# EXP-14b: deep XOR supervised with ample settling

- Date: 2026-10-08 | Cadence version: 0.78.0 (commit not recorded in the iter7 entry) | Program iter: iter7
- Script: `exp/EXP-14b_xor_supervised_deep.py` | Evidence: `exp/EXP-14b-supervised-deep.out`

## Summary
- Question: EXP-14's deep supervised XOR crashed just short of tolerance at 96 sweeps; with enough settling, does the deep stack fit XOR (meaning it was only settling-limited and Dongyang's cheap deep settling would be our lever)?
- Setup: deep Cortex perception(P) -> hidden(H) -> response(2), fan_in 9, parameter_prior 0.3, tolerance 1e-3. (P, H) in {(6, 6), (12, 12)}, settle_budget in {256, 512}, seeds 7, 11 (8 runs). `bootstrap` max_error 0.3, epochs 80, checks = the same 4 points. Predict budget = 3x settle_budget. Both bootstrap and predict wrapped in try/except.
- Prediction: written in the docstring before running, as a decision rule: "deep supervised XOR now PASSES / predicts [0,1,1,0] => substrate CAN represent XOR at depth, it was settling-limited => Dongyang's cheap-deep-settling IS our lever" versus "still fails after ample settling => XOR genuinely hard on this substrate; synergy needs a different irreducible task, and Dongyang is still not the lever."
- Result: all 8 runs passed=False, acc 0.50. Seed 7 preds 0.16/0.01/0.27/0.13 (P6-H6) and 0.15/0.00/0.31/0.08 (P12-H12); seed 11 preds 0.08/0.40/0.55/0.85 (P6-H6) and 0.07/0.40/0.54/0.86 (P12-H12). settle_budget 256 and 512 print identical preds in every cell. No crashes. Target [0, 1, 1, 0].
- Verdict: NULL (deep supervised does not fit XOR even with converged settling; the second pre-written branch). Not retracted. Together with EXP-14 it closed the iter7 RL caveat. Later context: iter12 (EXP-18) moved the XOR into the data channel instead, and iter13 (EXP-19) crossed M2-synergy for learned sequential composition; XOR-style representational synergy is still listed as open in the program header and HANDOFF.

## Reflection
- What it means: on the wirings tried (widths 6 and 12, one hidden column, fan_in 9), the equilibrium creature does not form the two-bit interaction XOR needs, under RL or supervised, shallow or deep. Faster settling cannot fix this, so Dongyang's method is not the lever. The preds are close to additive in the two bits (seed 11: 0.08 + 0.32 + 0.47 = 0.87 versus 0.85 observed; seed 7 is similarly additive), which is the signature of a model with no interaction term.
- Process: (1) The program log calls the sb256 and sb512 preds "BYTE-IDENTICAL"; the output only shows them equal to 2 decimals, so "identical to printed precision" is the supported claim. Identical results are consistent with the solve reaching tolerance before 256 sweeps, i.e. converged. (2) The program log summarizes seed 11 as about 0.4*b1 + 0.35*b2; the fitted coefficients from the printed preds are closer to an intercept of 0.08, about 0.47 for b1 and 0.32 for b2. The qualitative point (linear) holds; the formula is loose. (3) As in EXP-14, width barely matters (P6 and P12 nearly identical per seed), unexamined. (4) Two seeds, fixed epochs 80, no sweep of fan_in, parameter_prior, extra columns or training length; the iter7 log correctly limits the claim to "not on the architectures tried". (5) No held-out set by design (representability test).
- Lesson: when a run fails "just short" of a threshold, test the generous version before building a theory on it; here the near-miss crash suggested a settling limit that did not exist.
- Next: iter7 final verdict (Dongyang not our lever; XOR-channel synergy hits a substrate wall on the architectures tried). Led to the redesign of synergy tests that do not need the brain to host XOR (EXP-18 data-channel, EXP-19 learned sequential composition).