diff options
| -rw-r--r-- | TWO_STATE_BIAS_PROGRAM.md | 22 |
1 files changed, 19 insertions, 3 deletions
diff --git a/TWO_STATE_BIAS_PROGRAM.md b/TWO_STATE_BIAS_PROGRAM.md index c7a238f..6811739 100644 --- a/TWO_STATE_BIAS_PROGRAM.md +++ b/TWO_STATE_BIAS_PROGRAM.md @@ -305,6 +305,22 @@ Stop this paper direction if: overclamping analogue reaches zero error without neutral observations and beats SDIL everywhere. Therefore the mechanism test passes, but the physical strong-baseline gate does not; see `results/physical_bias/p1_surrogate.json`. -- Dual Prop same-path confirmation: active/supporting, not a passed result. -- EP/CpL adapters: not implemented under this bias model. -- Current score for this new paper framing: 5/10 until P1 passes. +- Dual Prop same-path confirmation: passed as a supporting mechanism result. + Across five frozen seeds, raw structured bias reduces validation accuracy to + `9.40%`, while same-path clean and innovation reach `82.86%` and `82.92%` + means. The paired innovation-minus-raw gain is `73.52` points (95% lower + bound `73.15`); the clean-minus-innovation 95% upper bound is `0.354` points, + and the innovation-versus-oracle absolute-mean bound is `0.566` points. The + maximum residual-bias ratio is `8.57e-8`, with zero instruction observations + used by the predictor; see `results/contrastive_bias/c1_gate.json`. Because + the bias is injected from the declared predictor feature, this confirms the + adapter and causal correction but is not independent real-world evidence. +- EP adapter: implemented against the pinned Rain author code with manual local + EP parameter gradients, autograd disabled for SDIL, local neutral LMS, replay + checks and clean/raw/constant/innovation/oracle/noise endpoints. CUDA + performance screening is active; no task-performance claim is recorded yet. +- Coupled-learning/DCHN adapter: not implemented under this bias model. +- Current score for this new paper framing: 5/10. Passing the artificial Dual + Prop confirmation does not repair the failed physical strong-baseline gate; + promotion requires an independently grounded bias source and a passed + cross-backbone confirmation. |