diff --git a/packages/gittensory-miner/lib/replay-task-generation.d.ts b/packages/gittensory-miner/lib/replay-task-generation.d.ts
index ced38e3253..02cb3941b4 100644
--- a/packages/gittensory-miner/lib/replay-task-generation.d.ts
+++ b/packages/gittensory-miner/lib/replay-task-generation.d.ts
@@ -72,10 +72,12 @@ export type ReplayTask = {
     commitT: string | null;
     contextTexts: string[];
   };
-  revealed: {
-    commitCount: number;
-    groundTruth: unknown;
-  };
+};
+
+export type ReplayScoringKey = {
+  eligible: true;
+  commitCount: number;
+  groundTruth: unknown;
 };
 
 export function detectForwardReferences(
@@ -108,3 +110,8 @@ export function generateReplayTask(
   context: ForwardRefContext | null | undefined,
   options: ReplayTaskOptions | null | undefined,
 ): ReplayTask | ReplayTaskRejected;
+
+export function generateReplayScoringKey(
+  candidate: FreezePointCandidate | null | undefined,
+  options: ReplayTaskOptions | null | undefined,
+): ReplayScoringKey | ReplayTaskRejected;
diff --git a/packages/gittensory-miner/lib/replay-task-generation.js b/packages/gittensory-miner/lib/replay-task-generation.js
index a861c3855f..f121b6b6e1 100644
--- a/packages/gittensory-miner/lib/replay-task-generation.js
+++ b/packages/gittensory-miner/lib/replay-task-generation.js
@@ -4,9 +4,10 @@
 // history on both sides to be worth scoring, and (b) nothing in the frozen context lets a replay run infer
 // the future by pattern-matching text rather than reasoning. This module selects calibration-worthy freeze
 // points, scrubs forward references out of the frozen context, tags each point's recency pool, and returns
-// the frozen snapshot and the revealed post-T ground truth as *separate* bundles so the replay pipeline never
-// holds both at once. Every function here is pure and deterministic — no clock, no randomness, no IO — so a
-// given (candidate, context) always yields an identical task.
+// the frozen replay task without the revealed post-T ground truth. Scoring data is exposed through a separate
+// function so replay execution never has to hold both sides at once. Every function here is pure and
+// deterministic — no clock, no randomness, no IO — so a given (candidate, context) always yields an
+// identical task.
 
 // What a scrubbed-away forward reference is replaced with. A fixed, self-delimiting token so the scrubbed
 // text stays readable and the substitution is itself deterministic.
@@ -168,9 +169,9 @@ export function classifyRecencyPool(candidate, options) {
   return lastActivityAt >= modelCutoffIso ? "recent" : "older";
 }
 
-// One-shot generator. Applies selection, then scrubs and lints the frozen context, then returns the frozen
-// snapshot and the revealed post-T ground truth as SEPARATE bundles — never merged — so a caller persists and
-// scopes them independently. An ineligible or un-scrubbable candidate is rejected without producing a task.
+// One-shot replay generator. Applies selection, then scrubs and lints the frozen context, then returns only
+// the frozen replay task. Revealed post-T ground truth is intentionally available only through
+// generateReplayScoringKey, so a replay worker/serializer/logger/model call never receives both sides.
 export function generateReplayTask(candidate, context, options) {
   const selection = selectFreezePoint(candidate, options?.thresholds);
   if (!selection.eligible) {
@@ -194,9 +195,29 @@ export function generateReplayTask(candidate, context, options) {
       commitT: typeof candidate?.commitT === "string" ? candidate.commitT : null,
       contextTexts: scrubbedTexts,
     },
-    revealed: {
-      commitCount: selection.revealedCommitCount,
-      groundTruth: candidate?.revealedGroundTruth ?? null,
-    },
+  };
+}
+
+// Scoring-only accessor. Call this from the isolated scorer path after replay execution has finished; do not
+// pass its result to replay workers. It deliberately shares only selection eligibility with generateReplayTask
+// and never carries frozen context.
+//
+// IMPORTANT: `eligible: true` here means only that selectFreezePoint accepted the candidate -- it does NOT
+// mean generateReplayTask would also produce a usable frozen task for it. generateReplayTask can still reject
+// a selection-eligible candidate afterward (`rejected: "unscrubbable_forward_reference"`, from
+// lintFrozenContext), because scrub/lint eligibility is about the FROZEN CONTEXT TEXT, which this function
+// never touches -- it only reveals commitCount/groundTruth, so lint/scrub has nothing to check here. A caller
+// must not assume a scoring key implies a replay task was ever generated for the same candidate; check
+// generateReplayTask's own result independently before treating the two as a matched pair.
+export function generateReplayScoringKey(candidate, options) {
+  const selection = selectFreezePoint(candidate, options?.thresholds);
+  if (!selection.eligible) {
+    return { eligible: false, rejected: "selection", reasons: selection.reasons };
+  }
+
+  return {
+    eligible: true,
+    commitCount: selection.revealedCommitCount,
+    groundTruth: candidate?.revealedGroundTruth ?? null,
   };
 }
diff --git a/test/unit/miner-replay-task-generation.test.ts b/test/unit/miner-replay-task-generation.test.ts
index 8fe77a2e50..4d2a8c72d6 100644
--- a/test/unit/miner-replay-task-generation.test.ts
+++ b/test/unit/miner-replay-task-generation.test.ts
@@ -4,6 +4,7 @@ import {
   RECENCY_POOLS,
   classifyRecencyPool,
   detectForwardReferences,
+  generateReplayScoringKey,
   generateReplayTask,
   lintFrozenContext,
   scrubForwardReferences,
@@ -137,7 +138,7 @@ describe("gittensory-miner leakage-safe replay task generation (#3011)", () => {
       modelCutoffIso: "2026-01-01T00:00:00Z",
     };
 
-    it("produces a scrubbed frozen bundle and a SEPARATE revealed bundle for an eligible clean point", () => {
+    it("produces only a scrubbed frozen bundle for an eligible clean point", () => {
       const task = generateReplayTask(
         { ...eligible, frozenContextTexts: ["intro references #300 and old #12"] },
         CONTEXT,
@@ -150,8 +151,8 @@ describe("gittensory-miner leakage-safe replay task generation (#3011)", () => {
         commitT: "abc1234def",
         contextTexts: [`intro references ${FORWARD_REF_PLACEHOLDER} and old #12`],
       });
-      // Ground truth lives only on the revealed side — never merged into the frozen bundle.
-      expect(task.revealed).toEqual({ commitCount: 10, groundTruth: { merged: true, approach: "refactor" } });
+      expect(task).not.toHaveProperty("revealed");
+      expect(JSON.stringify(task)).not.toContain("refactor");
       expect(task.frozen).not.toHaveProperty("groundTruth");
     });
 
@@ -187,5 +188,39 @@ describe("gittensory-miner leakage-safe replay task generation (#3011)", () => {
         generateReplayTask(input, CONTEXT, options),
       );
     });
+
+    it("exposes revealed ground truth only through the scoring-only accessor", () => {
+      const replayTask = generateReplayTask(
+        { ...eligible, frozenContextTexts: ["intro references #300"] },
+        CONTEXT,
+        options,
+      );
+      const scoringKey = generateReplayScoringKey(
+        { ...eligible, frozenContextTexts: ["intro references #300"] },
+        options,
+      );
+
+      if (!replayTask.eligible) {
+        throw new Error(`expected eligible replay task, got ${JSON.stringify(replayTask)}`);
+      }
+      if (!scoringKey.eligible) {
+        throw new Error(`expected eligible scoring key, got ${JSON.stringify(scoringKey)}`);
+      }
+      expect(replayTask).not.toHaveProperty("revealed");
+      expect(scoringKey).toEqual({
+        commitCount: 10,
+        eligible: true,
+        groundTruth: { merged: true, approach: "refactor" },
+      });
+      expect(scoringKey).not.toHaveProperty("frozen");
+    });
+
+    it("rejects a scoring key for a candidate that fails selection", () => {
+      expect(generateReplayScoringKey({ priorCommitCount: 2, revealedCommitCount: 1 }, options)).toEqual({
+        eligible: false,
+        rejected: "selection",
+        reasons: ["insufficient_prior_history", "insufficient_revealed_history"],
+      });
+    });
   });
 });