protocol: freeze dynamic standard-depth recovery

1 files changed, 91 insertions, 0 deletions

diff --git a/ORAL_A_RECOVERY.md b/ORAL_A_RECOVERY.md
new file mode 100644
index 0000000..55189b1
--- /dev/null
+++ b/ORAL_A_RECOVERY.md
@@ -0,0 +1,91 @@
+# Oral-A recovery: dynamic innovation on standard ResNet depth
+
+This branch does not reopen the failed learned-vectorizer A3/A4 protocol in
+`ORAL_A.md`. It asks a different question that became available only after the
+reciprocal Kolen--Pollack substrate and fast instruction-off neutral projection
+were established: does load-bearing somato-dendritic innovation retain
+standard-image performance, credit alignment, and a positive depth benefit as
+the same canonical CIFAR ResNet family grows?
+
+No recovery task endpoint may be generated until this protocol and its
+executable gate are committed. Execution is hard-gated on both a complete D4
+accept pass (`7/10`) and a complete oral-B R2 pass (`8/10`). Thus the required
+order remains accept bar, biological oral-B, then standard-depth oral-A.
+
+## Fixed panel and reuse boundary
+
+Use CIFAR-10 ResNet-20/32/56 with width 16, BatchNorm, option-A shortcuts,
+batch size 128, all 50,000 training examples, 200 epochs, SGD momentum 0.9,
+weight decay `1e-4`, base learning rate 0.1 with 10-fold drops at epochs 100
+and 150, no warmup, and model/data-loader seeds 10--14. Test is evaluated
+exactly once at the endpoint; validation and intermediate test evaluations are
+zero.
+
+The matched methods are:
+
+1. exact BP on the identical forward architecture;
+2. tuned DFA at its already selected hidden rate 0.03 and output rate 0.1;
+3. clean reciprocal Kolen--Pollack at hidden/output rate 0.1; and
+4. dynamic innovation on the same reciprocal substrate, with traffic ratio 4,
+   traffic seed `5000 + model_seed`, 64 instruction-off calibration examples,
+   one closed-form neutral projection, and no later predictor update.
+
+The already frozen D4 clean-KP and dynamic records at ResNet-20/seeds 10--14
+are reused verbatim rather than rerun. The recovery therefore adds exactly 50
+records: BP and DFA at depth 20, and all four methods at depths 32 and 56.
+Together with D4 this forms an exact 60-cell panel. All new records must share
+one clean tracked source revision; all D4 records retain their separate frozen
+source revision. No completed record may be deleted, replaced, or pooled with
+a pilot.
+
+Fixed HFA remains closed after its preregistered short gate failed, while FF,
+PEPITA, canonical EP, BurstCCN, and Dual Prop retain their method-native
+audits. They are reported as separate baselines rather than silently moved
+onto this matched standard-ResNet grid.
+
+## Frozen mechanism and cost invariants
+
+Every dynamic record must retain all D4 invariants: finite 200-epoch
+trajectory; mean early-third credit alignment measured only diagnostically;
+final feedback/forward cosine at least 0.98 and epoch-151--200 mean at least
+0.97; instruction-to-teaching RMS error at most `1e-4`; maximum post-projection
+traffic ratio and absolute soma slope at most `1e-5`; zero instruction leakage
+into neutral observations; exactly 64 predictor-update examples; exactly one
+neutral observation per ordinary example; and zero task-loss queries.
+
+At every depth, dynamic and clean KP must each use at most `1.34x` the matched
+BP MAC estimate. Dynamic elementwise projection work is reported separately
+and must be positive. Peak allocated memory may not exceed 8 GiB on the
+authorized GTX 1080 devices. Wall time is reported but is never substituted
+for MACs, queries, or memory.
+
+## Frozen oral-A gate
+
+The five model/data seeds are the independent paired units. One-sided 95%
+Student-t bounds use four degrees of freedom. Oral-A passes only if every
+record and every audited value is finite and all of the following hold:
+
+1. Mean BP accuracy is at least 90% at every depth and every dynamic endpoint
+   is at least 87%, preventing joint collapse from satisfying a relative gate.
+2. At each depth, mean dynamic accuracy is within 2 points of matched BP and
+   the one-sided 95% upper bound on the paired deficit is at most 3 points.
+3. At each depth, mean dynamic accuracy is within 1.5 points of clean KP and
+   the upper bound on that paired deficit is at most 2.5 points.
+4. At depth 56, dynamic exceeds tuned DFA by at least 2 points on average and
+   the one-sided lower bound on the paired advantage is at least 1 point.
+5. Dynamic gains at least 0.5 points from depth 20 to 56 on average, its
+   one-sided lower bound is nonnegative, and at least four of five paired
+   seeds improve. Its mean depth gain must also be no more than 1 point below
+   BP's paired depth gain.
+6. Mean dynamic early-third alignment is at least 0.85 at every depth; every
+   depth-56 seed is at least 0.80; and the depth-56 mean retains at least 90%
+   of the depth-20 mean.
+7. Every mechanism, query, MAC, elementwise-work, memory, source, split, and
+   evaluation invariant above passes without exception.
+
+The executable runner and immutable complete-grid analyzer are
+`experiments/oral_a_dynamic_scaling.py` and
+`experiments/analyze_oral_a_dynamic_scaling.py`. A complete pass raises the
+strict reviewer score from 8 to 9 and establishes oral-A standard-depth
+scaling for dynamic innovation. Any failure is retained, leaves the score at
+8, and closes this branch without depth, seed, optimizer, or threshold repair.