experiment: freeze contrastive state-bias screen

1 files changed, 121 insertions, 0 deletions

diff --git a/CONTRASTIVE_BIAS.md b/CONTRASTIVE_BIAS.md
new file mode 100644
index 0000000..5d6a718
--- /dev/null
+++ b/CONTRASTIVE_BIAS.md
@@ -0,0 +1,121 @@
+# Contrastive state-bias protocol
+
+## Question and prior-work boundary
+
+This protocol asks whether neuron-specific bias in a contrastive teaching
+difference can destroy Dual Propagation (DP), and whether a neutral-period
+somato-dendritic innovation removes that bias locally.
+
+This is not the finite-nudge estimator bias already studied for Equilibrium
+Propagation (EP).  Symmetric `+beta/-beta` EP reduces that numerical bias, and
+Dual Propagation already studies asymmetric nudging.  The candidate claim here
+is narrower: ordinary cell activity can produce a nonzero, state-dependent
+offset in the measured contrast even when the task instruction is absent.
+
+For DP population `l`, let
+
+```
+h_l = alpha * s_plus_l + (1 - alpha) * s_minus_l
+d_l = s_plus_l - s_minus_l
+```
+
+and let the observed teaching difference be
+
+```
+d_raw_l = d_l + n_l(h_l).
+```
+
+The raw condition uses `d_raw_l`.  The innovation condition receives one
+instruction-off observation of `n_l(h_l)`, fits a per-cell affine neutral
+prediction from `h_l`, and uses
+
+```
+d_innovation_l = d_raw_l - n_hat_l(h_l).
+```
+
+The oracle condition subtracts the generated nuisance exactly.  No task label,
+task loss, clean DP difference, downstream weight, or BP gradient enters the
+neutral fit.  The extra neutral observation is counted.
+
+A bias added identically to both compartments must cancel before this operation:
+
+```
+(s_plus_l + b_l) - (s_minus_l + b_l) = d_l.
+```
+
+That common-bias condition is a required negative control.  A claimed gain
+there would indicate an implementation error.
+
+## B0: equation and locality checks
+
+Before any task endpoint, deterministic tests must establish:
+
+1. zero bias makes clean, raw, innovation, and oracle differences identical;
+2. identical common bias cancels to numerical precision;
+3. a fixed differential offset is removed by the affine intercept;
+4. activity-dependent affine bias is removed by the per-cell slope;
+5. the innovation fit sees zero task-instruction observations;
+6. raw bias changes the local DP update while oracle and innovation recover the
+   clean update to numerical precision;
+7. the forward parameters, initial values, minibatches, and DP inference rule
+   are identical across conditions.
+
+## B1: frozen development screen
+
+B1 uses the patched author Dual Propagation implementation at frozen upstream
+revision `7b2595b34421e1483a721dbfdeff8cdabda3a1ff`, miniCNN on CIFAR-10,
+validation-only model seed `1988`, minibatch seed `1988`, `alpha=0`,
+`beta=0.1`, `fwK`, 16 inference passes, learning rate `0.025`, batch size 100,
+and 20 epochs.  Test is not evaluated.
+
+The complete cell set is:
+
+- one clean DP cell;
+- one ratio-4 common activity-bias cell using the raw rule;
+- raw, innovation, and oracle rules for differential fixed bias at ratios 1
+  and 4;
+- raw, innovation, and oracle rules for differential activity-dependent bias
+  at ratios 0.25, 1, and 4.
+
+The ratio is calibrated once at initialization as nuisance RMS divided by the
+clean DP state-difference RMS, separately for every hidden population.  It is
+then frozen.  A failed, nonfinite, or chance-level cell remains in the grid.
+
+B1 advances only if all mechanical checks pass and:
+
+- clean DP reaches at least 70% validation accuracy at epoch 20;
+- common bias stays within 0.2 accuracy points of clean DP and its teaching
+  difference relative error is at most `1e-6`;
+- at some activity-bias ratio, raw is at least 5 accuracy points below clean
+  or becomes nonfinite, while innovation is within 2 points of clean and
+  within 1 point of oracle;
+- innovation's post-subtraction nuisance RMS is at most `1e-3` of its raw
+  nuisance RMS;
+- every predictor report records zero task-instruction observations.
+
+The selected confirmation ratio is the largest activity-bias ratio satisfying
+the innovation conditions.  If none exists, no confirmation is opened.
+
+## B2: untouched confirmation
+
+If B1 passes, B2 freezes five new model/minibatch seeds before running clean,
+raw, innovation, and oracle at the selected activity-bias ratio for the full
+130-epoch author schedule.  Each seed uses identical initialization and batch
+order across conditions.  Accuracy, finite status, gradient alignment,
+teaching-difference error, wall time, task-loss queries, and neutral
+observations are reported.  Test is evaluated once only after the complete
+seed panel exists.
+
+The paper-facing claim requires the paired innovation-minus-raw accuracy gain
+to be positive in all five seeds, a one-sided 95% lower bound above 3 points,
+innovation within 1 point of oracle on average, and no nonfinite innovation
+run.  Failure narrows or closes the contrastive-bias claim; seeds and ratios
+are not removed after inspection.
+
+## Later extensions
+
+Random zero-mean noise, slowly drifting bias, missing neutral observations,
+and EP are separate experiments.  Random noise is expected not to be removed
+by the affine innovation and is a negative control.  EP task-accuracy evidence
+is allowed only after its clean matched implementation learns above chance;
+otherwise only equation-level EP checks may be reported.