| # Expected results — what a successful replication looks like |
| Our measured values (AGX Orin, 2026-07; single scene family unless noted). Your |
| rig/scene should reproduce the STRUCTURE (the PASS bars), not the exact numbers — |
| absolute R and the learnability threshold are scene-local by design. |
|
|
| ## The three PASS bars (what `replicate.py` checks) |
| | # | test | bar | our value(s) | |
| |---|------|-----|--------------| |
| | P1 | Spearman(R, feat-R) | >= 0.90 | 0.981 (home), 0.969 (2nd scene) | |
| | P2 | Spearman(R, learnability) — THE LAW | >= 0.85 | 0.988 (10-cell blind), 0.884 (feat-R), 0.80 (2nd scene, weak) | |
| | P3 | ignition: learnable cell escapes (loss < log 4 = 1.386), floor cell pinned at log 4 | present | confirmed on CNN + ViT, home + 2nd scene | |
|
|
| ## What we also established (context for interpreting your run) |
| - MECHANISM: R = init feature separability. The untrained encoder preserves & |
| amplifies the pixel signal/clutter ratio into feature space (that's P1). When a |
| model learns, the SIGNAL component of feat-R amplifies ~1e5x and the head's |
| margin ramps; when it collapses, signal stays ~0 and loss sits at log 4. |
| - CROSS-ARCH: the ignition fires in a ViT too (not a CNN quirk). |
| - CROSS-SCENE: R's ORDERING is scene-universal (Spearman .996 between two |
| scenes); the THRESHOLD is scene-local (a low-R cell can be learnable on an |
| easier scene). So compare ORDERING, never absolute thresholds, across rigs. |
|
|
| ## CAVEATS that will bite you if ignored (hard-won this session) |
| 1. THE WALL IS SEED-BIMODAL / SLOW. A cell "collapses" at 8k steps and ESCAPES |
| at 18k. Escape times spread 6k-18k. => ALWAYS multi-seed (>=2, ideally 3) and |
| train long (--full = 20k). Single-seed near the wall is noise, not signal. |
| 2. COLLAPSE = loss ~ log(K) = 1.386 exactly (uniform-guess basin), acc ~ 1/K. |
| Use this to detect collapse with zero eval and early-abort to save compute. |
| 3. feat-R (the ratio) is CONFOUNDED near collapse: it can spike when the encoder |
| degenerates (clutter->0), NOT from signal. Track the SIGNAL component, or just |
| the head margin/loss, to judge escape — not the raw ratio. |
| 4. A cell that floors on our scene may learn on yours (easier scene, lower |
| threshold). That is EXPECTED (scene-local threshold), not a contradiction. |
|
|
| ## Verdict rules |
| - P1 & P2 & P3 PASS on your independent rig+scene -> the law + mechanism |
| replicate. This is the result we most want confirmed or killed. |
| - P2 fails (ordering doesn't predict) -> R is scene-specific / overfit. A real, |
| publishable negative. Send it. |
| - P1 passes, P3 fails -> feature separability holds but learning doesn't ignite |
| -> optimization/scale issue on your rig; note steps, seeds, lr. |
| Send back results_replicate.json + your scene_hash. Disagreement is a RESULT. |
|
|
|
|
| ## Note on synthetic timestamps (kit internals) |
| `replicate.py` spreads your images over a synthetic ~96-hour timeline, because the |
| task dates changes over a 24-hour age horizon and needs memory slots at ages up to |
| 24h. If you build a bank yourself, give it a timestamp array spanning well beyond |
| 86,400,000 ms — 1-second spacing silently breaks episode sampling. |
|
|
