Add Phase 6.5A: same-batch linesearch REVISES Phase 6A conclusion

Phase 6A's "better credit → worse loss" was a protocol artifact caused by:
1. Credit normalization (inflated DFA, suppressed Vec magnitude ordering)
2. Held-out evaluation (measured generalization failure, not exploitability)
3. Gradient clamping

With strict same-batch evaluation:
- Oracle BP: dL_same = -0.406 (strongest descent)
- Vec_M4: dL_same = -0.135
- ScalarCB: dL_same = -0.025
- DFA: dL_same = -0.003
Same-batch loss decrease is MONOTONIC with credit quality.

But held-out loss INCREASES for all non-DFA methods (Case D: overfitting).
The bottleneck is batch-level generalization, not surrogate exploitability.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

1 files changed, 44 insertions, 0 deletions

diff --git a/report_explore/MEMO_6.5A_samebatch_linesearch.md b/report_explore/MEMO_6.5A_samebatch_linesearch.md
new file mode 100644
index 0000000..733db12
--- /dev/null
+++ b/report_explore/MEMO_6.5A_samebatch_linesearch.md
@@ -0,0 +1,44 @@
+# Phase 6.5A Memo: Same-Batch Linesearch
+
+**Date**: 2026-03-25
+
+## Question
+Under strict same-batch evaluation, does better credit produce better loss decrease?
+
+## Answer: YES.
+
+Phase 6A's conclusion was wrong due to protocol confounds. With same-batch evaluation:
+
+### All blocks, normalized credit (closest to Phase 6A protocol):
+
+| Method | best eta | dL_same | dL_held |
+|--------|---------|---------|---------|
+| DFA | 1e-2 | -0.003 | +0.004 |
+| ScalarCB | 3e-3 | -0.025 | +0.027 |
+| Vec_M4 | 3e-3 | **-0.135** | +0.045 |
+| Oracle BP | 1e-2 | **-0.406** | +0.094 |
+
+**Same-batch loss decreases monotonically with credit quality**: Oracle > Vec > ScalarCB > DFA.
+
+### But held-out loss increases for all non-DFA methods.
+
+This is **Case D**: the local surrogate correctly exploits credit to decrease training loss, but the update overfits to the batch. The better the credit, the more effective the overfitting.
+
+### Key confounds in Phase 6A:
+1. **Normalization**: Phase 6A always normalized credit, which amplified DFA's weak signals to the same magnitude as Vec's strong signals, erasing the natural ordering
+2. **Held-out evaluation**: Phase 6A evaluated on held-out batches, showing the generalization failure rather than the exploitability success
+3. **Gradient clamping**: Phase 6A clamped gradients to [-1, 1], further distorting the relationship
+
+### Raw vs Normalized (all blocks):
+| Method | raw dL_same (best) | norm dL_same (best) |
+|--------|--------------------|---------------------|
+| Vec_M4 | -0.005 | -0.135 |
+| Oracle | -0.003 | -0.406 |
+
+Raw credit produces tiny updates because BP gradients have RMS ≈ 0.00004. Normalization brings all methods to comparable magnitude but introduces overfitting.
+
+## Revised Diagnosis
+
+The bottleneck is NOT "local surrogate cannot exploit good credit" (Case B from Phase 6A). It IS:
+- **Generalization/overfitting**: local surrogate with good credit decreases train loss but increases held-out loss
+- This means the project direction should be about **regularizing local updates** rather than replacing the surrogate entirely