diff options
Diffstat (limited to 'report_explore/MEMO_10A6_structured_vs_semantic.md')
| -rw-r--r-- | report_explore/MEMO_10A6_structured_vs_semantic.md | 36 |
1 files changed, 36 insertions, 0 deletions
diff --git a/report_explore/MEMO_10A6_structured_vs_semantic.md b/report_explore/MEMO_10A6_structured_vs_semantic.md new file mode 100644 index 0000000..c63b5ea --- /dev/null +++ b/report_explore/MEMO_10A6_structured_vs_semantic.md @@ -0,0 +1,36 @@ +# Phase 10A.6 Memo: Structured vs Semantic Auxiliary + +**Date**: 2026-03-26 + +## Question +Is the blend gain from semantic credit, structured trainable signal, or trainable regularization? + +## Results + +| Branch | final | diff vs DFA | Interpretation | +|--------|-------|-------------|----------------| +| continue_DFA | 0.312 | — | baseline | +| random_trainable | 0.324 | +1.2% | works | +| shuffled_trainable | 0.325 | +1.4% | works (no semantics needed) | +| **zero_target** | **0.221** | **-9.1%** | **crashes** (must output non-zero) | +| fresh_random_target | 0.325 | +1.3% | works (stable targets not needed) | +| time_only | 0.321 | +1.0% | works (h_l not needed, just depth) | +| **constant_input** | **0.312** | **+0.0%** | **neutral** (needs at least depth info) | +| prefit60_frozen | 0.127 | -18.4% | crashes (frozen = bad) | +| prefit60_trainable | 0.321 | +1.0% | works but ≈ random_trainable | + +## Mechanism Identified + +The gain requires: **(1) trainable, (2) non-zero output, (3) at least depth-aware**. + +- **Not semantic credit**: shuffled and fresh_random targets work equally well +- **Not pure trainability**: zero_target crashes (the aux must actually output something) +- **Not state-dependent**: time_only (no h_l) works almost as well as full Vec +- **Depth-awareness matters**: constant_input (no depth info) doesn't help +- **Prefit adds nothing**: prefit60_trainable ≈ random_trainable + +## Conclusion + +The mechanism is a **depth-aware trainable auxiliary perturbation** that diversifies the block-local update directions beyond what DFA alone provides. It doesn't need to be semantically correct credit — it just needs to be a non-trivial, evolving, depth-dependent signal that prevents blocks from collapsing into the DFA-only fixed-direction regime. + +This is NOT evidence for the credit bridge hypothesis. The gain is an optimization dynamics phenomenon unrelated to credit quality. |