diff options
| author | YurenHao0426 <Blackhao0426@gmail.com> | 2026-08-06 15:56:27 -0500 |
|---|---|---|
| committer | YurenHao0426 <Blackhao0426@gmail.com> | 2026-08-06 15:56:27 -0500 |
| commit | fc42f93cee71210532983a3962f1cd227f96e5f3 (patch) | |
| tree | 524f33266f1505c7304c2e3ee484ad34eaf2057e /RESULTS.md | |
| parent | a23d761d7330be8d7ac1745be60d4291961087aa (diff) | |
results: reproduce physical structured-bias scaling
Diffstat (limited to 'RESULTS.md')
| -rw-r--r-- | RESULTS.md | 29 |
1 files changed, 29 insertions, 0 deletions
@@ -1,5 +1,34 @@ # SDIL — results log (audited sections report n; timan107 GTX-1080 / ep_pascal) +## Published physical structured-bias reproduction (P0; descriptive) + +The active two-state-bias program begins with a reanalysis of already +published physical measurements rather than a generated neural-network +corruption. We downloaded Dillavou et al.'s complete public artifact from +Zenodo record `15692914`, release `v1.0.1`, and analyzed the released +small-network tables and two-dimensional gate-voltage trajectories. This is a +post-publication descriptive reproduction, not a preregistered confirmation or +an SDIL learning result. + +For each of the three physical experiments, an ordinary log--log fit uses the +six lowest task-switching periods. Combined-error slopes are `0.061871`, +`0.288254`, and `-0.023506` (mean `0.108873`), reproducing a nonzero error +plateau under rapid switching. Squared-cycle-span slopes are `1.837374`, +`1.735974`, and `1.946804` (mean `1.840051`), corresponding to a mean span +slope of `0.920025`. Thus the distance traversed in a cycle shrinks roughly in +proportion to the period while the inferred distance per unit period remains +finite. Faster task averaging reduces the size of each cycle; it does not make +the deterministic physical drift rate vanish. + +`experiments/analyze_physical_bias_p0.py` records a SHA-256 hash for every +source CSV, audits all ten released bow-tie trajectories, emits +`results/physical_bias/p0_summary.json`, and renders +`results/figs/physical_bias_p0.png`. The result supports the existence and +non-averaging of structured bias in a real coupled-learning system. It does +not establish that SDIL removes this bias, that the effect grows with neural +network depth, or that the original paper's overclamping baseline can be +beaten. Those are separate P1 and cross-backbone gates. + ## 2026-07-21 audit and version boundary Git was initialized after inheriting the project. The original code/results are preserved at |
