protocol: freeze endogenous shared-feedback screen

1 files changed, 116 insertions, 0 deletions

diff --git a/SHARED_FEEDBACK.md b/SHARED_FEEDBACK.md
new file mode 100644
index 0000000..a7d34c2
--- /dev/null
+++ b/SHARED_FEEDBACK.md
@@ -0,0 +1,116 @@
+# Endogenous shared-feedback feasibility protocol
+
+## Question and claim boundary
+
+The existing clean-task scaling results cannot identify an SDIL scaling
+contribution.  When reciprocal Kolen--Pollack (KP) feedback contains only the
+instructional field, the neutral predictor is zero and SDIL is exactly clean
+KP.  Adding an otherwise unused nuisance field and removing it establishes
+robustness, not innovation-enabled scaling.
+
+This protocol tests a different system constraint before any further scaling
+claim is allowed: one apical pathway must carry both task-required contextual
+feedback during inference and an instructional field during learning.  The
+ordinary component is therefore present with no bias or nuisance intervention,
+and disabling it must damage the task.  SDIL may advance only if neutral-period
+prediction separates instruction from that shared field.  A clean/oracle KP
+condition is retained as an upper bound, but it receives a separate
+instruction-only observation that the shared-path conditions do not receive.
+
+Passing this feasibility screen is not a scaling result and cannot change the
+review score.  A later depth or architecture experiment may be opened only by
+the frozen gate below.
+
+## S0 contextual task and architecture
+
+Each example contains two independent standard-normal features `x0,x1` and a
+balanced binary context `z`.  The target is `1[x0 > 0]` when `z=0` and
+`1[x1 > 0]` when `z=1`.  Input features alone have a 75% Bayes ceiling because
+the context is independent and unavailable on the basal input path.
+
+The student is a two-hidden-layer width-64 tanh network.  Context enters every
+hidden population only as a fixed random apical field
+
+```text
+n_l(z) = C_l onehot(z)
+u_l    = W_l h_(l-1) + n_l(z)
+h_l    = tanh(u_l).
+```
+
+`C_l` is drawn once with elementwise standard deviation `1/sqrt(2)` and is
+identical across conditions.  There is no context input to the basal stream or
+readout.  The output is a two-class linear readout.  Removing `n_l(z)` at
+evaluation is the required context lesion.
+
+Forward and reciprocal weights use independently initialized modified-KP
+updates with equal momentum and decay.  For population `l`, let `t_l` be the
+instructional field transported by its reciprocal path.  The measured shared
+apical field is generated by the forward computation itself:
+
+```text
+a_l = n_l(z) + t_l.
+```
+
+No traffic ratio, additive offset, measurement error, or generated nuisance is
+introduced.
+
+## Conditions
+
+All conditions share examples, minibatch order, initial forward/reciprocal
+weights, fixed context projections, optimizer, and forward computation.
+
+1. `oracle`: use the separately observed instruction-only field `t_l`.
+2. `raw_shared`: use the shared field `a_l` directly.
+3. `innovation`: fit a per-cell affine neutral predictor
+   `n_hat_li = p_li h_li + b_li` from instruction-off observations and use
+   `a_l - n_hat_l`.
+4. `matched_raw`: preserve the direction of `a_l` but match its per-example
+   norm to the innovation, ruling out a magnitude-only explanation.
+
+The predictor sees context-bearing forward states and neutral apical activity,
+but no label, loss, instructional field, oracle difference, or downstream
+weight.  A fixed 512-example training-only neutral set is observed before each
+epoch.  These observations and their extra forward work are counted.
+
+## Frozen development screen
+
+Use 8,192 generated training examples and an independently generated balanced
+2,048-example validation set, batch size 128, 40 epochs, model/data seed 3101,
+SGD momentum 0.9, hidden/output learning rate 0.03, reciprocal learning rate
+0.03, and weight decay `1e-4`.  There is no learning-rate, context-scale,
+predictor-rate, width, depth, or epoch grid.  Validation is evaluated only at
+the endpoint.  No test split is generated in S0.
+
+Before the task run, deterministic mechanics tests must establish:
+
+- identical forward outputs and parameters across all four conditions;
+- exact context removal in the context lesion;
+- zero instructional observations during neutral predictor fitting;
+- `raw_shared = n + t`, oracle recovery under exact subtraction, and exact
+  per-example norm matching;
+- reciprocal updates do not read or copy forward weights or updates;
+- the zero-context limit makes oracle, raw, matched raw, and innovation
+  identical when the predictor is zero.
+
+S0 passes only if all mechanics checks pass and the single frozen run satisfies
+all of:
+
+- oracle validation accuracy is at least 90%;
+- the oracle context lesion loses at least 10 accuracy points;
+- neutral context RMS is nonzero in every hidden population;
+- raw shared feedback is at least 5 points below oracle or becomes nonfinite;
+- innovation is at least 5 points above raw and no more than 3 points below
+  oracle;
+- innovation is no more than 2 points from exact generated-context subtraction;
+- matched raw remains at least 3 points below innovation;
+- the neutral predictor explains at least 80% of per-cell context variance on
+  average and leaves at most 25% of raw context RMS;
+- all predictor reports contain zero task-instruction observations.
+
+Failure closes this realization without tuning it.  Passing permits one new,
+separately committed useful-depth protocol.  That protocol must demonstrate a
+BP/oracle benefit from depth, a growing raw shared-path deficit, recovery by
+innovation, context lesions, and an explicit accounting of the extra clean-KP
+instruction wire.  Injected-bias or clean-KP scaling results cannot substitute
+for those checks.
+