protocol: freeze post-failure Oral-A v2 funnel

1 files changed, 105 insertions, 0 deletions

diff --git a/ORAL_A_V2.md b/ORAL_A_V2.md
new file mode 100644
index 0000000..d4586f6
--- /dev/null
+++ b/ORAL_A_V2.md
@@ -0,0 +1,105 @@
+# Oral-A-v2: representable-subspace development protocol
+
+## Status and claim boundary
+
+This is a transparent post-failure development branch.  Frozen Oral-A-v1
+failed when its SDIL run became nonfinite at epoch 89; it remains failed and is
+never relabeled.  The executable diagnosis found that its unitwise K1 causal
+target was not captured by the translation-shared feedback model.  The v2
+change is restricted to that diagnosed interface: perturb the two basis fields
+that the channel-gated vectorizer can actually express, then apply the same
+expected full-field delta rule.  Architecture, task update, causal-query count,
+cadence, sigma, data split, and v1 optimization schedule do not change.
+
+No v2 GPU endpoint was observed before this protocol and its selectors were
+committed.  The CIFAR-10 test set and confirmation seeds 10--14 remain
+untouched.  A v2 result cannot be pooled with v1 or described as preregistered
+before the v1 failure.
+
+## V2-0: mechanics and mathematical identity
+
+`experiments/conv_local_smoke.py` must establish all of the following on CPU:
+
+- structured moment-estimator cosine above 0.985 and norm ratio in [0.90,1.10];
+- antithetic directional derivative relative error below `2e-6` at a
+  differentiable fixed ResNet point;
+- absolute error below `1e-14` between the implemented A/G update and the
+  analytically expected full-field delta-rule update;
+- all legacy convolutional, BatchNorm, predictor, and local-eligibility checks
+  remain green.
+
+This stage passed before the GPU funnel was opened: cosine `0.998217`, norm
+ratio `1.006989`, JVP relative error `1.69e-10`, and delta-rule absolute error
+`2.71e-20`.
+
+## V2-1: frozen-forward causal-capture screen
+
+Use seed 0 ResNet-20, the frozen 45,000/5,000 split, the first 10,000 training
+examples, batch 128, and 400 feedback-only minibatches.  Forward weights, BN
+state and affine parameters, and the readout remain bitwise fixed.  The
+vectorizer is `channel_gated`, starts at exactly zero (`a_scale=0`), and uses
+K1 antithetic calibration with `sigma=0.01`.  Cross:
+
+- estimator: legacy `unit_targets`, v2 `channel_subspace`;
+- apical rate: `0.01`, `0.1`, `1.0`.
+
+All candidates use identical data, perturbation seed 1000, and a fixed
+64-example training-prefix exact-gradient audit.  Autograd is used only after
+calibration to audit teaching alignment; it never updates a parameter.
+
+Within each estimator select maximum early-third teaching/negative-gradient
+cosine, then maximum all-layer cosine, then lower apical rate.  V2 advances
+only if all records are finite and the selected structured estimator has:
+
+1. early-third cosine at least `0.01`;
+2. all-layer cosine at least `0.01`;
+3. early-third cosine at least `0.01` above the best matched unit-target run.
+
+The estimator-specific prediction/target cosine is reported, but it is not
+compared across modes because one record is measured in the full hidden field
+and the other in its two channel-basis moments.  No rate or threshold is added
+after results are observed.
+
+## V2-2: full ResNet-20 validation gate
+
+Only if V2-1 passes, run one seed-0 SDIL model for 200 epochs on all 45,000
+development-training examples.  Copy Oral-A-v1 exactly: ResNet-20 width 16,
+batch 128, hidden LR `0.03`, output LR `0.1`, momentum `0.9`, weight decay
+`1e-4`, and 10x drops at epochs 100 and 150.  Use the V2-1-selected apical
+rate, zero-initialized channel-gated feedback, 400 feedback-only warmup steps,
+and structured K1/every-4 calibration at `sigma=0.01`.
+
+The already frozen BP (`91.62%`) and DFA (`33.06%`) records remain the matched
+references.  V2 passes only if:
+
+1. every loss and metric is finite;
+2. final validation accuracy is within 5 points of BP and at least 2 points
+   above DFA;
+3. early-third teaching alignment is at least `0.05`;
+4. estimated total training MACs do not exceed BP.
+
+Failure closes v2; there is no LR, warmup, cadence, sigma, direction-count, or
+vectorizer recovery branch.
+
+## V2-3: untouched depth confirmation
+
+Only after V2-2 passes, copy its complete hyperparameters without depth tuning
+to ResNet-20/32/56 and model/data-loader seeds 10--14.  Refit on all 50,000
+training examples.  Each run evaluates CIFAR-10 test exactly once at the end;
+no intermediate test endpoint is logged.  Exact BP and the v1-selected DFA are
+run on the identical panel.
+
+The confirmation gate is unchanged from Oral-A-v1: all 45 trajectories finite;
+SDIL within 2 points of BP at every depth; SDIL at least 2 points above DFA at
+depth 56; paired SDIL depth-20 to depth-56 change no worse than -2 points;
+depth-56 early-third alignment at least `0.05` and at least 30% of depth-20;
+and SDIL nondominated in accuracy versus estimated MACs, causal queries, and
+peak memory.  Wall time is descriptive only.
+
+## Reviewer-score rule
+
+V2-0 and V2-1 cannot raise the strict ICLR score because they establish
+mechanics and causal capture rather than standard-scale task success.  A V2-2
+pass can move the score from 5 only after its full audit is committed.  A V2-3
+multi-depth, multi-seed pass is required for an oral-level scaling claim.
+