youssefhassan13/exp02-form-constant-generator

Exp 02 — Ermentrout–Cowan form-constant generator (Stage A)

Pre-registered experiment from the Operating System Hypothesis project. The only experiment in the program that runs the biological mechanism as code: a scalar neural field on a periodic cortical sheet, pushed through its Turing bifurcation by the gain μ and rendered to the visual field through the inverse complex-log map. Then the archived Exp 01 judge (Claude Sonnet 4.6, rubric unchanged) scored every render blind, mixed with blank renders and ordinary generated scenes, as a positive control for the Exp 01 null.

Pre-registered 2026-06-29; run 2026-09-03 under two dated amendments committed before any scoring. The commit dates in the public repo are the proof.

What is here that is not on GitHub

• renders/ — 160 visual-field renders, 512 × 512, {regime}_mu{μ/μ_c}_s{seed}.png
• fields/ — 160 steady cortical fields, 256 × 256 float32 .npy
• judge_set/ — the 200 blind images exactly as the judge saw them (b000…), mapped by judge_manifest.json
• figures/ — contact sheet and the curve / control figure

Every text artifact is also in the public repo.

Design

• Model: τ ∂a/∂t = −a + f(μ (w ∗ a)); difference-of-Gaussians kernel σ_e = 1, σ_i = 2; sheet 256², side 12 λ_c
• Regimes: hex (asymmetric sigmoid, pure-noise start → hexagonal lattices) and stripe (odd sigmoid, oriented seed at noise amplitude → rolls whose orientation gives tunnels, funnels, spirals)
• Sweep: μ ∈ {0.8, 0.9, 0.95, 1.0, 1.05, 1.1, 1.25, 1.5} μ_c × 10 seeds × 2 regimes
• Negatives: 40 sub-threshold blanks; 40 images from the Exp 03 corpus at guidance 7 and 11

Headline, including the miss

Replication: blank below μ_c, all four Klüver classes above it, monotone in μ. Three of four gates pass; the planform → class specificity gate fails (the judge is multi-label under the log-polar map).

Positive control for the Exp 01 judge:

Set	n	judged structured
rendered form constants	80	80
blank renders	40	0
ordinary generated scenes	40	18

The pre-specified control failed on specificity: the binary class flags fire on literal grids (brick walls, window mullions, tiles). On the score Exp 01 actually used, no ordinary scene reaches the top level and every form constant does. Both readings are in analysis.md; the second is labelled exploratory.

Provenance

• Code, pre-registration, write-up: https://github.com/youssefhassan/operating-system-hypothesis-public/tree/main/experiments/02_form_constant_generator
• Git commit: b0c52dcf91373cdc93828977c269e1d66ca09b1e
• Exported: 2026-09-02T23:56:01.279363+00:00
• Files: {"render_png": 160, "field_npy": 160, "judge_set_png": 200, "figure_png": 2, "text": 12}

License

MIT, as the parent repository.