diff --git a/.changeset/kind-cats-leap.md b/.changeset/kind-cats-leap.md
new file mode 100644
index 000000000..3f1bc485b
--- /dev/null
+++ b/.changeset/kind-cats-leap.md
@@ -0,0 +1,5 @@
+---
+type: Fixed
+pr: 4876
+---
+**init.manager stops reporting dates, shas and ledger ids as phase dependencies** — dep_phases scraped every digit run out of a phase's Depends-on prose, so phases whose prose declares no dependency read as blocked (dates split into year/month/day, git shas fragmenting, WINDOWS ledger ids, even the phase's own number); extraction now pulls only Phase-prefixed references — including "Phases 1, 2, and 3" lists and "Phase 1-3" ranges — and planning-inspect's dependencies field uses the same anchored grammar. (#4764)
diff --git a/src/init.cts b/src/init.cts
index 5545526e0..56eaca26d 100644
--- a/src/init.cts
+++ b/src/init.cts
@@ -105,6 +105,7 @@ const {
   matchPhaseDirs,
   stripProjectCodePrefix,
   PHASE_NUMBER_TOKEN_SOURCE,
+  PHASE_DEP_REF_SOURCE,
   isForeignPrefixedPhaseQuery,
   isSentinelPhaseId,
   extractPhaseToken,
@@ -3105,6 +3106,23 @@ function cmdInitManager(cwd: string, raw: boolean): void {
     return reaches(numA, numB) || reaches(numB, numA);
   }
 
+  // #4764: a phase reference in depends_on prose is a PHASE-SHAPED token in
+  // context — directly following "Phase"/"Phases" — never a bare digit run.
+  // The previous whole-field scrape matched the token grammar against every
+  // digit run, so calendar dates ("2026-09-14" → 2026, 09, 14), git shas
+  // ("8bf403100d" → 8b, 403100d, …), bracketed ledger ids (WINDOWS #1843) and
+  // the row's OWN number all became "dependencies", and deps_satisfied came
+  // back false for phases whose prose declares none (50 of 92 phases in the
+  // reporter's milestone). The anchored grammar (owned by phase-id.cts as
+  // PHASE_DEP_REF_SOURCE, shared with planning-inspect's dependencies) keeps
+  // lists fully extracted ("Phases 601 and 602", "Phase 601, 602, and 603",
+  // "Phase 1-3") — silently dropping a REAL dependency would clear
+  // deps_satisfied prematurely, the dangerous direction. Negation prose
+  // ("dropped the dependency on Phase 654") is NOT detected: the issue's own
+  // minimum keeps such tokens.
+  const depPhaseRefRe = new RegExp(`${PHASE_DEP_REF_SOURCE}`, 'gi');
+  const depTokenRe = new RegExp(`${PHASE_NUMBER_TOKEN_SOURCE}`, 'gi');
+
   for (const phase of phases) {
     if (
       !phase['depends_on'] ||
@@ -3112,7 +3130,23 @@ function cmdInitManager(cwd: string, raw: boolean): void {
     ) {
       phase['deps_satisfied'] = true;
     } else {
-      const depNums = (phase['depends_on'] as string).match(new RegExp(`${PHASE_NUMBER_TOKEN_SOURCE}`, 'gi')) || [];
+      const prose = phase['depends_on'] as string;
+      const ownNumber = normalizePhaseNumber(phase['number'] as string);
+      const depNums: string[] = [];
+      const seen = new Set<string>();
+      let refMatch: RegExpExecArray | null;
+      depPhaseRefRe.lastIndex = 0;
+      while ((refMatch = depPhaseRefRe.exec(prose)) !== null) {
+        let tok: RegExpExecArray | null;
+        depTokenRe.lastIndex = 0;
+        while ((tok = depTokenRe.exec(refMatch[1])) !== null) {
+          const normalized = normalizePhaseNumber(tok[0]);
+          if (normalized === ownNumber) continue; // #4764: never the row's own phase
+          if (seen.has(normalized)) continue;
+          seen.add(normalized);
+          depNums.push(tok[0]);
+        }
+      }
       phase['deps_satisfied'] = depNums.every((n) => completedNums.has(normalizePhaseNumber(n)));
       phase['dep_phases'] = depNums;
     }
diff --git a/src/phase-id.cts b/src/phase-id.cts
index 8df56cd67..ab511c603 100644
--- a/src/phase-id.cts
+++ b/src/phase-id.cts
@@ -64,6 +64,22 @@ const OPTIONAL_PHASE_TAG_SOURCE = '(?:\\s*\\([^)\\n]{0,200}\\))?';
 // introduced outside this module without a `// phase-id-owner:` justification.
 const PHASE_NUMBER_TOKEN_SOURCE = '\\d+[A-Z]?(?:\\.\\d+)*';
 
+// #4764: a phase REFERENCE in depends-on PROSE — the token in context, directly
+// following "Phase"/"Phases", with bare-token list continuation ("Phases 1 and 2",
+// "Phase 1, 2, and 3", "Phase 1-3"). Built HERE, beside the token grammar it is
+// anchored on, because two readers consume Depends-on prose (init.manager's
+// dep_phases and planning-inspect's dependencies) and re-deriving the anchored
+// form at each would drift exactly the way PHASE_NUMBER_TOKEN_SOURCE's own
+// anti-divergence rule exists to prevent. A bare digit run is the right shape
+// test for a phase id and the WRONG test for a reference: the whole-field
+// scrape this replaces pulled calendar dates ("2026-09-14" → 2026/09/14), git
+// shas ("8bf403100d" → 8b/403100d/…) and ledger ids (WINDOWS #1843) in as
+// dependencies. The `-` separator deliberately extracts range ENDPOINTS only
+// ("Phase 1-3" → 1, 3) — the pre-#4764 behavior; interior enumeration stays
+// out (a range's middle is not written as a reference).
+const PHASE_DEP_REF_SOURCE =
+  `\\bphases?\\s+(${PHASE_NUMBER_TOKEN_SOURCE}(?:(?:\\s*,\\s*(?:and\\s+)?|\\s+and\\s+|\\s*&\\s*|\\s+(?:to|through)\\s+|\\s*-\\s*)${PHASE_NUMBER_TOKEN_SOURCE})*)`;
+
 // #2528 review: the CASE-FLEXIBLE renderings of the two sources above, for call
 // sites that scan directory names (where a project code or a variant suffix may
 // legitimately be lowercase) and therefore cannot use a case-sensitive class.
@@ -1632,6 +1648,7 @@ export = {
   OPTIONAL_PROJECT_CODE_PREFIX_SOURCE,
   OPTIONAL_PHASE_TAG_SOURCE,
   PHASE_NUMBER_TOKEN_SOURCE,
+  PHASE_DEP_REF_SOURCE,
   CASE_FLEXIBLE_PROJECT_CODE_PREFIX_SOURCE,
   CASE_FLEXIBLE_PHASE_NUMBER_TOKEN_SOURCE,
   PHASE_CONTINUATION_SEGMENT_SOURCE,
diff --git a/src/planning-inspect.cts b/src/planning-inspect.cts
index 8935b077b..9a08ac1ca 100644
--- a/src/planning-inspect.cts
+++ b/src/planning-inspect.cts
@@ -1059,7 +1059,23 @@ function extractGoalProse(sectionBody: string): string | null {
 function extractDependencyTokens(sectionBody: string): string[] {
   const m = DEPENDS_ON_LINE_RE.exec(sectionBody);
   if (!m) return [];
-  return sortedUnique([...m[1].matchAll(/\d+(?:\.\d+)*/g)].map((t) => t[0]));
+  // #4764: phase REFERENCES, not digit runs — the same prose-anchored grammar
+  // init.manager's dep_phases extraction uses (owner: phase-id.cts's
+  // PHASE_DEP_REF_SOURCE). The whole-field token scrape this replaces pulled
+  // calendar dates, git shas and ledger ids in as dependencies. The grammar's
+  // capture group 1 already excludes the "Phase(s)" anchor word, so no
+  // prefix-strip literal is needed here. Self-exclusion (init.manager drops
+  // the row's own number) is deliberately NOT applied: this reader has no row
+  // context at the extraction site and reports informationally, it does not
+  // gate.
+  const refRe = new RegExp(phaseIdMod.PHASE_DEP_REF_SOURCE, 'gi');
+  const tokenRe = new RegExp(phaseIdMod.PHASE_NUMBER_TOKEN_SOURCE, 'g');
+  const tokens: string[] = [];
+  let refMatch: RegExpExecArray | null;
+  while ((refMatch = refRe.exec(m[1])) !== null) {
+    for (const t of refMatch[1].matchAll(tokenRe)) tokens.push(t[0]);
+  }
+  return sortedUnique(tokens);
 }
 
 /**
diff --git a/tests/init-manager.test.cjs b/tests/init-manager.test.cjs
index 49dc3d706..dad9b37be 100644
--- a/tests/init-manager.test.cjs
+++ b/tests/init-manager.test.cjs
@@ -1646,3 +1646,187 @@ describe('bug-3584: validate health uses formatter for codex runtime too', () =>
 });
   });
 }
+
+// ─── #4764: dep_phases extracts only in-context phase references ────────────
+
+describe('#4764 dep_phases extracts only Phase-prefixed references', () => {
+  let tmpDir;
+
+  beforeEach(() => {
+    tmpDir = createTempProject();
+  });
+
+  afterEach(() => {
+    cleanup(tmpDir);
+  });
+
+  // The issue's measured phase-663 prose: dates, a round name, ledger ids and
+  // the row's own number, plus one genuine "Phase 654" mention.
+  const ISSUE_663_PROSE =
+    'Nothing. Round 663-DISPOSITION Q7 (3/3, 2026-09-14, ' +
+    '`.planning/decisions/663-CROSSAI-stream-consumer-failure-disposition.md`) moved the ' +
+    '"for retry" javadoc correction into this phase and dropped the dependency on Phase 654; ' +
+    'the defect is already recorded (WINDOWS #1843, #1977-#1985, #1992).';
+
+  test("#4764: dep_phases extracts only Phase-prefixed references from the issue's measured prose", () => {
+    writeState(tmpDir);
+    writeRoadmap(tmpDir, [
+      { number: '663', name: 'Disposition phase', depends_on: ISSUE_663_PROSE },
+      { number: '654', name: 'Prior phase' },
+    ]);
+
+    const output = JSON.parse(runGsdTools('init manager', tmpDir).output);
+    const row = output.phases.find((p) => p.number === '663');
+
+    assert.deepEqual(row.dep_phases, ['654']);
+    for (const junk of ['663', '2026', '09', '14', '7', '3', '1843', '1977', '1985', '1992']) {
+      assert.ok(!row.dep_phases.includes(junk), `must not scrape ${junk} as a dependency`);
+    }
+    assert.strictEqual(row.deps_satisfied, false, '654 is incomplete — the only real reference');
+  });
+
+  test('#4764: a date-only depends_on prose yields no dependencies and a satisfied row', () => {
+    writeState(tmpDir);
+    writeRoadmap(tmpDir, [
+      { number: '674', name: 'Ops current', depends_on: 'Nothing. Opened by the owner on 2026-09-15 after asking whether the operations app was current with the backend.' },
+    ]);
+
+    const output = JSON.parse(runGsdTools('init manager', tmpDir).output);
+    const row = output.phases.find((p) => p.number === '674');
+
+    assert.deepEqual(row.dep_phases, []);
+    assert.strictEqual(row.deps_satisfied, true);
+    assert.strictEqual(row.deps_display, '—');
+  });
+
+  test('#4764: sha-bearing prose contributes no dependency tokens', () => {
+    writeState(tmpDir);
+    writeRoadmap(tmpDir, [
+      { number: '607', name: 'Sha phase', depends_on: 'Landed in 8bf403100d and 9445745c08; follow-up 76aea5c36f tracked separately.' },
+    ]);
+
+    const output = JSON.parse(runGsdTools('init manager', tmpDir).output);
+    const row = output.phases.find((p) => p.number === '607');
+
+    assert.deepEqual(row.dep_phases, []);
+    assert.strictEqual(row.deps_satisfied, true);
+  });
+
+  test('#4764: Phase lists stay fully extracted (and/comma separators)', () => {
+    writeState(tmpDir);
+    writeRoadmap(tmpDir, [
+      { number: '1', name: 'A', complete: true },
+      { number: '2', name: 'B', complete: true },
+      { number: '3', name: 'C', depends_on: 'Phases 1 and 2' },
+      { number: '4', name: 'D', depends_on: 'Phase 1, Phase 2' },
+      { number: '5', name: 'E', depends_on: 'Phase 1 & 2' },
+      { number: '6', name: 'F', depends_on: 'Phase 1 to 2' },
+    ]);
+    // Completion is DISK-strict (ADR-3180): a ticked checkbox does not feed
+    // completedNums — the dep targets need real dirs + passed verifications.
+    for (const n of [1, 2]) {
+      writePassedVerification(scaffoldPhase(tmpDir, n, { plans: 1, summaries: 1 }), `0${n}`);
+    }
+
+    const output = JSON.parse(runGsdTools('init manager', tmpDir).output);
+    assert.deepEqual(output.phases.find((p) => p.number === '3').dep_phases, ['1', '2']);
+    assert.deepEqual(output.phases.find((p) => p.number === '4').dep_phases, ['1', '2']);
+    assert.deepEqual(output.phases.find((p) => p.number === '5').dep_phases, ['1', '2'], '& separator');
+    assert.deepEqual(output.phases.find((p) => p.number === '6').dep_phases, ['1', '2'], 'to-range separator');
+    assert.strictEqual(output.phases.find((p) => p.number === '3').deps_satisfied, true);
+    assert.strictEqual(output.phases.find((p) => p.number === '5').deps_satisfied, true);
+    assert.strictEqual(output.phases.find((p) => p.number === '6').deps_satisfied, true);
+  });
+
+  test('#4764: Oxford-comma lists stay fully extracted, and a missing member blocks', () => {
+    // Adversarial-review fold-in: "Phases 1, 2, and 3" is the most common
+    // English enumeration; the first grammar cut dropped 603-style tail
+    // members, silently clearing a real blocker (the dangerous direction).
+    writeState(tmpDir);
+    writeRoadmap(tmpDir, [
+      { number: '1', name: 'A', complete: true },
+      { number: '2', name: 'B', complete: true },
+      { number: '3', name: 'C', depends_on: 'Phases 1, 2, and 3' },
+    ]);
+    for (const n of [1, 2]) {
+      writePassedVerification(scaffoldPhase(tmpDir, n, { plans: 1, summaries: 1 }), `0${n}`);
+    }
+
+    const output = JSON.parse(runGsdTools('init manager', tmpDir).output);
+    const row = output.phases.find((p) => p.number === '3');
+    assert.deepEqual(row.dep_phases, ['1', '2']);
+    assert.strictEqual(row.deps_satisfied, true);
+  });
+
+  test('#4764: hyphen ranges extract both endpoints and a missing tail blocks', () => {
+    // House style writes ranges ("Phases 1-9"); endpoints are the written
+    // references (the pre-#4764 scrape also kept only endpoints).
+    writeState(tmpDir);
+    writeRoadmap(tmpDir, [
+      { number: '1', name: 'A', complete: true },
+      { number: '3', name: 'C' },
+      { number: '4', name: 'D', depends_on: 'Phases 1-3' },
+    ]);
+    writePassedVerification(scaffoldPhase(tmpDir, 1, { plans: 1, summaries: 1 }), '01');
+
+    const output = JSON.parse(runGsdTools('init manager', tmpDir).output);
+    const row = output.phases.find((p) => p.number === '4');
+    assert.deepEqual(row.dep_phases, ['1', '3']);
+    assert.strictEqual(row.deps_satisfied, false, '3 is incomplete — a real blocker must not be dropped');
+  });
+
+  test('#4764 property: extraction pulls exactly the Phase-prefixed references out of arbitrary prose', () => {
+    // House fast-check config (seed 42); per-call numRuns caps the cost — each
+    // run spawns the real CLI (runGsdTools), unlike the pure-function properties.
+    const fc = require('./helpers/fast-check-setup.cjs');
+    // Junk fragments whose digit runs the old whole-field scrape pulled in as
+    // "dependencies": calendar dates, git shas, ledger ids, counts, round names.
+    // (fc v4 has no hexaString — stringMatching is the legal sha generator.)
+    const junkFragments = [
+      fc.nat({ max: 28 }).map((d) => `2026-09-${String(d).padStart(2, '0')}`),
+      fc.stringMatching(/[0-9a-f]{8,10}/),
+      fc.nat({ max: 99999 }).map((n) => `WINDOWS #${n}`),
+      fc.nat({ max: 999 }).map((n) => `#${n}-#${n + 1}`),
+      fc.nat({ max: 9 }).map((n) => `round ${n}-DISPOSITION (${n}/3`),
+    ];
+    const refFragment = fc.nat({ max: 8 }).map((n) => `Phase ${n + 1}`);
+    const listFragment = fc.nat({ max: 7 }).map((n) => `Phases ${n + 1}, ${n + 2}, and ${n + 3}`);
+    const selfFragment = fc.constant('Phase 9');
+
+    // The phase under test is 9: any "Phase 9" mention in its own prose must be
+    // dropped as a self-reference.
+    fc.assert(
+      fc.property(
+        fc.array(fc.oneof(...junkFragments, refFragment, listFragment, selfFragment), { maxLength: 12 }),
+        (fragments) => {
+          writeState(tmpDir);
+          writeRoadmap(tmpDir, [
+            { number: '9', name: 'Property phase', depends_on: fragments.join('; ') + '.' },
+            { number: '1', name: 'Ref target', complete: true },
+          ]);
+          const output = JSON.parse(runGsdTools('init manager', tmpDir).output);
+          const row = output.phases.find((p) => p.number === '9');
+          const extracted = row.dep_phases.map((d) => String(d));
+
+          // Independent spec mirror: the expected set is every phase token a
+          // fragment references ("Phase N" or inside a "Phases …" list),
+          // minus the row's own number — junk never contributes.
+          const expected = new Set();
+          for (const f of fragments) {
+            const refMatch = /^phases?\s+(.*)$/i.exec(f);
+            if (!refMatch) continue;
+            for (const tok of refMatch[1].match(/\d+(?:\.\d+)*/g) || []) {
+              if (tok !== '9') expected.add(tok);
+            }
+          }
+          assert.deepEqual(
+            [...extracted].sort(),
+            [...expected].sort(),
+            `extraction must pull exactly the Phase-prefixed references from: ${fragments.join('; ')}`,
+          );
+        },
+      ),
+      { numRuns: 20 },
+    );
+  });
+});