diff --git a/.github/workflows/governance-acceptance.yml b/.github/workflows/governance-acceptance.yml
index 5b84d16..c1e85c4 100644
--- a/.github/workflows/governance-acceptance.yml
+++ b/.github/workflows/governance-acceptance.yml
@@ -10,6 +10,10 @@ on:
         description: Same immutable central commit as the reusable workflow reference
         required: true
         type: string
+      require-codex-marker:
+        description: Fail this observation when the exact-head Codex prerequisite is absent
+        default: false
+        type: boolean
       mode:
         default: SHADOW
         type: string
@@ -31,6 +35,7 @@ jobs:
           PR_NUMBER: ${{ inputs.pr-number }}
           POLICY_REVISION: ${{ inputs.policy-revision }}
           MODE: ${{ inputs.mode }}
+          REQUIRE_CODEX_MARKER: ${{ inputs.require-codex-marker }}
         with:
           script: |
             // Trusted reusable workflow code only; never load candidate source.
@@ -40,10 +45,13 @@ jobs:
             const number = Number(process.env.PR_NUMBER);
             const fail = code => { throw new Error(code); };
             const hash = value => typeof value === 'string' && /^[0-9a-f]{40}$/.test(value);
+            const markerInput = process.env.REQUIRE_CODEX_MARKER;
+            const markerRequired = markerInput === 'true';
             let report = {schema_version: 1, mode: 'SHADOW', merge_authorized: false,
+              codex_prerequisite: markerRequired || !['true', 'false'].includes(markerInput) ? 'BLOCKED' : 'NOT_REQUIRED',
               observation: 'UNKNOWN', reason: 'READ_OR_INTEGRITY_FAILURE'};
             try {
-              if (process.env.MODE !== 'SHADOW' || !hash(revision)
+              if (!['true', 'false'].includes(markerInput) || process.env.MODE !== 'SHADOW' || !hash(revision)
                   || !Number.isSafeInteger(number) || number < 1) fail('INVALID_INPUT');
               const {data: content} = await github.rest.repos.getContent({owner: 'QuantStrategyLab',
                 repo: '.github', path: '.github/governance/policy.json', ref: revision});
@@ -197,6 +205,12 @@ jobs:
                 semantic_classification: 'HEURISTIC_NOT_ACCEPTANCE', required_checks: 'NOT_EVALUATED',
                 event_freshness: 'SNAPSHOT_ONLY', baseline_exception: 'UNSUPPORTED',
                 codex_observation: codexObservation(evidence.comments)};
+              if (markerRequired) {
+                report.codex_prerequisite = ['NO_REPORTED_FINDINGS', 'FINDINGS_REPORTED']
+                  .includes(report.codex_observation) ? 'SATISFIED' : 'BLOCKED';
+                if (report.codex_prerequisite !== 'SATISFIED')
+                  core.setFailed('Exact-head Codex review prerequisite is not satisfied');
+              }
             } catch {
               // Deliberately omit exception text, paths, patches, user bodies and API response data.
               report = {...report, observation: 'UNKNOWN', reason: 'READ_OR_INTEGRITY_FAILURE'};
diff --git a/README.md b/README.md
index 703b654..3316d5e 100644
--- a/README.md
+++ b/README.md
@@ -23,3 +23,5 @@ Organization-wide files for QuantStrategyLab repositories.
 Open a pull request. Describe the problem, the change, and the effect on existing repositories. AI assistants and maintainers follow [AGENTS.md](AGENTS.md).
 
 - [Governance shadow engine](docs/governance-shadow.md) — reusable report-only review observations.
+
+- [Collaboration governance](docs/collaboration-governance.md)
diff --git a/README.zh-CN.md b/README.zh-CN.md
index e632add..a1ec362 100644
--- a/README.zh-CN.md
+++ b/README.zh-CN.md
@@ -23,3 +23,5 @@ QuantStrategyLab 各仓库共用的组织级文件。
 提交 pull request，说明要解决的问题、改了什么，以及对现有仓库的影响。AI 助手与维护者遵循 [AGENTS.md](AGENTS.md)。
 
 - [Governance shadow engine](docs/governance-shadow.md) — reusable report-only review observations.
+
+- [协作规则草案](docs/collaboration-governance.md)
diff --git a/docs/collaboration-governance.md b/docs/collaboration-governance.md
new file mode 100644
index 0000000..f662218
--- /dev/null
+++ b/docs/collaboration-governance.md
@@ -0,0 +1,69 @@
+# Collaboration governance
+
+## Merge prerequisites
+
+The dedicated review-and-merge bot owns PR acceptance and merge disposition.
+Before any merge it reads the OPEN PR, full current head and base, changed scope,
+native required checks and applicable owner authorization. It also requires a
+trusted Codex App comment containing exactly one `codex-review-state:v1` record.
+The `code_review.commit_sha` must equal the current full PR head,
+`status` must be `completed`, and `has_unresolved_findings` must be `false`.
+Both finding flags must be booleans. Findings previously reported and resolved
+still require the bot's substantive disposition. Labels, descriptions and a
+worker's own receipt cannot substitute for authenticated review evidence.
+
+Missing, stale, malformed, ambiguous, running or unavailable evidence blocks
+merge. Re-read the candidate and evidence immediately before disposition;
+changes invalidate the previous snapshot. A head-bound marker does not prove
+base or scope acceptance. Independent review, native gates and owner approval
+remain separate requirements. Never bypass protection to satisfy this policy.
+
+Review happens before merge. Never request a new review on a merged PR,
+backfill a marker, or treat a later review as retrospective merge permission.
+A historical missing-review exception stays recorded. Preserve existing
+findings and evidence; an authorized correction uses a new OPEN PR with its
+own review. A clean historical report does not prove deployment or runtime.
+
+## Progress and daily brief
+
+PR state is the source for source-delivery progress. Read current state, Draft
+flag, full head/base, marker binding and required checks from the PR. Describe
+OPEN/Draft as work under review, CLOSED without merge as closed, and MERGED as
+source integrated. In the same record state whether its exact head has a
+completed marker and unresolved findings. Unknown evidence remains unknown.
+
+Issue #36 becomes a daily brief derived from those readbacks. It is not a
+second mutable progress authority. Each entry links its PR where public and
+records its observation time, owner, blocker and exact resume condition.
+Preserve unresolved production incidents until production readback proves a
+fix or the owner explicitly closes them. Source integration cannot close an
+incident, PAPER qualification or runtime acceptance by itself. Private work is
+shown only as PAS (private), with no private repository link or identity.
+
+Issue #38 is retired as an instruction and progress channel. Historical content
+remains historical; it cannot launch work or authorize an action. Use the
+existing coordinator and typed callback receipts for authorized handoffs.
+Daily brief describes cadence only: this proposal installs no timer, schedule,
+issue editor or external message sender.
+
+## Workflow prerequisite proposal
+
+The reusable [observer](../.github/workflows/governance-acceptance.yml) accepts
+`require-codex-marker: true` as an opt-in check. It fails unless authenticated,
+current-head completed evidence has no unresolved findings. The existing pinned
+policy supplies trusted App identities; callers cannot supply their own.
+Read or integrity failures fail the check and publish only fixed diagnostic
+fields. Two reads reject candidate or evidence drift. Each run is stateless.
+
+The default remains `false` for existing shadow callers. The workflow always
+returns `merge_authorized: false`; a successful prerequisite is necessary
+evidence, never merge authority. This proposal changes no caller, trusted
+identity, ruleset, branch protection or required-check list. Actual enforcement
+requires a separately authorized integration with the dedicated bot and native
+controls. Pin both the reusable workflow and policy to the same reviewed
+immutable revision before an authorized caller adopts the prerequisite.
+
+The [skill candidate](skills/quant-collaboration/SKILL.md) contains the matching
+operator guidance. Repository presence does not install or update a host skill.
+The [legacy exception queue](legacy-governance-exceptions.md) preserves the bounded
+historical evidence request without requesting reviews on merged PRs.
diff --git a/docs/governance-shadow.md b/docs/governance-shadow.md
index 30db82e..b281a12 100644
--- a/docs/governance-shadow.md
+++ b/docs/governance-shadow.md
@@ -81,3 +81,11 @@ Run `node --test scripts/test-governance-shadow.cjs`, the existing Python checke
 and secret-scan tests, and `python scripts/check_style.py --phase 2`. The new
 `Governance shadow tests` workflow exercises the exact inline reusable code with
 fake API responses. Offline tests do not prove hosted execution or adoption.
+
+## Optional exact-head prerequisite
+
+[Collaboration governance](collaboration-governance.md) defines the proposed
+`require-codex-marker: true` input. It fails the existing observation job when
+trusted completed current-head evidence is absent or unresolved. Existing
+callers keep the default `false`. Passing the opt-in check never authorizes
+merge; no caller or required-check setting is changed by this source proposal.
diff --git a/docs/legacy-governance-exceptions.md b/docs/legacy-governance-exceptions.md
new file mode 100644
index 0000000..8e8b2b0
--- /dev/null
+++ b/docs/legacy-governance-exceptions.md
@@ -0,0 +1,46 @@
+# Legacy missing-review exception queue
+
+This retained governance exception record tracks the requested eleven-item
+historical batch. It does not authorize new reviews on merged PRs, backfilled
+markers, corrective implementation or external actions. The first four
+references and their order are supplied by the bounded task. The source that
+identifies the remaining seven references is missing, so those slots remain
+unassigned. This is an incomplete batch, not a claim of eleven verified PRs.
+
+## Priority and evidence
+
+The required first-four order is preserved. Production impact determines the
+remaining order once their identities are available; do not guess candidates
+from a larger search result. Each named item is MERGED with no matching
+current-head marker observed in the readback. This says nothing about whether
+an earlier independent review occurred or whether the change was deployed.
+
+| Priority | Reference | Production impact | Evidence disposition | Next condition |
+| --- | --- | --- | --- | --- |
+| 1 | InteractiveBrokersPlatform #962 | Heartbeat diagnostics and notification policy | Missing current-head marker; runtime policy fix unproved | Retrieve existing evidence; preserve incident until authentic runtime readback |
+| 2 | LongBridgePlatform #702 | PAPER identity validation and target selection | Missing current-head marker; historical report has a P2 target-selection finding | Preserve finding and existing repair ownership |
+| 3 | QuantPlatformKit #771 | Shared setup contract and release bindings | Missing current-head marker; historical report has two P2 findings | Retain contract/lockfile findings and existing successor evidence |
+| 4 | QuantAIService #462 | Monitor configuration and operational diagnosis | Missing current-head marker; historical report is CLEAN, runtime configuration still unproved | Retain clean report separately from the open production incident |
+| 5 | Unassigned | Unknown | Original batch reference unavailable | Original batch identity source required |
+| 6 | Unassigned | Unknown | Original batch reference unavailable | Original batch identity source required |
+| 7 | Unassigned | Unknown | Original batch reference unavailable | Original batch identity source required |
+| 8 | Unassigned | Unknown | Original batch reference unavailable | Original batch identity source required |
+| 9 | Unassigned | Unknown | Original batch reference unavailable | Original batch identity source required |
+| 10 | Unassigned | Unknown | Original batch reference unavailable | Original batch identity source required |
+| 11 | Unassigned | Unknown | Original batch reference unavailable | Original batch identity source required |
+
+## Batch handling contract
+
+The established review-and-merge bot owns historical evidence disposition.
+Retrieve already-existing reports and bind them to the original head and merge
+commit; retain missing evidence explicitly. This queue requests evidence
+reconciliation only. It does not dispatch review-after-merge. Concrete known
+findings may inform a separately authorized new OPEN corrective PR, which must
+pass current pre-merge review and gates. Never rewrite historical authorization.
+
+A completed queue record needs the authenticated original eleven references,
+production-impact order after the first four, current source state, original
+head/merge bindings in approved evidence storage, existing findings, owner,
+blocker and resume condition. Public records show any private reference only
+as PAS (private), without links or private identity. No private evidence or
+operational resource values belong in this document.
diff --git a/docs/skills/quant-collaboration/SKILL.md b/docs/skills/quant-collaboration/SKILL.md
new file mode 100644
index 0000000..6fbb09e
--- /dev/null
+++ b/docs/skills/quant-collaboration/SKILL.md
@@ -0,0 +1,28 @@
+---
+name: quant-collaboration
+description: Apply PR review prerequisites and daily-brief rules when coordinating QuantStrategyLab PR delivery. Does not authorize dispatch, merge or production changes.
+---
+# Quant collaboration
+
+Follow the repository [AGENTS.md](../../../AGENTS.md) and
+[collaboration governance](../../collaboration-governance.md). This is a skill
+candidate for review; publishing this file does not install it on a host.
+
+- Use the current PR state, Draft flag, head/base and checks as source progress.
+  Issue #36 is a daily derived brief; issue #38 is retired.
+- The dedicated review-and-merge bot owns acceptance. Before merge require an
+  authenticated `codex-review-state:v1` marker with `code_review.commit_sha`
+  equal to the current head, `status: completed`, boolean finding flags and
+  `has_unresolved_findings: false`. Re-read state before disposition. Marker
+  completion alone does not satisfy base/scope review, CI or owner approval.
+- Unknown, stale, absent or malformed evidence blocks merge. Never replace
+  evidence with a label, worker report or edited PR description.
+- Review before merge. Never request review on merged PRs or backfill markers.
+  Retain historical exceptions and findings; corrections need an authorized
+  new OPEN PR and its own review.
+- Keep runtime, PAPER and production incident evidence separate from PR state.
+  Preserve unresolved production incidents in the daily brief. Display private
+  work only as PAS (private), without private links or repository identities.
+- Use only existing authorized handoff paths. Distinguish dispatched, received
+  and acknowledged receipts. This skill creates no scheduler, dispatch rights,
+  configuration changes, merge rights or production authority.
diff --git a/scripts/test-governance-shadow.cjs b/scripts/test-governance-shadow.cjs
index dbce45d..950a722 100644
--- a/scripts/test-governance-shadow.cjs
+++ b/scripts/test-governance-shadow.cjs
@@ -29,7 +29,7 @@ async function observe(f = fixture()) {
   const core = {setFailed: () => {state.failed = true;}, setOutput: () => {}, summary: {
     addHeading() {return this;}, addCodeBlock(s) {state.summary = s; return this;}, async write() {}}};
   const report = await execute(github, {repo: {owner: '<PROJECT_ID>', repo: 'sample'}}, core, require,
-    {env: {MODE: f.mode || 'SHADOW', POLICY_REVISION: R, PR_NUMBER: '1'}});
+    {env: {MODE: f.mode || 'SHADOW', POLICY_REVISION: R, PR_NUMBER: '1', REQUIRE_CODEX_MARKER: f.requireMarker === undefined ? 'false' : f.requireMarker}});
   assert.equal(report.merge_authorized, false);
   assert.ok(!state.summary.includes('private token'));
   return {report, state};
@@ -230,3 +230,51 @@ test('Optional style advice does not require repair', async () => {
   record.findings = [{kind: 'advisory'}]; f.evidence.reviews[0].body = JSON.stringify(record);
   assert.equal((await observe(f)).report.observation, 'PENDING_OWNER');
 });
+
+// The optional prerequisite is necessary evidence only, never merge authority.
+test('Required Codex prerequisite fails closed when unconfigured', async () => {
+  const f = fixture(); f.requireMarker = 'true';
+  const {report, state} = await observe(f);
+  assert.equal(state.failed, true);
+  assert.equal(report.codex_prerequisite, 'BLOCKED');
+});
+test('Required Codex prerequisite accepts an authenticated completed exact-head marker', async () => {
+  const {f} = codexFixture(); f.requireMarker = 'true';
+  const {report, state} = await observe(f);
+  assert.equal(state.failed, false);
+  assert.equal(report.codex_prerequisite, 'SATISFIED');
+  assert.equal(report.merge_authorized, false);
+});
+for (const mutation of ['missing', 'untrusted', 'stale', 'running', 'unresolved', 'malformed', 'duplicate', 'merged', 'base-drift', 'head-drift', 'invalid-input', 'api']) {
+  test(`Required marker blocks unsafe evidence: ${mutation}`, async () => {
+    const {f, entry, record} = codexFixture(); f.requireMarker = 'true';
+    if (mutation === 'missing') f.evidence.comments = [];
+    if (mutation === 'untrusted') entry.performed_via_github_app.id = 54321;
+    if (mutation === 'stale') record.code_review.commit_sha = 'd'.repeat(40);
+    if (mutation === 'running') record.code_review.status = 'running';
+    if (mutation === 'unresolved') record.code_review.has_unresolved_findings = true;
+    entry.body = mutation === 'malformed' ? '<!-- codex-review-state:v1 {bad} -->'
+      : `<!-- codex-review-state:v1 ${JSON.stringify(record)} -->`;
+    if (mutation === 'duplicate') entry.body += entry.body;
+    if (mutation === 'merged') f.pr.state = 'closed';
+    if (mutation.endsWith('-drift')) {
+      f.current = structuredClone(f.pr);
+      f.current[mutation === 'base-drift' ? 'base' : 'head'].sha = 'd'.repeat(40);
+    }
+    if (mutation === 'invalid-input') f.requireMarker = 'unknown';
+    if (mutation === 'api') f.apiError = true;
+    const {report, state} = await observe(f);
+    assert.equal(state.failed, true);
+    assert.equal(report.codex_prerequisite, 'BLOCKED');
+    assert.ok(!state.summary.includes('private token'));
+  });
+}
+test('Parallel observations keep no acceptance cache across a moved head', async () => {
+  const {f} = codexFixture(); f.requireMarker = 'true';
+  const moved = structuredClone(f); moved.current = structuredClone(f.pr);
+  moved.current.head.sha = 'd'.repeat(40);
+  const [clean, stale] = await Promise.all([observe(f), observe(moved)]);
+  assert.equal(clean.report.codex_prerequisite, 'SATISFIED');
+  assert.equal(stale.report.codex_prerequisite, 'BLOCKED');
+  assert.equal(stale.state.failed, true);
+});