@smartmemory/compose 0.3.5-beta → 0.3.7

package/bin/judgment-import.js

@@ -0,0 +1,572 @@
+#!/usr/bin/env node
+/**
+ * bin/judgment-import.js — one-time markdown → records import (S07/W3).
+ *
+ * Parses the hand-written judgment canon (REGISTER.md, LEDGER.md,
+ * OBJECTIVE.md, positions/*.md) into records and writes them THROUGH the
+ * writer with `via: 'import'` — historical transcription, not new
+ * authorship: original dates and `[ASSERT]` grounding are preserved
+ * (ASSERT claims gain a citation-elicitation pointing at the imported
+ * source), `minutes` costs map to `hours` with a note flag, and curated
+ * banners/prose become `note` records with placement anchors.
+ *
+ * Ledger entries that predate the schema's kind-specific requirements and
+ * cannot be mapped losslessly (e.g. the slot-assignment `rank` entry) fall
+ * back to `note` records that keep the original heading in the title —
+ * deterministic, honest, and reviewable in the cutover diff.
+ *
+ * Dry-run stages the import in a temp workspace and diffs the regenerated
+ * projections against the current hand-written files; the real run is
+ * refused if records already exist (import is one-time). Kept for provider
+ * migrations (PROVIDER-SEAM: switching is a re-import, never a sync).
+ */
+import { readFileSync, readdirSync, existsSync, mkdtempSync } from 'node:fs';
+import { join, basename } from 'node:path';
+import { tmpdir } from 'node:os';
+
+import {
+  judgmentPositionCreate,
+  judgmentJointAdd,
+  judgmentLedgerAppend,
+} from '../lib/judgment-writer.js';
+import {
+  computeRecordHashes,
+  writeManifest,
+} from '../lib/judgment-attest.js';
+import { getJudgmentValidator } from '../lib/judgment/schema.js';
+
+const METHODS = ['EXT', 'INT', 'CONSTRUCT', 'ASSERT', 'STRADDLE'];
+const RULED_COSTS = ['hours', 'days', 'weeks', 'months'];
+const DEFAULT_DATE = '2026-07-20';
+
+/** Map a raw cost cell onto the ruled COARSE-BUCKETS. */
+export function mapCost(raw) {
+  const cost = String(raw ?? '').toLowerCase().trim();
+  if (RULED_COSTS.includes(cost)) return { cost, note: null };
+  if (cost === 'minutes') return { cost: 'hours', note: 'import: cost "minutes" mapped to "hours" (COARSE-BUCKETS)' };
+  return { cost: 'hours', note: `import: cost "${raw}" is not a ruled bucket — defaulted to "hours"` };
+}
+
+// ---------------------------------------------------------------------------
+// Shared parsing helpers
+// ---------------------------------------------------------------------------
+
+function cells(line) {
+  return line.replace(/^\|/, '').replace(/\|\s*$/, '').split('|').map((c) => c.trim());
+}
+
+function plain(cell) {
+  return cell.replace(/\*\*/g, '').replace(/`/g, '').replace(/~~/g, '').trim();
+}
+
+function ts(date) {
+  return `${date}T12:00:00Z`;
+}
+
+function convictionFrom(text) {
+  const t = (text ?? '').toLowerCase();
+  const level = ['high', 'medium', 'low'].find((l) => t.includes(l)) ?? 'medium';
+  const source = t.includes('inferred') ? 'inferred' : (t.includes('stated') || t.includes('owner') ? 'stated' : 'inferred');
+  return { level, source };
+}
+
+function isTableRow(line) {
+  return line.startsWith('|') && !/^\|[\s\-:|]+\|?\s*$/.test(line);
+}
+
+// ---------------------------------------------------------------------------
+// REGISTER.md → joints + anchored notes
+// ---------------------------------------------------------------------------
+
+function parseRegister(text, objectiveQuestions) {
+  const lines = text.split('\n');
+  const joints = [];
+  const notes = [];
+  const struckRows = new Map(); // slug → question (for dissolved joints)
+
+  let section = 'banner';
+  const bannerLines = [];
+  const footerBlocks = []; // { title, body[] }
+  let currentFooter = null;
+
+  const sectionOf = (heading) => {
+    if (/^## Under test/.test(heading)) return { name: 'table', state: 'under_test', rank: 'high' };
+    if (/^## Open — high/.test(heading)) return { name: 'table', state: 'open', rank: 'high' };
+    if (/^## Open — medium/.test(heading)) return { name: 'table', state: 'open', rank: 'medium' };
+    if (/^## Resolved/.test(heading)) return { name: 'resolved' };
+    if (/^## Dissolved/.test(heading)) return { name: 'dissolved' };
+    return { name: 'footer', title: heading.replace(/^##\s*/, '') };
+  };
+
+  let mode = { name: 'banner' };
+  let dissolvedBody = [];
+
+  for (const raw of lines) {
+    const line = raw.trimEnd();
+    if (/^# /.test(line)) continue;
+    if (/^## /.test(line)) {
+      mode = sectionOf(line);
+      if (mode.name === 'footer') {
+        currentFooter = { title: mode.title, body: [] };
+        footerBlocks.push(currentFooter);
+      }
+      continue;
+    }
+    if (mode.name === 'banner') {
+      bannerLines.push(line);
+      continue;
+    }
+    if (mode.name === 'footer') {
+      currentFooter.body.push(line);
+      continue;
+    }
+    if (mode.name === 'dissolved') {
+      dissolvedBody.push(line);
+      continue;
+    }
+    if (mode.name === 'table' || mode.name === 'resolved') {
+      if (!isTableRow(line)) {
+        // inter-table prose (Moved here…, Re-weighted…, demotion notes)
+        if (line.trim()) footerBlocks.push({ title: 'register prose', body: [line] });
+        continue;
+      }
+      const row = cells(line);
+      if (row.length < 4 || plain(row[0]).toLowerCase() === 'joint') continue;
+      if (raw.includes('~~')) {
+        struckRows.set(plain(row[0]), plain(row[1]));
+        continue;
+      }
+      if (mode.name === 'resolved') {
+        const [slugCell, dateCell, byCell, outcomeCell] = row;
+        const slug = plain(slugCell);
+        const date = plain(dateCell);
+        const question = objectiveQuestions.get(slug) ?? '(question not recorded at import)';
+        const refMatch = /\(`?(decide: [a-z0-9-]+)`?\)/.exec(outcomeCell);
+        const { cost, note } = mapCost('minutes'); // the register records these as minutes-scale ASSERT rulings
+        joints.push({
+          date,
+          record: {
+            slug,
+            question,
+            branch_true: '(not recorded at import)',
+            branch_false: '(not recorded at import)',
+            resolve_by: METHODS.find((m) => byCell.includes(m)) ?? 'ASSERT',
+            cost,
+            rank: 'medium',
+            state: 'resolved',
+            flags: [note].filter(Boolean),
+            resolution: {
+              outcome: 'resolved',
+              evidence: plain(outcomeCell),
+              elicitation: {
+                asked: question,
+                answered_at: ts(date),
+                answer_ref: refMatch ? refMatch[1] : 'docs/judgment/LEDGER.md',
+              },
+            },
+          },
+        });
+        continue;
+      }
+      // open / under_test tables
+      const [slugCell, qCell, tCell, fCell, byCell, costCell] = row;
+      const slug = plain(slugCell);
+      // Strip flag tokens BEFORE method detection — 'EXT-UNREACHABLE' must not
+      // read as method EXT (already-knew resolves by INT).
+      const byClean = plain(byCell ?? '').replace(/EXT-UNREACHABLE/g, '').replace(/blocked on sharpening/gi, '');
+      const method = (/\b(EXT|INT|CONSTRUCT|ASSERT|STRADDLE)\b/.exec(byClean) ?? [])[1] ?? 'INT';
+      const { cost, note } = mapCost(plain(costCell ?? ''));
+      const flags = [];
+      if (/EXT-UNREACHABLE/.test(byCell ?? '')) flags.push('EXT-UNREACHABLE');
+      if (/blocked on sharpening/i.test(byCell ?? '')) flags.push('blocked-on-sharpening');
+      const methodNote = byClean
+        .replace(new RegExp(`\\b${method}\\b`), '')
+        .replace(/\s{2,}/g, ' ')
+        .replace(/^[\s—·-]+|[\s—·-]+$/g, '');
+      if (methodNote) flags.push(`method-note: ${methodNote}`);
+      if (note) flags.push(note);
+      joints.push({
+        date: DEFAULT_DATE,
+        record: {
+          slug,
+          question: plain(qCell),
+          branch_true: plain(tCell),
+          branch_false: plain(fCell),
+          resolve_by: method,
+          cost,
+          rank: mode.rank,
+          state: mode.state,
+          flags,
+        },
+      });
+    }
+  }
+
+  // Dissolved joints from struck rows + the Dissolved section prose.
+  const dissolvedText = dissolvedBody.join('\n');
+  for (const [slug, question] of struckRows) {
+    const halves = [...new Set([...dissolvedText.matchAll(/\bthe ([A-Z][a-z]+)\b/g)].map((m) => m[1].toLowerCase()))];
+    joints.push({
+      date: DEFAULT_DATE,
+      record: {
+        slug,
+        question: question || '(question not recorded at import)',
+        branch_true: '(not recorded at import)',
+        branch_false: '(not recorded at import)',
+        resolve_by: 'CONSTRUCT',
+        cost: 'hours',
+        rank: 'medium',
+        state: 'dissolved',
+        flags: ['import: resolve_by/cost not recorded on the dissolved row — defaulted'],
+        dissolution: { decomposed_into: halves.length ? halves : ['recorded-in-dissolution-note'] },
+      },
+    });
+  }
+
+  notes.push({
+    date: '2026-07-22',
+    event: { kind: 'note', title: 'Register banner (imported)', body: bannerLines.filter(Boolean).join('\n'), anchor: 'register-header' },
+  });
+  if (dissolvedText.trim()) {
+    footerBlocks.push({ title: 'Dissolved', body: dissolvedBody });
+  }
+  for (const block of footerBlocks) {
+    const body = block.body.filter(Boolean).join('\n').trim();
+    if (!body) continue;
+    notes.push({
+      date: '2026-07-22',
+      event: { kind: 'note', title: `Register: ${block.title} (imported)`, body, anchor: 'register-footer' },
+    });
+  }
+  return { joints, notes };
+}
+
+// ---------------------------------------------------------------------------
+// OBJECTIVE.md → position 'objective' + question map
+// ---------------------------------------------------------------------------
+
+function parseObjective(text) {
+  const questions = new Map();
+  for (const line of text.split('\n')) {
+    if (!isTableRow(line)) continue;
+    const row = cells(line);
+    if (row.length >= 3 && /`[a-z0-9-]+`/.test(row[0])) {
+      questions.set(plain(row[0]), plain(row[1]));
+    }
+  }
+  const primary = /\*\*Primary:\*\*\s*([\s\S]*?)\n\n/.exec(text)?.[1]?.replace(/\n/g, ' ').trim()
+    ?? '(objective not parseable at import)';
+  // The file's own header tags it `[ASSERT]` — preserved verbatim (binding:
+  // import preserves grounding). The elicitation citation records honestly
+  // that this was back-inferred, never elicited; the health warning survives
+  // in the anchored prose note.
+  const position = {
+    date: DEFAULT_DATE,
+    slug: 'objective',
+    record: {
+      slug: 'objective',
+      claims: [{
+        id: 'c1',
+        text: primary,
+        grounding: 'ASSERT',
+        supports: [],
+        elicitation: {
+          asked: 'What are we optimizing for?',
+          answered_at: ts(DEFAULT_DATE),
+          answer_ref: 'import:docs/judgment/OBJECTIVE.md (back-inferred draft — NOT owner-confirmed, see health warning)',
+        },
+      }],
+      conviction: { level: 'low', source: 'inferred' },
+    },
+  };
+  const note = {
+    date: DEFAULT_DATE,
+    event: { kind: 'note', title: 'OBJECTIVE.md (imported prose)', body: text.trim(), anchor: 'position:objective' },
+  };
+  return { position, note, questions };
+}
+
+// ---------------------------------------------------------------------------
+// positions/*.md → position records + prose notes
+// ---------------------------------------------------------------------------
+
+function parsePositionFile(slug, text, sourceRel) {
+  const date = /\*\*Held since:\*\*\s*(\d{4}-\d{2}-\d{2})/.exec(text)?.[1]
+    ?? /(\d{4}-\d{2}-\d{2})/.exec(text)?.[1] ?? DEFAULT_DATE;
+  const conviction = convictionFrom(/\*\*Conviction:\*\*([^\n]*)/.exec(text)?.[1] ?? '');
+
+  const claims = [];
+  const cite = (asked) => ({ asked, answered_at: ts(date), answer_ref: `import:${sourceRel}` });
+
+  // Numbered argument table: | # | Step | Grounding |
+  for (const line of text.split('\n')) {
+    if (!isTableRow(line)) continue;
+    const row = cells(line);
+    if (row.length >= 3 && /^\d+$/.test(plain(row[0]))) {
+      const id = `c${plain(row[0])}`;
+      const step = plain(row[1]);
+      const groundingCell = row[2];
+      const derivedFrom = /derived from (\d+)/i.exec(groundingCell);
+      let claim;
+      if (derivedFrom) {
+        claim = { id, text: step, grounding: 'DERIVED', supports: [`c${derivedFrom[1]}`] };
+      } else {
+        // Preserve the recorded grounding verbatim (binding); AGENT only when
+        // the cell names none of the vocabulary.
+        const grounding = (/\b(EXT|INT|ASSERT|DERIVED|AGENT)\b/.exec(groundingCell) ?? [])[1] ?? 'AGENT';
+        claim = { id, text: step, grounding, supports: [] };
+        if (grounding === 'ASSERT') claim.elicitation = cite(step);
+      }
+      if (/owner-locked/.test(groundingCell)) claim.owner_locked = true;
+      claims.push(claim);
+    }
+  }
+  // Evidence bullets: - `[EXT]` text
+  if (claims.length === 0) {
+    let n = 0;
+    for (const line of text.split('\n')) {
+      const m = /^-\s*`\[(EXT|INT|ASSERT|AGENT|DERIVED)\]`\s*(.+)$/.exec(line.trim());
+      if (!m) continue;
+      n += 1;
+      const claim = { id: `c${n}`, text: plain(m[2]), grounding: m[1], supports: [] };
+      if (m[1] === 'ASSERT') claim.elicitation = cite(claim.text);
+      if (/owner-locked/.test(line)) claim.owner_locked = true;
+      claims.push(claim);
+    }
+  }
+  if (claims.length === 0) {
+    const claimSection = /## (?:The )?[Cc]laim\n+([\s\S]*?)(?:\n## |$)/.exec(text)?.[1]?.replace(/\n/g, ' ').trim();
+    claims.push({ id: 'c1', text: claimSection ?? '(claim not parseable at import)', grounding: 'AGENT', supports: [] });
+  }
+
+  return {
+    position: { date, slug, record: { slug, claims, conviction } },
+    note: {
+      date,
+      event: { kind: 'note', title: `${basename(sourceRel)} (imported prose)`, body: text.trim(), anchor: `position:${slug}` },
+    },
+  };
+}
+
+// ---------------------------------------------------------------------------
+// LEDGER.md → ledger events (schema-mapped, note fallback)
+// ---------------------------------------------------------------------------
+
+const LEDGER_KINDS = ['decide', 'kill', 'override', 'escalate', 'calibrate', 'postmortem', 'rank', 'note', 'correct', 'open'];
+
+function parseLedger(text) {
+  const lines = text.split('\n');
+  const entries = [];
+  let banner = [];
+  let sawSection = false;
+  let currentDate = DEFAULT_DATE;
+  let current = null;
+
+  const flush = () => {
+    if (current) entries.push(current);
+    current = null;
+  };
+
+  for (const raw of lines) {
+    const line = raw.trimEnd();
+    const heading = /^###\s+([a-z]+):\s*(.+)$/.exec(line);
+    if (/^## /.test(line)) {
+      flush();
+      sawSection = true;
+      const date = /(\d{4}-\d{2}-\d{2})/.exec(line)?.[1];
+      if (date) currentDate = date;
+      entries.push({ kind: 'note', title: `Ledger section: ${line.replace(/^##\s*/, '')} (imported)`, body: [], anchor: 'ledger', date: currentDate });
+      continue;
+    }
+    if (heading && LEDGER_KINDS.includes(heading[1])) {
+      flush();
+      const rest = heading[2].trim();
+      const date = /(\d{4}-\d{2}-\d{2})/.exec(rest)?.[1] ?? currentDate;
+      current = { kind: heading[1], title: rest, body: [], anchor: null, date };
+      continue;
+    }
+    if (line === '---') { flush(); continue; }
+    if (!sawSection && !current) { banner.push(line); continue; }
+    if (current) current.body.push(line);
+    else if (line.trim()) entries.push({ kind: 'note', title: 'Ledger prose (imported)', body: [line], anchor: 'ledger', date: currentDate });
+  }
+  flush();
+
+  const validator = getJudgmentValidator();
+  const probeProvenance = { actor: 'agent', written_at: ts(DEFAULT_DATE), via: 'import' };
+
+  const events = [];
+  if (banner.filter(Boolean).length) {
+    events.push({
+      date: DEFAULT_DATE,
+      event: { kind: 'note', title: 'Ledger banner (imported)', body: banner.filter(Boolean).join('\n'), anchor: 'ledger-header' },
+    });
+  }
+  for (const entry of entries) {
+    const body = (entry.body ?? []).filter(Boolean).join('\n').trim();
+    const event = { kind: entry.kind, title: entry.title, body };
+    if (entry.anchor) event.anchor = entry.anchor;
+    if (entry.kind === 'decide' || entry.kind === 'kill') {
+      const rejectedLine = /\*\*Rejected[^:]*:\*\*\s*([\s\S]*?)(?:\n\*\*|$)/.exec(body)?.[1];
+      if (entry.kind === 'decide') {
+        event.rejected = rejectedLine
+          ? [{ what: rejectedLine.replace(/\n/g, ' ').trim(), why: 'recorded in the imported entry body' }]
+          : [];
+        event.conviction = convictionFrom(/\*\*Conviction:\*\*([^\n]*)/.exec(body)?.[1] ?? '');
+      }
+    }
+    if (entry.kind === 'override') {
+      event.reason = body.split('\n')[0] || entry.title;
+    }
+    if (!body) delete event.body;
+
+    const { valid } = validator.validate('ledger_event', { ...event, provenance: probeProvenance });
+    if (valid) {
+      events.push({ date: entry.date, event });
+    } else {
+      // pre-schema entry shape → honest note fallback, original heading kept
+      events.push({
+        date: entry.date,
+        event: { kind: 'note', title: `${entry.kind}: ${entry.title} (imported as note)`, body: body || entry.title, anchor: 'ledger' },
+      });
+    }
+  }
+  return events;
+}
+
+// ---------------------------------------------------------------------------
+// The import
+// ---------------------------------------------------------------------------
+
+function readSources(cwd) {
+  const dir = join(cwd, 'docs', 'judgment');
+  const positionsDir = join(dir, 'positions');
+  const positionFiles = existsSync(positionsDir)
+    ? readdirSync(positionsDir).filter((f) => f.endsWith('.md')).sort()
+    : [];
+  return {
+    register: readFileSync(join(dir, 'REGISTER.md'), 'utf8'),
+    ledger: readFileSync(join(dir, 'LEDGER.md'), 'utf8'),
+    objective: readFileSync(join(dir, 'OBJECTIVE.md'), 'utf8'),
+    positions: positionFiles.map((f) => ({
+      slug: f.replace(/\.md$/, ''),
+      rel: `docs/judgment/positions/${f}`,
+      text: readFileSync(join(positionsDir, f), 'utf8'),
+    })),
+  };
+}
+
+function buildPlan(sources) {
+  const objective = parseObjective(sources.objective);
+  const register = parseRegister(sources.register, objective.questions);
+  const ledger = parseLedger(sources.ledger);
+
+  const positions = [objective.position];
+  const notes = [objective.note];
+  for (const file of sources.positions) {
+    const parsed = parsePositionFile(file.slug, file.text, file.rel);
+    positions.push(parsed.position);
+    notes.push(parsed.note);
+  }
+  // Event order: ledger banner + entries (chronology), then register notes,
+  // then position prose notes.
+  const events = [...ledger, ...register.notes, ...notes];
+  return { positions, joints: register.joints, events };
+}
+
+async function executePlan(cwd, plan) {
+  for (const p of plan.positions) {
+    await judgmentPositionCreate(cwd, p.record, { via: 'import', writtenAt: ts(p.date) });
+  }
+  for (const j of plan.joints) {
+    await judgmentJointAdd(cwd, j.record, { via: 'import', writtenAt: ts(j.date) });
+  }
+  for (const e of plan.events) {
+    await judgmentLedgerAppend(cwd, e.event, { via: 'import', writtenAt: ts(e.date) });
+  }
+}
+
+function diffProjections(stagingCwd, cwd) {
+  const stagingDir = join(stagingCwd, 'docs', 'judgment');
+  const files = ['REGISTER.md', 'LEDGER.md', 'OBJECTIVE.md', 'index.md',
+    ...readdirSync(join(stagingDir, 'positions')).filter((f) => f.endsWith('.md')).map((f) => `positions/${f}`)];
+  return files.map((rel) => {
+    const generated = readFileSync(join(stagingDir, rel), 'utf8');
+    const currentPath = join(cwd, 'docs', 'judgment', rel);
+    const current = existsSync(currentPath) ? readFileSync(currentPath, 'utf8') : null;
+    return { file: `docs/judgment/${rel}`, changed: current !== generated, new: current === null };
+  });
+}
+
+/**
+ * Run the import. Dry-run: stage in a temp workspace, report counts + the
+ * projection diff, write nothing into cwd. Real run: refuse if records
+ * exist, then write records through the writer (projections cut over as a
+ * side-effect of each write).
+ *
+ * @param {string} cwd
+ * @param {{ dryRun?: boolean }} [opts]
+ */
+export async function runJudgmentImport(cwd, { dryRun = false } = {}) {
+  // Preflight the configured canon provider: if judgment.provider selects an
+  // unimplemented/unknown backend this throws HERE, before any staging or
+  // copying — floor records must never land under a config that rejects them.
+  const { createJudgmentStore } = await import('../lib/judgment/store/index.js');
+  createJudgmentStore(cwd);
+  if (existsSync(join(cwd, 'docs', 'judgment', 'records'))) {
+    throw new Error(
+      'judgment records already exist under docs/judgment/records/ — import is one-time (remove records/ to re-run; '
+      + 'if a cutover was interrupted after records landed, regenerate projections via lib/judgment-gen.js#regenerateProjections)',
+    );
+  }
+  const sources = readSources(cwd);
+  const plan = buildPlan(sources);
+
+  // Stage the FULL import in a temp workspace: every record goes through the
+  // writer there, and the staged projections are what the human gate diffs
+  // against the current hand-written files (staging path is returned so the
+  // diff is actually inspectable).
+  const staging = mkdtempSync(join(tmpdir(), 'judgment-import-staging-'));
+  await executePlan(staging, plan);
+  const diff = diffProjections(staging, cwd);
+
+  if (!dryRun) {
+    // Crash-safe cutover: the writer-produced records land in cwd via copy →
+    // single rename (no partial records/ tree on interruption), then the
+    // projections regenerate from them. A crash after the rename self-heals:
+    // every subsequent writer op regenerates all projections.
+    const { cpSync, renameSync: rename } = await import('node:fs');
+    const { regenerateProjections } = await import('../lib/judgment-gen.js');
+    const dst = join(cwd, 'docs', 'judgment', 'records');
+    const tmpDst = `${dst}.importing.${process.pid}`;
+    cpSync(join(staging, 'docs', 'judgment', 'records'), tmpDst, { recursive: true });
+    rename(tmpDst, dst);
+    regenerateProjections(cwd);
+    // Trust-on-first-use baseline for the promoted TARGET records. Never use
+    // the staging manifest here: its workspace paths are disposable.
+    writeManifest(cwd, computeRecordHashes(cwd));
+  }
+
+  return {
+    dryRun,
+    staging,
+    counts: {
+      positions: plan.positions.length,
+      joints: plan.joints.length,
+      ledger_events: plan.events.length,
+    },
+    diff,
+  };
+}
+
+// CLI: node bin/judgment-import.js [--dry-run] [cwd]
+if (import.meta.url === `file://${process.argv[1]}`) {
+  const dryRun = process.argv.includes('--dry-run');
+  const cwd = process.argv.slice(2).find((a) => !a.startsWith('--')) ?? process.cwd();
+  runJudgmentImport(cwd, { dryRun }).then((result) => {
+    console.log(JSON.stringify(result, null, 2));
+  }).catch((err) => {
+    console.error(err.message);
+    process.exit(1);
+  });
+}

package/contracts/feature-json.schema.json

@@ -40,6 +40,11 @@
       "maximum": 4,
       "description": "COMP-TRIAGE-5: raw machine tier (0-4) from the front estimator/refinement. Maps to `complexity` via tierToComplexity."
     },
+    "triageConfidence": {
+      "type": "string",
+      "enum": ["high", "medium", "low"],
+      "description": "COMP-TRIAGE-6: confidence from the front scope estimate."
+    },
     "estimateSource": {
       "type": "string",
       "enum": ["front", "refined", "escalated"],