Fix #2364: [Bug] MIN_ADAPTIVE_BASELINE floor untuned for v2.0.7 normalized backprop reward

apps/memos-local-plugin/core/memory/l2/README.md

@@ -122,10 +122,16 @@ and reward backprop spreads similar V values across all step traces, so
 `mean(V_without)` collapses onto `mean(V_with)` and the V7 contrast
 formula evaluates to ≈ 0 for every policy regardless of its actual
 utility. Anchoring the without-set against a neutral 0.5 baseline
-guarantees that genuinely-useful policies (V_with ≈ 0.7-0.85) score
-positive and net-neutral or harmful ones don't. As real comparable
-without-evidence accumulates, the prior gracefully dilutes and we
-recover the original V7 §0.6 contrast formulation.
+guarantees that genuinely-useful policies score positive and
+net-neutral or harmful ones don't. Post-v2.0.7 the normalized
+credit-assignment backprop (see `core/reward/backprop.ts`) leaves
+per-trace V values in a ~0.02–0.5 band (matching the
+`minTraceValue: 0.005` docstring in `core/config/defaults.ts`), so the
+`adaptiveBaseline` floor (`gain.ts::MIN_ADAPTIVE_BASELINE = 0.005`)
+lets `poolMean` flow through as the actual anchor across that band
+rather than being pinned at a stale pre-v2.0.7 constant. As real
+comparable without-evidence accumulates, the prior gracefully dilutes
+and we recover the original V7 §0.6 contrast formulation.
 
 Use cases:
 

apps/memos-local-plugin/core/memory/l2/gain.ts

@@ -60,7 +60,21 @@ export const V7_NEUTRAL_BASELINE = 0.5;
  * samples. Five is roughly "one short episode worth" of signal.
  */
 export const WITHOUT_PRIOR_PSEUDOCOUNT = 5;
-export const MIN_ADAPTIVE_BASELINE = 0.2;
+/**
+ * Protective floor for {@link adaptiveBaseline}. Only guards against a
+ * degenerate near-zero `poolMean` (empty / all-noise pool) — the normal
+ * v2.0.7 backprop V distribution sits in ~0.02–0.5 (see
+ * `core/config/defaults.ts` `minTraceValue` docstring), so the adaptive
+ * branch of `adaptiveBaseline(poolMean) = poolMean` must be reachable
+ * across that entire band. Pre-v2.0.7 this constant was 0.2, which was
+ * appropriate for the old right-to-left decay reward scale where pool
+ * means clustered around 0.5–0.85; after the normalized-credit backprop
+ * rewrite it clamped ~every pool to a stale 0.2 and effectively froze
+ * policy/skill promotion (see issue #2364). 0.005 matches the v2.0.7
+ * `minTraceValue` floor and only kicks in when the pool is truly empty
+ * or all traces sit below the trace-retention threshold.
+ */
+export const MIN_ADAPTIVE_BASELINE = 0.005;
 
 export interface ComputeGainOpts {
   tauSoftmax: number;

apps/memos-local-plugin/tests/unit/memory/l2/gain.test.ts

@@ -4,7 +4,7 @@
 
 import { describe, expect, it, vi } from "vitest";
 
-import { applyGain, computeGain, nextStatus, smoothGain } from "../../../../core/memory/l2/gain.js";
+import { adaptiveBaseline, applyGain, computeGain, MIN_ADAPTIVE_BASELINE, nextStatus, smoothGain } from "../../../../core/memory/l2/gain.js";
 import type { PolicyId, TraceRow } from "../../../../core/types.js";
 
 function mkTrace(value: number): TraceRow {
@@ -60,6 +60,47 @@ describe("memory/l2/gain", () => {
     expect(g.baseline).toBeCloseTo(0.5, 5);
   });
 
+  it("adaptiveBaseline tracks poolMean across the v2.0.7 backprop V distribution (regression for #2364)", () => {
+    // v2.0.7 backprop V values cluster in the 0.02–0.5 band (see
+    // core/config/defaults.ts "V values for typical multi-step turns are
+    // clustered around 0.02–0.5"). At these poolMeans the baseline must
+    // *actually* adapt instead of clamping to a fixed floor.
+    expect(adaptiveBaseline(0.02)).toBeCloseTo(0.02, 6);
+    expect(adaptiveBaseline(0.05)).toBeCloseTo(0.05, 6);
+    expect(adaptiveBaseline(0.15)).toBeCloseTo(0.15, 6);
+    // Bounds still hold at both ends.
+    expect(adaptiveBaseline(0.6)).toBeCloseTo(0.5, 6);
+    expect(adaptiveBaseline(0)).toBeCloseTo(MIN_ADAPTIVE_BASELINE, 6);
+    expect(adaptiveBaseline(Number.NaN)).toBeCloseTo(0.5, 6);
+    // The floor must be low enough that the entire v2.0.7 V band
+    // (documented as 0.02–0.5) is inside the adaptive branch, not below
+    // it — otherwise adaptiveBaseline collapses to a constant floor and
+    // policy/skill promotion stalls fleet-wide.
+    expect(MIN_ADAPTIVE_BASELINE).toBeLessThanOrEqual(0.02);
+  });
+
+  it("computeGain follows poolMean when the pool sits in the v2.0.7 backprop V band (regression for #2364)", () => {
+    // Typical v2.0.7 pool: with-set ~0.05–0.08, without-set ~0.02–0.05.
+    // Before the fix, baseline clamped to 0.2 and gain went strongly
+    // negative for these pools even though the with-set outperforms the
+    // without-set by the reward distribution's own scale.
+    const g = computeGain(
+      {
+        policyId: "po_v207" as PolicyId,
+        withTraces: [mkTrace(0.06), mkTrace(0.05)],
+        withoutTraces: [mkTrace(0.03), mkTrace(0.04)],
+      },
+      { tauSoftmax: 0.5 },
+    );
+    expect(g.poolMean).toBeCloseTo(0.045, 5);
+    expect(g.baseline).toBeCloseTo(0.045, 5);
+    // With baseline = 0.045 the blended without-mean sits close to the
+    // empirical mean instead of being dragged up to 0.2, so the sign of
+    // the gain reflects the with vs. without contrast rather than the
+    // stale floor.
+    expect(g.gain).toBeGreaterThan(0);
+  });
+
   it("uses value-weighted mean for the with-set when count ≥ 3", () => {
     const g = computeGain(
       {