of ctx.util.descendants(block.owner)) {
+			if (node.attrs?.["data-translates"] === undefined || seenSpans.has(node)) continue
+			seenSpans.add(node)
+			const pair = translationPair(node, rows, filled, slug, reportLang)
+			if (pair) out.push(pair)
+		}
+	}
+	return out.map((pair, index) => ({ pair_id: `P${index + 1}`, ...pair }))
+}

package/packages/omo-codex/plugin/skills/ulw-research/scripts/fact-recompute.mjs

@@ -0,0 +1,89 @@
+// Iterative postorder verifier: one snapshot and one arithmetic evaluation per reachable node.
+import { compute } from "./fact-derive.mjs"
+import { parseExpr } from "./fact-expr.mjs"
+import { isLegacyDerived, kindRuleErrors, readFlaws, readRow } from "./fact-ledger-kinds.mjs"
+import { rowErrors } from "./fact-ledger-schema.mjs"
+import { refuse } from "./unit-algebra.mjs"
+import { refusal } from "./unit-errors.mjs"
+import { object, sameDeclared } from "./unit-metadata.mjs"
+import { expectedUnit, malformed, ruleOf, unitMismatch } from "./unit-rules.mjs"
+
+const derived = (row) => row.kind === "derived" || row.kind === "converted"
+const fail = (detail, shape, expected) => ({ ok: false, detail, shape, expected })
+function refuseLegacy(row) {
+	if (isLegacyDerived(row)) refuse(`${row.fact_id} was written before precision, compatibility, and context existed; regenerate it with derive`, undefined, "legacy derived row", "a derived row regenerated by derive")
+}
+
+function verifyRow(row, rows) {
+	const rule = ruleOf(row.operation)
+	const ast = rule.expression ? parseExpr(row.formula) : null
+	const context = { ...row.context, rationale: row.rationale }
+	const unit = expectedUnit(row.operation, rows, { ast, context })
+	const expected = { unit, ...compute(row.operation, rows, unit, ast, context, row.formula), derived_from: rows.map((r) => r.fact_id) }
+	if (row.unit !== unit) return fail(`${row.fact_id}: ${unitMismatch(row.operation, rows, row.unit, unit)}`, "result_unit", expected)
+	for (const field of Object.keys(expected)) {
+		if (!sameDeclared(row[field], expected[field])) return fail(`${row.fact_id}: stored ${field} differs from the recomputed result`, `${field}_mismatch`, expected)
+	}
+	return { ok: true, expected }
+}
+
+export function verifyGraph(ledger, roots, requireActive = false, override = null) {
+	try {
+		if (!object(ledger) || !Array.isArray(ledger.rows)) refuse("ledger rows must be an array")
+		const byId = new Map()
+		for (const r of ledger.rows) {
+			if (!object(r) || typeof r.fact_id !== "string" || byId.has(r.fact_id)) refuse("ledger rows need distinct fact ids", undefined, "identity", "distinct fact ids")
+			byId.set(r.fact_id, r)
+		}
+		if (override) byId.set(override.fact_id, override)
+		const states = new Map()
+		const snapshots = new Map()
+		const results = new Map()
+		const stack = roots.map((id) => ({ id, finish: false }))
+		while (stack.length) {
+			const { id, finish } = stack.pop()
+			if (states.get(id) === "done") continue
+			if (!finish) {
+				if (states.get(id) === "open") refuse(`${id}: operands form a cycle`, undefined, "cycle", "an acyclic dependency graph")
+				const original = byId.get(id)
+				if (!original) refuse(`no ledger row ${id}`, undefined, "missing_operand", "an existing fact id")
+				const row = { ...readRow(original) }
+				snapshots.set(id, row)
+				if (requireActive && row.status !== "active") refuse(`${id} is ${row.status}`, undefined, "inactive_dependency", "active dependencies at every depth")
+				const [flaw] = readFlaws(row)
+				if (flaw) refuse(`${id}: ${flaw.detail}`, undefined, flaw.shape, flaw.expected)
+				if (derived(row)) {
+					refuseLegacy(row)
+					const shape = malformed(row)
+					if (shape) refuse(`${id}: ${shape}`, undefined, "malformed_row", "the operation's registry row shape")
+				}
+				const errors = derived(row) ? kindRuleErrors(row, id) : rowErrors(row)
+				if (errors.length) refuse(errors.join("; "), undefined, "row_schema", "a canonical row matching its kind and operation")
+				if (row.exact === null || row.value === null) refuse(`${id} has no exact value to derive from`)
+				states.set(id, "open")
+				stack.push({ id, finish: true })
+				if (derived(row)) for (const dependency of row.operands) stack.push({ id: dependency, finish: false })
+			} else {
+				const row = snapshots.get(id)
+				if (derived(row)) {
+					const result = verifyRow(row, row.operands.map((ref) => snapshots.get(ref)))
+					if (!result.ok) return result
+					const errors = rowErrors(row)
+					if (errors.length) refuse(errors.join("; "), undefined, "row_schema", "a complete canonical row")
+					results.set(id, result)
+				}
+				states.set(id, "done")
+			}
+		}
+		return results.get(roots[0]) ?? { ok: true }
+	} catch (error) { return { ok: false, ...refusal(error) } }
+}
+
+export function recompute(ledger, row) {
+	try {
+		refuseLegacy(row)
+		const shape = malformed(row)
+		if (shape) refuse(shape, undefined, "malformed_row", "a valid derived or converted row")
+		return verifyGraph(ledger, [row.fact_id], false, row)
+	} catch (error) { return { ok: false, ...refusal(error) } }
+}

package/packages/omo-codex/plugin/skills/ulw-research/scripts/fidelity-binding.mjs

@@ -0,0 +1,130 @@
+// fidelity-binding.mjs - sentence binding and claim-status checks of the fidelity gate.
+//
+// proseSentences walks every leaf prose element (p, li) outside figures, tables and page chrome,
+// keeps the offsets of each data-fact element inside its text, and splits it into sentences, so
+// every check sees which ledger rows a sentence cites. bindingChecks raises unbound_sentence (a
+// findings sentence with no fact, placeholder, citation, or data-claim), claim_status_invalid (a
+// data-claim id that is unknown, refuted or unresolved outside a Gaps or Contradictions section,
+// or high-risk and missing from verified-claims), and conflicting_readings (two active rows with
+// one value and unit but different periods or qualifiers cited in one sentence). Plain Node ESM.
+
+import { proseText } from "./html-lite.mjs"
+import { fromDecimal, eq } from "./rational.mjs"
+import { splitSentences } from "./sentence-split.mjs"
+import { normalizeUnit } from "./units.mjs"
+
+const SKIP = new Set(["code", "pre", "kbd", "samp", "script", "style", "svg", "template", "noscript"])
+const CONTAINERS = new Set(["figure", "table", "nav", "header", "footer"])
+/** data-kind values whose sentences need no binding (opinion and recommendation blocks). */
+export const EXEMPT_KINDS = Object.freeze(["opinion", "recommendation"])
+export const isExemptKind = (kind) => EXEMPT_KINDS.includes(kind)
+const DECIMAL = /^-?\d+(?:\.\d+)?$/
+/** Gaps and Contradictions headings (English, and the Korean gap/limit/contradiction/conflict words). */
+const GAPS_HEADING = /\b(?:gaps?|contradictions?|unresolved)\b|\uACF5\uBC31|\uD55C\uACC4|\uBAA8\uC21C|\uC0C1\uCDA9/i
+
+/** Canonical unit token, or the raw unit without spaces, lowercased; "" for none. */
+export function unitKey(unit) {
+	if (unit === null || unit === undefined) return ""
+	return normalizeUnit(String(unit)) ?? String(unit).replace(/\s+/g, "").toLowerCase()
+}
+
+/** Exact decimal equality; false when either side is not a plain decimal. */
+export function sameValue(a, b) {
+	const [x, y] = [String(a ?? "").replace(/,/g, ""), String(b ?? "").replace(/,/g, "")]
+	return DECIMAL.test(x) && DECIMAL.test(y) && eq(fromDecimal(x), fromDecimal(y))
+}
+
+function serialize(node, out) {
+	for (const child of node.children) {
+		if (child.tag === "#text") out.text += child.text
+		else if (child.tag === "br") out.text += " "
+		else if (SKIP.has(child.tag)) continue
+		else if (child.attrs["data-fact"] !== undefined) {
+			const text = proseText(child)
+			out.facts.push({ id: child.attrs["data-fact"], start: out.text.length, end: out.text.length + text.length, text, node: child })
+			out.text += text
+		} else serialize(child, out)
+	}
+	return out
+}
+
+/**
+ * Leaf prose elements with their sentences and the facts each sentence cites.
+ * @returns {{ owner: object, scope: object, kind: string|null, sentences: { text: string, start: number, end: number, facts: object[] }[] }[]}
+ */
+export function proseSentences(ctx) {
+	const { closest, descendants } = ctx.util
+	const blocks = []
+	for (const { node, scope } of ctx.elements) {
+		if (node.tag !== "p" && node.tag !== "li") continue
+		if (node.tag === "li" && descendants(node).some((child) => child.tag === "p" || child.tag === "li")) continue
+		if (closest(node, (cur) => CONTAINERS.has(cur.tag))) continue
+		const { text, facts } = serialize(node, { text: "", facts: [] })
+		const sentences = splitSentences(text).map((s) => ({ ...s, facts: facts.filter((f) => f.start < s.end && f.end > s.start) }))
+		blocks.push({ owner: node, scope, kind: ctx.kindOf(scope), sentences })
+	}
+	return blocks
+}
+
+function claimList(claims) {
+	const list = Array.isArray(claims) ? claims : (claims?.claims ?? [])
+	return new Map(list.map((claim) => [claim.claim_id, claim]))
+}
+
+const selfOrUp = (ctx, node, test) => test(node) || Boolean(ctx.util.closest(node, test))
+
+function claimDefects(ctx, claims) {
+	const byId = claimList(claims)
+	const out = []
+	for (const { node, scope } of ctx.elements) {
+		const attr = node.attrs["data-claim"]
+		if (attr === undefined) continue
+		const heading = scope.section?.heading
+		const inGaps = Boolean(heading) && GAPS_HEADING.test(proseText(heading))
+		for (const id of attr.split(/[\s,]+/).filter(Boolean)) {
+			const claim = byId.get(id)
+			let why = null
+			if (!claim) why = `data-claim ${id} is not in the claim index`
+			else if (/refuted|unresolved/i.test(claim.status) && !inGaps) why = `claim ${id} is ${claim.status} but asserted outside Gaps or Contradictions`
+			else if (!inGaps && /high/i.test(claim.risk) && !claim.in_verified) why = `high-risk claim ${id} is not in verified-claims`
+			if (why) out.push(ctx.defect("claim_status_invalid", node, why))
+		}
+	}
+	return out
+}
+
+function conflict(rows, facts) {
+	const cited = [...new Set(facts.map((f) => f.id))].map((id) => rows.get(id)).filter((row) => row && row.status === "active" && row.kind !== "quote")
+	for (let i = 0; i < cited.length; i++) {
+		for (let k = i + 1; k < cited.length; k++) {
+			const [a, b] = [cited[i], cited[k]]
+			if (!sameValue(a.value, b.value) || unitKey(a.unit) !== unitKey(b.unit)) continue
+			if (a.valid_at !== b.valid_at || a.qualifier !== b.qualifier) return [a, b]
+		}
+	}
+	return null
+}
+
+/**
+ * @param {object} ctx buildContext(...) from gates-static.mjs
+ * @param {{ claims: object[] } | object[] | null} claims indexClaims(...) output or its claims array
+ * @param {{ rows: object[] } | null} ledger
+ */
+export function bindingChecks(ctx, claims, ledger) {
+	const rows = new Map((ledger?.rows ?? []).map((row) => [row.fact_id, row]))
+	const out = []
+	for (const block of proseSentences(ctx)) {
+		const findings = Boolean(block.scope.section) && block.kind === null
+		const exempt = selfOrUp(ctx, block.owner, (cur) => isExemptKind(cur.attrs["data-kind"]) || cur.attrs["data-claim"] !== undefined)
+		for (const sentence of block.sentences) {
+			const bound = sentence.facts.length > 0 || ctx.util.citationMarker.test(sentence.text) || sentence.text.includes("{{")
+			if (findings && !exempt && !bound) out.push(ctx.defect("unbound_sentence", block.owner, `no fact, citation, or data-claim binds: "${sentence.text.slice(0, 80)}"`))
+			const pair = conflict(rows, sentence.facts)
+			if (pair) {
+				const [a, b] = pair
+				out.push(ctx.defect("conflicting_readings", block.owner, `${a.fact_id} (${a.valid_at ?? "no period"}) and ${b.fact_id} (${b.valid_at ?? "no period"}) share ${a.value} ${a.unit ?? ""} in one sentence`))
+			}
+		}
+	}
+	return [...out, ...claimDefects(ctx, claims)]
+}

package/packages/omo-codex/plugin/skills/ulw-research/scripts/fidelity-brief.mjs

@@ -0,0 +1,150 @@
+// fidelity-brief.mjs - parse the `## Fidelity` block of brief.md and decide whether it is locked.
+//
+// Grammar (fable-interview.md section 3, "Parser contract"): the block starts at a line
+// `## Fidelity` and ends at the next `## ` heading. It holds the seven field lines below, in any
+// order, each `<field>: <value> (answered_by: user|default|request|pending)`, plus an optional
+// `asked: <field>, <field>` history line. Values carry no parentheses, and `answered_by: pending`
+// is legal only with the value `pending`. Trailing whitespace on a line is ignored; a second
+// `## Fidelity` block is a parse error. This module parses; it never interprets a value, which
+// is each consumer's job. Plain Node ESM, node: builtins only.
+
+import { readFileSync } from "node:fs"
+
+export const FIDELITY_FIELDS = Object.freeze([
+	"as of",
+	"meaning check",
+	"source floor",
+	"derived numbers",
+	"languages",
+	"identifier classes",
+	"fallback",
+])
+export const FIDELITY_ANSWERED_BY = Object.freeze(["user", "default", "request", "pending"])
+const LOCKED_ANSWERED_BY = new Set(["user", "default", "request"])
+
+export const FIDELITY_INCOMPLETE = "fidelity_interview_incomplete"
+const INCOMPLETE_HINT =
+	"A fidelity field is pending or missing in brief.md; answer the open question or record its default, then rerun."
+
+const BLOCK_HEADING = /^## Fidelity\s*$/
+const NEXT_HEADING = /^## /
+const FIELD_LINE = /^([a-z][a-z ]*[a-z]): ([^()]+) \(answered_by: (user|default|request|pending)\)$/
+const ASKED_LINE = /^asked:(.*)$/
+
+function blockLines(markdown) {
+	const lines = String(markdown)
+		.split(/\r?\n/)
+		.map((line) => line.trimEnd())
+	const headings = lines.flatMap((line, index) => (BLOCK_HEADING.test(line) ? [index] : []))
+	if (headings.length === 0) return null
+	const [start, ...duplicates] = headings
+	const body = []
+	for (let index = start + 1; index < lines.length; index += 1) {
+		if (NEXT_HEADING.test(lines[index])) break
+		body.push({ text: lines[index], line: index + 1 })
+	}
+	return { body, duplicates: duplicates.map((index) => index + 1) }
+}
+
+function parseAsked(rest, line, text, errors) {
+	const names = rest
+		.split(",")
+		.map((name) => name.trim())
+		.filter(Boolean)
+	const unknown = names.filter((name) => !FIDELITY_FIELDS.includes(name))
+	if (unknown.length > 0) {
+		errors.push(`line ${line}: malformed asked line, unknown field ${unknown.join(", ")}: ${text}`)
+		return []
+	}
+	return names
+}
+
+function parseFieldLine({ text, line }, parsed, errors) {
+	const match = FIELD_LINE.exec(text)
+	if (!match) return errors.push(`line ${line}: malformed fidelity line: ${text}`)
+	const [, field, rawValue, answeredBy] = match
+	const value = rawValue.trim()
+	if (!FIDELITY_FIELDS.includes(field)) return errors.push(`line ${line}: malformed fidelity line, unknown field "${field}": ${text}`)
+	if (Object.hasOwn(parsed.fields, field)) return errors.push(`line ${line}: malformed fidelity line, duplicate field "${field}": ${text}`)
+	if ((answeredBy === "pending") !== (value === "pending")) {
+		return errors.push(`line ${line}: malformed fidelity line, value pending and answered_by: pending go together: ${text}`)
+	}
+	parsed.fields[field] = value
+	parsed.answered_by[field] = answeredBy
+}
+
+/**
+ * @param {string} briefMd the whole brief.md text
+ * @returns {{ present: boolean, fields: Record<string, string>, answered_by: Record<string, string>,
+ *   asked: string[], missing: string[], pending: string[], errors: string[] }}
+ */
+export function parseFidelityBlock(briefMd) {
+	const parsed = { present: false, fields: {}, answered_by: {}, asked: [], missing: [], pending: [], errors: [] }
+	const block = blockLines(briefMd)
+	if (block !== null) {
+		const { body, duplicates } = block
+		parsed.present = true
+		for (const line of duplicates) parsed.errors.push(`line ${line}: duplicate ## Fidelity block, only one is allowed`)
+		let sawAsked = false
+		for (const entry of body) {
+			if (entry.text.trim() === "") continue
+			const asked = ASKED_LINE.exec(entry.text)
+			if (!asked) {
+				parseFieldLine(entry, parsed, parsed.errors)
+				continue
+			}
+			if (sawAsked) parsed.errors.push(`line ${entry.line}: malformed asked line, duplicate: ${entry.text}`)
+			else parsed.asked = parseAsked(asked[1], entry.line, entry.text, parsed.errors)
+			sawAsked = true
+		}
+	}
+	parsed.missing = FIDELITY_FIELDS.filter((field) => !Object.hasOwn(parsed.fields, field))
+	parsed.pending = FIDELITY_FIELDS.filter((field) => parsed.answered_by[field] === "pending")
+	return parsed
+}
+
+/** Locked = no parse error and all seven fields present with answered_by user, default, or request. */
+export function isLocked(parsed) {
+	if (!parsed || parsed.errors.length > 0) return false
+	return FIDELITY_FIELDS.every((field) => Object.hasOwn(parsed.fields, field) && LOCKED_ANSWERED_BY.has(parsed.answered_by[field]))
+}
+
+/** The brief's `meaning check` field turns the meaning verifier off only when it reads "off". */
+export function meaningCheckEnabled(parsed) {
+	return !/^off\b/iu.test(parsed?.fields?.["meaning check"] ?? "")
+}
+
+function incompleteDefect(briefPath, parsed) {
+	const reasons = []
+	if (!parsed.present) reasons.push("no ## Fidelity block")
+	else if (parsed.missing.length > 0) reasons.push(`missing: ${parsed.missing.join(", ")}`)
+	if (parsed.pending.length > 0) reasons.push(`pending: ${parsed.pending.join(", ")}`)
+	return {
+		code: FIDELITY_INCOMPLETE,
+		severity: "blocker",
+		integrity: true,
+		selector: null,
+		line: null,
+		message: `${briefPath}: fidelity interview incomplete (${reasons.join("; ")})`,
+		hint: INCOMPLETE_HINT,
+	}
+}
+
+/**
+ * The lock check the reader-facing CLI stages run before writing anything.
+ * exitCode 0 = locked; 1 = pending, missing, or no brief (one `fidelity_interview_incomplete`
+ * defect); 2 = malformed block (usage error, `errors` name the lines, no defect).
+ * @param {string} briefPath
+ */
+export function requireLockedBrief(briefPath) {
+	let markdown = ""
+	try {
+		markdown = readFileSync(briefPath, "utf8")
+	} catch (error) {
+		if (error?.code !== "ENOENT") throw error
+	}
+	const parsed = parseFidelityBlock(markdown)
+	if (parsed.errors.length > 0) return { locked: false, exitCode: 2, defects: [], errors: parsed.errors, parsed }
+	if (isLocked(parsed)) return { locked: true, exitCode: 0, defects: [], errors: [], parsed }
+	return { locked: false, exitCode: 1, defects: [incompleteDefect(briefPath, parsed)], errors: [], parsed }
+}

package/packages/omo-codex/plugin/skills/ulw-research/scripts/fidelity-claims.mjs

@@ -0,0 +1,89 @@
+// fidelity-claims.mjs - the claim-kind checks of the deterministic fidelity gate.
+//
+// A mark (the data-claim-kind attribute) binds per cited fact: a data-fact element takes the kind
+// of its nearest mark (its own, else the closest marked ancestor), and only that kind is checked
+// against its ledger row (claim-kinds.mjs CLAIM_KINDS); a row that fails is the kind's defect
+// (unsupported_verification for `verified`). A fact under two marks of different kinds is a
+// claim_kind_mismatch, and a mark that owns no fact of its own kind (every fact under it cites
+// nothing or is shadowed by another kind) is its kind's defect: shadowing never satisfies a parent.
+// An unmarked fact citing a row neither fact nor quote accepts (an assumed row) is a
+// claim_kind_mismatch naming the mark it needs. A sentence that matches a kind's unmarked patterns warns (minor, verification_claim_unmarked)
+// unless every fact it cites is marked that kind; a sentence citing no fact warns unless its block
+// is marked. No model, no network.
+
+import { CLAIM_ATTRIBUTE, CLAIM_KINDS, claimSatisfied, unmarkedKinds, unmarkedRequires } from "./claim-kinds.mjs"
+import { proseSentences } from "./fidelity-binding.mjs"
+import { clip } from "./fidelity-trace.mjs"
+import { proseText } from "./html-lite.mjs"
+
+const marked = (node) => node.attrs?.[CLAIM_ATTRIBUTE] !== undefined
+const kindOf = (node) => node.attrs[CLAIM_ATTRIBUTE].trim()
+const isFact = (node) => node.attrs["data-fact"] !== undefined
+
+/** The marks over a node, nearest first: its own mark, then each marked ancestor. */
+function marksOver(node) {
+	const marks = []
+	for (let cur = node; cur; cur = cur.parent) if (marked(cur)) marks.push(cur)
+	return marks
+}
+
+const nearestKind = (node) => {
+	const [mark] = marksOver(node)
+	return mark ? kindOf(mark) : null
+}
+
+/** Whether `mark` binds `fact`: every mark from the fact up to `mark` names the same kind. */
+function binds(mark, fact) {
+	const marks = marksOver(fact)
+	const at = marks.indexOf(mark)
+	return at >= 0 && marks.slice(0, at + 1).every((cur) => kindOf(cur) === kindOf(mark))
+}
+
+/**
+ * @param {object} ctx buildContext(...) from gates-static.mjs
+ * @param {Map<string, object>} rows ledger rows by fact_id
+ * @param {(row: object) => string | undefined} provenanceOf meta.json kind of the snapshot holding the row's own locator artifact
+ * @param {(code: string, node: object, message: string) => void} flag
+ */
+export function claimChecks(ctx, rows, provenanceOf, flag) {
+	for (const node of ctx.all.filter(marked)) {
+		const name = kindOf(node)
+		const entry = CLAIM_KINDS[name]
+		if (!entry) {
+			flag("claim_kind_mismatch", node, `${CLAIM_ATTRIBUTE}="${name}" is not a claim kind (${Object.keys(CLAIM_KINDS).join(", ")})`)
+			continue
+		}
+		const cited = [node, ...ctx.util.descendants(node)].filter(isFact)
+		if (cited.some((fact) => binds(node, fact))) continue
+		const why = cited.length === 0 ? "cites no fact" : "cites no fact of its own (a nested mark of another kind never satisfies it)"
+		flag(entry.defect, node, `"${clip(proseText(node))}" is marked ${name} but ${why}`)
+	}
+	for (const fact of ctx.all.filter(isFact)) {
+		const [nearest, ...outer] = marksOver(fact)
+		const id = fact.attrs["data-fact"]
+		if (!nearest) {
+			const need = unmarkedRequires(rows.get(id))
+			const mark = need && `{{${CLAIM_KINDS[need].prefix}:${id}}} (${CLAIM_ATTRIBUTE}="${need}")`
+			if (need) flag("claim_kind_mismatch", fact, `${id} is a ${rows.get(id).kind} row cited with no mark; an unmarked numeral reads as a fact, so mark it ${mark}`)
+			continue
+		}
+		const name = kindOf(nearest)
+		const conflict = outer.find((mark) => kindOf(mark) !== name)
+		if (conflict) flag("claim_kind_mismatch", fact, `${id} is marked ${name} inside a ${kindOf(conflict)} mark; a fact carries one claim kind`)
+		const entry = CLAIM_KINDS[name]
+		const row = rows.get(id)
+		if (!entry || !row || claimSatisfied(name, row, provenanceOf(row))) continue
+		const source = entry.sourceKinds ? ` and not one whose own locator artifact sits in a ${entry.sourceKinds.join(" or ")} snapshot` : ""
+		flag(entry.defect, fact, `${row.fact_id} is marked ${name}, but it is a ${row.kind} row${source}`)
+	}
+	for (const block of proseSentences(ctx)) {
+		if (block.kind === "sources" || block.kind === "toc") continue
+		for (const sentence of block.sentences) {
+			for (const name of unmarkedKinds(sentence.text)) {
+				const covered = sentence.facts.length > 0 ? sentence.facts.every((fact) => nearestKind(fact.node) === name) : nearestKind(block.owner) === name
+				if (covered) continue
+				flag(CLAIM_KINDS[name].unmarked.code, block.owner, `"${clip(sentence.text)}" reads as a ${name} claim but cites a fact with no ${CLAIM_ATTRIBUTE}="${name}" mark`)
+			}
+		}
+	}
+}