diff --git a/packages/gittensory-engine/src/index.ts b/packages/gittensory-engine/src/index.ts
index 880938724b..f6800de26f 100644
--- a/packages/gittensory-engine/src/index.ts
+++ b/packages/gittensory-engine/src/index.ts
@@ -217,6 +217,17 @@ export {
   type SelfReviewSlopInput,
   type SelfReviewVerdict,
 } from "./miner/self-review-adapter.js";
+export {
+  decideNextAction,
+  decideNextActionWithReason,
+  deriveSelfReviewOutcome,
+  type AbandonReason,
+  type HandoffPacket,
+  type IterateLoopAction,
+  type IterateLoopDecision,
+  type IterationState,
+  type SelfReviewOutcome,
+} from "./miner/iterate-policy.js";
 export {
   codingAgentModeExecutes,
   isGlobalMinerCodingAgentPause,
diff --git a/packages/gittensory-engine/src/miner/iterate-policy.ts b/packages/gittensory-engine/src/miner/iterate-policy.ts
new file mode 100644
index 0000000000..d3dab777b2
--- /dev/null
+++ b/packages/gittensory-engine/src/miner/iterate-policy.ts
@@ -0,0 +1,159 @@
+// Iterate-loop stop/abandon/handoff policy (#2335): the explicit POLICY the orchestration loop's control flow
+// (#2333, sibling issue) consults each iteration to decide among exactly three outcomes. Deliberately split
+// from the loop MECHANICS (#2333) so the actual thresholds/rules are one small, individually-reviewable, pure
+// artifact -- `decideNextAction` needs no driver, no worktree, no IO to test.
+//
+// STRATEGIC CONSTRAINTS this policy encodes (gittensory-miner-autonomy-roadmap):
+//   - "never auto-submit (P4) before governor+caps (P5)" -- a mandatory clean predicted-gate PASS is the ONLY
+//     path to `"handoff"`; an ambiguous or errored self-review downgrades to abandon, never optimistically
+//     hands off.
+//   - "disengage SILENTLY on rejection" (the Matplotlib "MJ Rathbun" cautionary tale) -- a rejection signal
+//     wins over EVERYTHING else, including a self-review that would otherwise pass. Continuing to submit to a
+//     repo that has already shown it does not want automated contributions is the exact anti-pattern this
+//     guards against, regardless of how good any individual attempt looks.
+//   - "reward MERGED net-positive (never submission volume)" -- the no-progress detector and iteration ceiling
+//     exist so a stuck loop stops wasting turns chasing a submission that was never going to land, rather than
+//     grinding toward *a* submission for its own sake.
+//
+// AUTONOMY DIAL (not yet wired): `src/settings/autonomy.ts`'s `resolveAutonomy`/`isActingAutonomyLevel` is the
+// existing reusable deny-by-default pattern this policy's eventual live autonomy-level check should consult
+// once `.gittensory-miner.yml` defines autonomy fields -- that wiring is explicitly left to a later phase; this
+// module's `decideNextAction` is autonomy-level-agnostic today.
+
+import type { SelfReviewVerdict } from "./self-review-adapter.js";
+
+/** The three outcomes `decideNextAction` may reach. */
+export type IterateLoopAction = "continue" | "handoff" | "abandon";
+
+/** Every distinct reason `decideNextAction` can abandon for -- kept as a closed literal union so a caller
+ *  recording the decision (the attempt-log primitive, per #2333) has a stable, exhaustive vocabulary. */
+export type AbandonReason = "rejection_signaled" | "self_review_ambiguous" | "max_iterations_reached" | "no_progress";
+
+/**
+ * The self-review outcome as the policy needs it -- narrower than the full {@link SelfReviewVerdict} (self-
+ * review-adapter.ts, #2334) so `IterationState` stays a minimal, cheap-to-construct synthetic fixture in
+ * tests. `"ambiguous"` is NOT something {@link deriveSelfReviewOutcome} ever produces from a successfully
+ * returned verdict (a real verdict object always has a definite conclusion) -- it is constructed directly by
+ * the caller's own error handling when the self-review call itself throws (e.g. a calculator inside it errors),
+ * per #2333's own "self-review itself errors" abandon trigger.
+ */
+export type SelfReviewOutcome =
+  | { readonly kind: "pass" }
+  | { readonly kind: "fail"; readonly blockerCodes: readonly string[] }
+  | { readonly kind: "ambiguous"; readonly reason?: string | undefined };
+
+/** Derive the policy-relevant {@link SelfReviewOutcome} from a real, successfully computed
+ *  {@link SelfReviewVerdict}. Only ever returns `"pass"` or `"fail"` -- see that variant's own doc comment for
+ *  why `"ambiguous"` is constructed elsewhere. */
+export function deriveSelfReviewOutcome(verdict: SelfReviewVerdict): SelfReviewOutcome {
+  if (verdict.passesPredictedGate) return { kind: "pass" };
+  return { kind: "fail", blockerCodes: verdict.predictedGateVerdict.blockers.map((blocker) => blocker.code) };
+}
+
+/**
+ * Everything `decideNextAction` needs for one iteration's decision. Deliberately minimal and synthetic-
+ * fixture-friendly -- no driver, no worktree, no IO.
+ */
+export type IterationState = {
+  /** 1-indexed count of iterations attempted so far, INCLUDING this one. */
+  iterationNumber: number;
+  /** Hard ceiling enforced INSIDE this policy (#2333's own deliverable: not left to an external caller to
+   *  remember to enforce). `iterationNumber >= maxIterations` abandons regardless of self-review outcome. */
+  maxIterations: number;
+  selfReview: SelfReviewOutcome;
+  /** The prior iteration's `fail` blocker codes, for the no-progress detector -- `null` when there is no prior
+   *  iteration to compare (the first iteration, or the prior iteration did not reach a `fail` outcome). */
+  previousBlockerCodes: readonly string[] | null;
+  /** True when the target repo (or this contributor's history with it) has signaled it does not want
+   *  automated/AI-authored contributions -- an explicit AI-usage-policy ban, or a prior submission from this
+   *  same miner was closed/rejected on this exact repo. The caller resolves this (e.g. via the AI-policy-map
+   *  signals or the rejection-state-machine primitive already shipped in `packages/gittensory-miner/lib/`) and
+   *  passes it in; this policy does not compute it itself. */
+  rejectionSignaled: boolean;
+};
+
+/** Forward-looking INTERFACE for Phase 4 (submission), not an implementation of it -- Phase 4 lands as a later,
+ *  separate issue. Gives it a stable target instead of reverse-engineering the shape from the loop's internals. */
+export type HandoffPacket = {
+  /** Absolute path to (or a branch ref identifying) the worktree holding the passing attempt's changes. */
+  worktreePath: string;
+  branchRef?: string | undefined;
+  /** Human-readable summary of the final diff, for the submission's own PR description. */
+  diffSummary: string;
+  /** The PASSING self-review verdict that authorized this handoff -- always has `passesPredictedGate: true`
+   *  (constructing a packet from anything else is a caller bug, not something this type can prevent statically,
+   *  since `decideNextAction` is the actual enforcement point). */
+  selfReviewVerdict: SelfReviewVerdict;
+  /** Reference into the attempt-log primitive (`packages/gittensory-engine/src/miner/attempt-log.ts`) for this
+   *  attempt's full decision trail. */
+  attemptLogReference: string;
+};
+
+export type IterateLoopDecision = {
+  action: IterateLoopAction;
+  /** Machine-stable, human-readable reason -- always populated, including for `"continue"` and `"handoff"`, so
+   *  every decision (not just abandons) has an auditable reason string for the attempt-log. */
+  reason: string;
+  /** Populated only when `action === "abandon"`. */
+  abandonReason?: AbandonReason | undefined;
+};
+
+function blockerSetsEqual(current: readonly string[], previous: readonly string[]): boolean {
+  if (current.length !== previous.length) return false;
+  const currentSet = new Set(current);
+  const previousSet = new Set(previous);
+  if (currentSet.size !== previousSet.size) return false;
+  for (const code of currentSet) if (!previousSet.has(code)) return false;
+  return true;
+}
+
+/**
+ * Decide the next action for one iteration. Pure; identical inputs always yield the identical decision.
+ *
+ * Precedence (each check short-circuits the ones below it):
+ * 1. `rejectionSignaled` -- ALWAYS abandons, even over an otherwise-passing self-review (disengage silently).
+ * 2. `selfReview.kind === "ambiguous"` -- abandons; never optimistically continues or hands off on ambiguity.
+ * 3. `selfReview.kind === "pass"` -- the ONLY path to `"handoff"`.
+ * 4. `iterationNumber >= maxIterations` -- abandons at the hard ceiling regardless of whether the blocker set
+ *    was still changing (genuine incremental progress does not buy unlimited iterations).
+ * 5. The current `fail` blocker set is identical to `previousBlockerCodes` -- abandons (no progress, stop
+ *    wasting turns).
+ * 6. Otherwise -- continue.
+ */
+export function decideNextActionWithReason(state: IterationState): IterateLoopDecision {
+  if (state.rejectionSignaled) {
+    return { action: "abandon", abandonReason: "rejection_signaled", reason: "Repo or contributor has signaled it does not want automated contributions; disengaging silently rather than retry-hammering." };
+  }
+  if (state.selfReview.kind === "ambiguous") {
+    return {
+      action: "abandon",
+      abandonReason: "self_review_ambiguous",
+      reason: `Self-review could not conclusively determine pass/fail${state.selfReview.reason ? `: ${state.selfReview.reason}` : "."} Downgrading to abandon rather than optimistically handing off.`,
+    };
+  }
+  if (state.selfReview.kind === "pass") {
+    return { action: "handoff", reason: "Self-review reached a clean predicted-gate pass." };
+  }
+  if (state.iterationNumber >= state.maxIterations) {
+    return {
+      action: "abandon",
+      abandonReason: "max_iterations_reached",
+      reason: `Reached the iteration ceiling (${state.maxIterations}) without a clean predicted-gate pass.`,
+    };
+  }
+  if (state.previousBlockerCodes !== null && blockerSetsEqual(state.selfReview.blockerCodes, state.previousBlockerCodes)) {
+    return {
+      action: "abandon",
+      abandonReason: "no_progress",
+      reason: `Blocker set unchanged from the prior iteration (${state.selfReview.blockerCodes.join(", ") || "no blockers listed"}); stopping rather than repeating an attempt that is not converging.`,
+    };
+  }
+  return { action: "continue", reason: "Self-review still failing but the blocker set changed since the prior iteration; continuing." };
+}
+
+/** The bare `decideNextAction(state) -> "continue" | "handoff" | "abandon"` signature this issue's deliverable
+ *  calls for. For the WHY behind a decision (the attempt-log needs a reason string, not just the action), use
+ *  {@link decideNextActionWithReason} -- this is a thin projection over the same logic. */
+export function decideNextAction(state: IterationState): IterateLoopAction {
+  return decideNextActionWithReason(state).action;
+}
diff --git a/packages/gittensory-engine/test/iterate-policy.test.ts b/packages/gittensory-engine/test/iterate-policy.test.ts
new file mode 100644
index 0000000000..8b79eab7a7
--- /dev/null
+++ b/packages/gittensory-engine/test/iterate-policy.test.ts
@@ -0,0 +1,168 @@
+import assert from "node:assert/strict";
+import { test } from "node:test";
+
+import {
+  decideNextAction,
+  decideNextActionWithReason,
+  deriveSelfReviewOutcome,
+  type IterationState,
+  type SelfReviewVerdict,
+} from "../dist/index.js";
+
+function baseState(overrides: Partial<IterationState> = {}): IterationState {
+  return {
+    iterationNumber: 1,
+    maxIterations: 5,
+    selfReview: { kind: "fail", blockerCodes: ["missing_linked_issue"] },
+    previousBlockerCodes: null,
+    rejectionSignaled: false,
+    ...overrides,
+  };
+}
+
+test("barrel: the public entrypoint re-exports the iterate-loop policy (#2335)", () => {
+  assert.equal(typeof decideNextAction, "function");
+  assert.equal(typeof decideNextActionWithReason, "function");
+  assert.equal(typeof deriveSelfReviewOutcome, "function");
+});
+
+test("continue: a still-failing self-review whose blocker set changed since the prior iteration", () => {
+  const state = baseState({
+    selfReview: { kind: "fail", blockerCodes: ["missing_test_evidence"] },
+    previousBlockerCodes: ["missing_linked_issue"],
+  });
+  assert.equal(decideNextAction(state), "continue");
+  const decision = decideNextActionWithReason(state);
+  assert.equal(decision.action, "continue");
+  assert.ok(decision.reason.length > 0, "even a continue decision must carry a populated reason string");
+  assert.equal(decision.abandonReason, undefined);
+});
+
+test("continue: a failing first iteration with no prior blocker set to compare against", () => {
+  const state = baseState({ previousBlockerCodes: null });
+  assert.equal(decideNextAction(state), "continue");
+});
+
+test("handoff: the ONLY path is a clean self-review pass", () => {
+  const state = baseState({ selfReview: { kind: "pass" } });
+  const decision = decideNextActionWithReason(state);
+  assert.equal(decision.action, "handoff");
+  assert.ok(decision.reason.length > 0);
+});
+
+test("abandon (rejection_signaled): wins over EVERYTHING else, including an otherwise-passing self-review", () => {
+  const state = baseState({ selfReview: { kind: "pass" }, rejectionSignaled: true });
+  const decision = decideNextActionWithReason(state);
+  assert.equal(decision.action, "abandon");
+  assert.equal(decision.abandonReason, "rejection_signaled");
+});
+
+test("abandon (self_review_ambiguous): never optimistically continues or hands off on ambiguity", () => {
+  const state = baseState({ selfReview: { kind: "ambiguous", reason: "predicted-gate calculator threw" } });
+  const decision = decideNextActionWithReason(state);
+  assert.equal(decision.action, "abandon");
+  assert.equal(decision.abandonReason, "self_review_ambiguous");
+  assert.match(decision.reason, /predicted-gate calculator threw/);
+});
+
+test("abandon (self_review_ambiguous): the reason is optional and the decision still abandons without it", () => {
+  const decision = decideNextActionWithReason(baseState({ selfReview: { kind: "ambiguous" } }));
+  assert.equal(decision.action, "abandon");
+  assert.equal(decision.abandonReason, "self_review_ambiguous");
+});
+
+test("abandon (max_iterations_reached): the hard ceiling stops the loop even if the blocker set is still changing", () => {
+  const state = baseState({
+    iterationNumber: 5,
+    maxIterations: 5,
+    selfReview: { kind: "fail", blockerCodes: ["a_brand_new_blocker_never_seen_before"] },
+    previousBlockerCodes: ["missing_linked_issue"],
+  });
+  const decision = decideNextActionWithReason(state);
+  assert.equal(decision.action, "abandon");
+  assert.equal(decision.abandonReason, "max_iterations_reached");
+});
+
+test("abandon (max_iterations_reached): fires at or beyond the ceiling, not only exactly at it", () => {
+  assert.equal(decideNextAction(baseState({ iterationNumber: 6, maxIterations: 5 })), "abandon");
+});
+
+test("continue: one iteration below the ceiling still continues", () => {
+  const state = baseState({ iterationNumber: 4, maxIterations: 5, previousBlockerCodes: ["something_else"] });
+  assert.equal(decideNextAction(state), "continue");
+});
+
+test("abandon (no_progress): an identical blocker set to the prior iteration stops wasting turns", () => {
+  const state = baseState({
+    iterationNumber: 2,
+    maxIterations: 10,
+    selfReview: { kind: "fail", blockerCodes: ["missing_linked_issue", "missing_test_evidence"] },
+    previousBlockerCodes: ["missing_linked_issue", "missing_test_evidence"],
+  });
+  const decision = decideNextActionWithReason(state);
+  assert.equal(decision.action, "abandon");
+  assert.equal(decision.abandonReason, "no_progress");
+});
+
+test("abandon (no_progress): the comparison is a SET, not an ordered array -- reordered-but-identical blockers still count as no progress", () => {
+  const state = baseState({
+    iterationNumber: 2,
+    maxIterations: 10,
+    selfReview: { kind: "fail", blockerCodes: ["b_code", "a_code"] },
+    previousBlockerCodes: ["a_code", "b_code"],
+  });
+  assert.equal(decideNextAction(state), "abandon");
+});
+
+test("continue: a duplicate blocker code does not falsely widen the set and mask real progress", () => {
+  // ["a","a"] and ["a","b"] must NOT compare equal as sets even though naive array-length comparison alone
+  // could be fooled by the duplicate.
+  const state = baseState({
+    iterationNumber: 2,
+    maxIterations: 10,
+    selfReview: { kind: "fail", blockerCodes: ["a_code", "a_code"] },
+    previousBlockerCodes: ["a_code", "b_code"],
+  });
+  assert.equal(decideNextAction(state), "continue");
+});
+
+test("continue: a different-length blocker set is never confused with no-progress (the fast-path length check)", () => {
+  const state = baseState({
+    iterationNumber: 2,
+    maxIterations: 10,
+    selfReview: { kind: "fail", blockerCodes: ["a_code", "b_code", "c_code"] },
+    previousBlockerCodes: ["a_code"],
+  });
+  assert.equal(decideNextAction(state), "continue");
+});
+
+test("abandon (no_progress): an empty blocker set unchanged from the prior (also empty) iteration still reads as no progress", () => {
+  const state = baseState({
+    iterationNumber: 2,
+    maxIterations: 10,
+    selfReview: { kind: "fail", blockerCodes: [] },
+    previousBlockerCodes: [],
+  });
+  const decision = decideNextActionWithReason(state);
+  assert.equal(decision.action, "abandon");
+  assert.equal(decision.abandonReason, "no_progress");
+  assert.match(decision.reason, /no blockers listed/);
+});
+
+test("deriveSelfReviewOutcome: a passing verdict maps to pass with no blocker codes", () => {
+  const verdict = { passesPredictedGate: true, predictedGateVerdict: { blockers: [] } } as unknown as SelfReviewVerdict;
+  assert.deepEqual(deriveSelfReviewOutcome(verdict), { kind: "pass" });
+});
+
+test("deriveSelfReviewOutcome: a failing verdict maps to fail with the real blocker codes extracted", () => {
+  const verdict = {
+    passesPredictedGate: false,
+    predictedGateVerdict: {
+      blockers: [
+        { code: "duplicate_pr_risk", title: "t1", detail: "d1" },
+        { code: "missing_linked_issue", title: "t2", detail: "d2" },
+      ],
+    },
+  } as unknown as SelfReviewVerdict;
+  assert.deepEqual(deriveSelfReviewOutcome(verdict), { kind: "fail", blockerCodes: ["duplicate_pr_risk", "missing_linked_issue"] });
+});