OCTAVE — SCHOOL TRAINING AND BENCHMARK RESULTS

Actual execution: 800.8 seconds recurrent experiment; 2.5 seconds dense comparison. Training plus final evaluation, excluding added diagnostics.
53,200 recurrent parameters; 74,306 dense parameters. Three training seeds. 600 base L1 updates + 600 per cloned branch; batch 16. 5,400 gradient updates / 86,400 example presentations per architecture across the experiment. Both branches consume the same fresh worlds; 57,600 unique generated training worlds per architecture.
The existing runnable model was trained from scratch on the broader school. The scaled world-model proposal is not implemented or trained.

Mean scores across three trained models. The same 512 final test worlds are reused across models/seeds; they are not 1,536 independent worlds.
Model | Branch | Exam | A bits % | B bits % | Entire forecast % | Hidden current %
Octave | continued_level1 | level1_trained_bands | 85.22 | -- | 12.83 | 97.14
Octave | continued_level1 | level1_heldout | 88.87 | -- | 11.98 | 96.81
Octave | continued_level1 | level2_trained_bands | 82.81 | 49.95 | 0.00 | 50.20
Octave | continued_level1 | level2_heldout | 84.39 | 51.03 | 0.00 | 44.34
Octave | continued_level1 | level2_outside_range | 57.03 | 50.89 | 0.00 | 51.24
Octave | continued_level1 | level2_noise | 83.58 | 51.16 | 0.00 | 45.57
Octave | level2_with_level1_replay | level1_trained_bands | 85.76 | -- | 19.40 | 96.48
Octave | level2_with_level1_replay | level1_heldout | 90.53 | -- | 14.32 | 97.33
Octave | level2_with_level1_replay | level2_trained_bands | 86.48 | 86.58 | 7.75 | 95.90
Octave | level2_with_level1_replay | level2_heldout | 92.48 | 92.91 | 9.90 | 97.79
Octave | level2_with_level1_replay | level2_outside_range | 70.17 | 70.26 | 0.00 | 82.36
Octave | level2_with_level1_replay | level2_noise | 91.64 | 92.33 | 6.84 | 97.01
Dense | continued_level1 | level1_trained_bands | 94.82 | -- | 26.76 | 96.81
Dense | continued_level1 | level1_heldout | 86.93 | -- | 15.49 | 93.29
Dense | continued_level1 | level2_trained_bands | 86.18 | 48.52 | 0.00 | 46.03
Dense | continued_level1 | level2_heldout | 78.65 | 48.57 | 0.00 | 45.51
Dense | continued_level1 | level2_outside_range | 54.02 | 50.41 | 0.00 | 29.30
Dense | continued_level1 | level2_noise | 78.18 | 48.40 | 0.00 | 47.33
Dense | level2_with_level1_replay | level1_trained_bands | 92.87 | -- | 23.83 | 95.96
Dense | level2_with_level1_replay | level1_heldout | 82.30 | -- | 11.07 | 94.27
Dense | level2_with_level1_replay | level2_trained_bands | 92.44 | 91.86 | 7.36 | 94.79
Dense | level2_with_level1_replay | level2_heldout | 85.24 | 83.08 | 4.23 | 94.21
Dense | level2_with_level1_replay | level2_outside_range | 52.62 | 50.56 | 0.00 | 74.80
Dense | level2_with_level1_replay | level2_noise | 84.57 | 82.59 | 3.39 | 92.97

Level 2 -> Level 1 transfer, held-out period bands, paired percentage-point changes:
Seed 117: bits +3.59; exact +10.94
Seed 228: bits +0.75; exact -1.95
Seed 339: bits +0.63; exact -1.95
Mean: bits +1.66; exact +2.34

Integer-school exams of these newly trained checkpoints:
117_continued_level1 / basic_eight_contexts: bits 53.12%, exact 0.00% (8 exhaustive contexts)
117_continued_level1 / all_1044_integer_pulse_contexts: bits 92.95%, exact 13.60% (1044 exhaustive contexts)
117_level2_with_level1_replay / basic_eight_contexts: bits 51.56%, exact 0.00% (8 exhaustive contexts)
117_level2_with_level1_replay / all_1044_integer_pulse_contexts: bits 95.17%, exact 42.43% (1044 exhaustive contexts)
228_continued_level1 / basic_eight_contexts: bits 49.22%, exact 0.00% (8 exhaustive contexts)
228_continued_level1 / all_1044_integer_pulse_contexts: bits 92.18%, exact 18.97% (1044 exhaustive contexts)
228_level2_with_level1_replay / basic_eight_contexts: bits 44.14%, exact 0.00% (8 exhaustive contexts)
228_level2_with_level1_replay / all_1044_integer_pulse_contexts: bits 92.64%, exact 32.28% (1044 exhaustive contexts)
339_continued_level1 / basic_eight_contexts: bits 52.34%, exact 0.00% (8 exhaustive contexts)
339_continued_level1 / all_1044_integer_pulse_contexts: bits 94.19%, exact 22.41% (1044 exhaustive contexts)
339_level2_with_level1_replay / basic_eight_contexts: bits 52.34%, exact 0.00% (8 exhaustive contexts)
339_level2_with_level1_replay / all_1044_integer_pulse_contexts: bits 94.82%, exact 28.93% (1044 exhaustive contexts)

Interpretation limits
This is a preliminary fixed-budget experiment, not a mastery or frontier-performance certificate. More updates could change the comparison. Same data/update budget does not mean same compute: architectures, learning rates and gradient clipping differ. Eight-observer superiority requires matched observer-count ablations. Complete L2 forecasts require all 64 A+B bits correct, whereas complete L1 forecasts require 32 A bits. Log loss and Brier scores appear in raw results. No uncertainty calibration is claimed.
Historical recurrence/copy reference uses strong integer-lag structure and separately trained weights. It is included to expose performance limits, not as a matched learned baseline.
Inference lesions demonstrate reliance on communication/routing/feedback for seed 117, but are not retrained ablations or evidence of physical-phase alignment.
Final exams did not select checkpoints, alter teaching or drive training. Train bands, held-out bands, noise protocol and reproduction commands are documented in README.txt.
No level is automatically unlocked by these results.

Measured conclusions
L2 improves held-out L1 bit accuracy in all three recurrent seeds: mean +1.66 percentage points. Exact L1 improves in seed 117 but declines in seeds 228 and 339: mean +2.34 points, driven by one seed. No reliable exact-transfer claim.
Held-out L2 combined A+B bit accuracy: recurrent 92.70%, dense 84.16%; complete forecasts: recurrent 9.90%, dense 4.23%. Dense performs better on trained-band bit forecasts, and on exact L1 forecasts in its L1-only branch.
All six new recurrent final checkpoints score 0/8 exact basic-pulse forecasts. Broad continuous training from scratch is insufficient to establish the easiest-pattern mastery. This is not forgetting an inherited basic checkpoint: these models did not load one. Explicit basic-pattern curriculum coverage and retention tests are necessary.
The older recurrence/copy model remains stronger on held-out L1: 98.19% bits / 58.20% exact. Its structural bias and historical training differ.
Compute: recurrent experiment 800.82 seconds; dense 2.51 seconds on this NumPy runtime. Includes final evaluation; not a pure training-throughput comparison. Current implementation costs much more compute.