exp: freeze nonlinear CLLN hardware confirmation

1 files changed, 52 insertions, 0 deletions

diff --git a/HARDWARE_CONFIRMATION.md b/HARDWARE_CONFIRMATION.md
new file mode 100644
index 0000000..1748d85
--- /dev/null
+++ b/HARDWARE_CONFIRMATION.md
@@ -0,0 +1,52 @@
+# Frozen nonlinear CLLN confirmation
+
+## Purpose
+
+Confirm the hardware-realistic result on new component draws after the
+nonlinear model, local predictor, baselines, task set, and endpoints have been
+fixed. This is an in-silico component confirmation, not a fabricated-chip
+experiment.
+
+## Frozen protocol
+
+- Simulator: `sdil/physical_grid.py` and
+  `experiments/physical_grid_bias_p5.py`.
+- Topology: released 4-by-4 periodic CLLN with 32 learnable edges.
+- Tasks: all 40 released ring-classification tasks: five input diameters and
+  eight label rotations.
+- New component seeds: `20260833,20260834,20260835`.
+- Published-scale component model: differential gain standard deviation
+  `0.01`, twin input mismatch `0.001 V`, multiplier offset `2.3 V/s`.
+- Methods: clean, raw imperfection, 16-observation constant calibration,
+  16-observation degree-2 SDIL predictor, overclamping, and overclamping plus
+  the same SDIL predictor.
+- Standard learning: 600 epochs and `0.001 s` learning exposure per active
+  update.
+- Overclamping: at most 1,000 epochs, released nudging, `0.0025 s/V`, and the
+  existing three-perfect-checkpoint early stop.
+- No BP, autograd, device constant, or oracle-neutral method is used.
+- Pairing: all six methods receive the same task, initial gates, and component
+  draw.
+
+## Endpoints and statistics
+
+The primary endpoint is final classification error. Stable zero-error fraction,
+wall time, clamp displacement, and held-out predictor RMSE are secondary. The
+40 tasks are the bootstrap unit; the three component draws are averaged within
+task before percentile 95% intervals are computed.
+
+The result supports confirmation only if:
+
+1. every method completes all 120 task/device cells with finite task metrics;
+2. standard SDIL improves raw error in every task-averaged pair and closes at
+   least 95% of the raw-to-clean mean error gap;
+3. standard SDIL has no more than one percentage point excess mean error over
+   clean;
+4. overclamping plus SDIL closes at least 95% of the overclamp-to-clean-
+   overclamp mean error gap;
+5. constant calibration and uncorrected overclamping remain in the result
+   regardless of outcome.
+
+The output is
+`results/physical_bias/p11_nonlinear_hardware_confirmation.json`. Development
+seeds `20260829` through `20260832` remain separate.