From 41555c80d365eb23801e21e7795e894eacadde4e Mon Sep 17 00:00:00 2001 From: YurenHao0426 Date: Thu, 6 Aug 2026 17:23:43 -0500 Subject: docs: record Rain EP bias evidence --- TWO_STATE_BIAS_PROGRAM.md | 33 ++++++++++++++++++++++++++++----- 1 file changed, 28 insertions(+), 5 deletions(-) diff --git a/TWO_STATE_BIAS_PROGRAM.md b/TWO_STATE_BIAS_PROGRAM.md index 6811739..c8ed934 100644 --- a/TWO_STATE_BIAS_PROGRAM.md +++ b/TWO_STATE_BIAS_PROGRAM.md @@ -317,10 +317,33 @@ Stop this paper direction if: 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. + checks and clean/raw/constant/innovation/oracle/noise endpoints. The first + parameter-measurement adapter is closed: fixed bias makes raw EP nonfinite + while same-RMS noise remains trainable, but online, 64-batch and matched + 1000-batch affine calibration all fail to recover clean learning; see + `results/ep_bias/s0_summary.json`. +- EP neuron-state adapter: mechanism-positive but not yet confirmed. The S1 + development screen reaches `78.05/49.30/62.95/77.10/78.10%` for + clean/raw/constant/innovation/oracle with 128 matched neutral observations + and no extra equilibrium, but its frozen five-seed C1 gate fails. Innovation + beats raw in every seed by `43.14` points on average (95% lower bound + `24.35`), and its residual is below constant in every seed, but one seed + reverses the accuracy comparison against constant; clean/oracle/noise endpoint + bounds also fail under the unstable three-epoch Rain dynamics. See + `results/ep_bias/c1_gate.json`. +- A more hardware-grounded S3 bias is normalized only by the initial free-state + activity and is bitwise independent of beta sign and task difference. In its + single-seed development screen, clean/raw/constant/innovation reach + `79.15/62.75/67.10/77.75%`; innovation is `+15.00/+10.65` points over + raw/constant and is `1.40` points below clean with essentially zero measured + wall overhead. This is not a confirmation. Random-sign beta is not yet a + functioning comparator in the pinned positive-EP configuration: both random + sign and fixed negative beta remain at `9.9%` even with zero injected bias, + so negative-beta stability must be solved independently before comparison; + see `results/ep_bias/s3_summary.json`. - 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. + Prop confirmation and the single-seed Rain S3 screen do not repair the failed + physical strong-baseline gate or the failed Rain C1 confirmation. Promotion + requires an independently grounded bias source, a stable EP protocol with a + functioning random-sign comparator, and a passed cross-backbone confirmation. -- cgit v1.2.3