protocol: freeze hierarchical causal-capture funnel

1 files changed, 94 insertions, 0 deletions

diff --git a/ORAL_A_V4.md b/ORAL_A_V4.md
new file mode 100644
index 0000000..db1af1a
--- /dev/null
+++ b/ORAL_A_V4.md
@@ -0,0 +1,94 @@
+# Oral-A-v4: causal calibration of hierarchical feedback maps
+
+## Status and claim boundary
+
+The output-error-only V2/V3 feedback families failed their frozen early-layer
+causal-capture gates. A held-out oracle then showed that ordinary downstream
+activations are not enough, whereas gated 3x3 maps of the actual residual-DAG
+child error fields can reconstruct early credit almost exactly. Fixed random
+hierarchical FA subsequently improved the matched short ResNet-20 endpoint to
+43.52% from DFA's 37.16%, but failed its frozen 50% full-run gate.
+
+V4 is the first trainable use of that diagnosed spatial hierarchy. Every
+feedback convolution consumes its local child teaching field. One independent
+Rademacher tensor is drawn in each feedback-parameter space; all induced parent
+fields are injected in a single antithetic pair, and the scalar task-loss
+derivative estimates each edge's causal target moment. The locally computed
+prediction moment is subtracted before updating Q/R. Forward weights are never
+copied or read by the learning rule, and no reverse-mode differentiation is
+used.
+
+Recursive learned feedback is not claimed as the Harnett-specific novelty.
+V4 is an engineering prerequisite for a later hierarchical innovation model
+and must be compared with fixed HFA, learned FA/weight mirroring, and BurstCCN.
+The frozen Oral-A-v1 failure remains failed; V4 cannot retroactively open A4.
+No V4 GPU endpoint was observed before this protocol and its selector were
+committed. CIFAR-10 test and confirmation seeds 10--14 remain untouched.
+
+## V4-0: mechanics gate
+
+Before any endpoint, `experiments/conv_local_smoke.py` must verify:
+
+- the simultaneous BatchNorm-coupled antithetic JVP to relative error below
+  `2e-6`;
+- under an audit-only symmetric Q=W and R=-Wout-transpose copy, every predicted
+  local feedback moment equals its exact causal delta-rule target to relative
+  error below `2e-12`;
+- the symmetric hierarchical field and local parameter update continue to
+  reproduce exact BP under the pre-existing thresholds;
+- all legacy convolutional checks remain green.
+
+This gate passed with JVP error `6.16e-10`, delta-rule relative error
+`2.30e-16`, hidden-field relative error `4.97e-16`, and parameter-update
+absolute error `1.49e-8`. Actual training never makes the symmetric copy.
+
+## V4-1: frozen-forward causal-capture gate
+
+Use seed-0 ResNet-20, the first 10,000 development-training examples, batch
+128, the frozen 45k/5k split, random hierarchical feedback scale 1, one
+direction, `sigma=0.01`, perturbation seed 1000, and a fixed 64-example exact
+gradient audit. Forward weights, readout, BatchNorm state, and affine
+parameters remain bitwise fixed.
+
+Record one uncalibrated HFA reference and calibrate Q/R for 400 feedback-only
+minibatches at `eta_A in {0.1, 1.0, 10.0}`. Select maximum early-third
+teaching/negative-gradient cosine, then maximum all-layer cosine, then lower
+rate. V4-1 passes only if all four records are finite and the selected method:
+
+1. reaches early-third alignment at least `0.05`;
+2. reaches all-layer alignment at least `0.10`;
+3. exceeds the matched uncalibrated HFA early alignment by at least `0.04`;
+4. keeps every feedback/forward audit norm ratio in `[0.1, 3.0]`.
+
+The norm gate is audit-only and prevents a scale explosion from being hidden
+by cosine. No extra rate, query count, initialization, warmup, normalization,
+or per-layer schedule is added after endpoints are observed.
+
+## V4-2: bounded short accuracy gate
+
+Only a V4-1 pass opens a 10k-example, 20-epoch ResNet-20 validation screen.
+Copy A2b exactly: batch 128, cosine decay, no warmup, momentum 0.9, weight decay
+`1e-4`, output LR 0.1, and hidden LR `{0.03, 0.1}`. Use the V4-1-selected
+feedback rate, 400 feedback-only warmup minibatches, then K1/every-4 joint
+calibration. Select validation accuracy, then lower MACs/rate.
+
+Advance only if both runs are finite, selected V4 reaches at least 60%, exceeds
+fixed HFA's 43.52% by at least 10 points, retains early alignment at least
+0.05, and does not exceed `1.5x` exact-BP estimated MACs. Failure closes V4
+without a full run or an expanded grid.
+
+## V4-3: full validation and confirmation rule
+
+Only a V4-2 pass opens one 200-epoch seed-0 validation run on all 45,000
+development-training examples. Copy the selected hidden rate, output rate 0.1,
+400-step feedback warmup, K1/every-4 joint calibration, and the A1 step drops
+at epochs 100/150. It must be finite, finish within 5 points of BP 91.62%, beat
+DFA 33.06% and fixed HFA 43.52% by at least 2 points, retain early alignment
+0.05, and remain within `1.5x` BP MACs. There is no recovery branch.
+
+A complete V4-3 validation pass may move the strict reviewer forecast from 5
+to 6 and permits a new, separately frozen hierarchical-SDIL residualization
+test. It does not itself establish Harnett-specific novelty. Only after that
+innovation test passes may an independent ResNet-20/32/56 confirmation be
+specified; no test endpoint is authorized by V4-1 or V4-2.
+