38

Source document

[2026-03-19 00:19] xenonfun: ``` ⏺ PD-HAM at step 15K — the protect gate has not specialized. - protect_gate = 0.506 (same ~50% as step 6K — no movement) - eff_gate = 0.278 (same) - write_fraction = 1.0 (every token still writes) - slot_drift = 0.30 (low — memory is stable, but because writes are weak everywhere, not because distractors are suppressed) The protect gate learned a uniform blanket suppression (~50%) rather than discriminating between entity tokens and distractors. This is the "diffuse gating" failure mode predicted in the plan. S2 = 1.9% at 15K, same as chance. ~15 min to finish. Summary across all 6 mechanism tests so far: ``` [2026-03-19 00:19] xenonfun: ``` ┌────────────┬───────────┬───────────────────────────────────────────────┐ │ Mechanism │ S2 at 20K │ Key insight │ ├────────────┼───────────┼───────────────────────────────────────────────┤ │ Baseline │ 1.6% │ No mechanism — S1=80% but S2=chance │ ├────────────┼───────────┼───────────────────────────────────────────────┤ │ AnchorKAN │ 1.8% │ Soft attractors: active but useless │ ├────────────┼───────────┼───────────────────────────────────────────────┤ │ VQ │ 1.0% │ Hard codes: active but useless │ ├────────────┼───────────┼───────────────────────────────────────────────┤ │ HAM soft │ 1.6% │ Diffuse reads/writes (entropy 99% max) │ ├────────────┼───────────┼───────────────────────────────────────────────┤ │ HAM sparse │ 1.2% │ Top-1 write: selective but still contaminated │ ├────────────┼───────────┼───────────────────────────────────────────────┤ │ PD-HAM │ ~1.9% │ Protect gate uniform, not discriminating │ └────────────┴───────────┴───────────────────────────────────────────────┘ None have moved S2. The protect gate is the right idea but it's not learning to discriminate — likely because the gradient signal for "this is a distractor, don't write" is too weak at the entity curriculum fraction we're using (10%). ```