An outside read of Stella 2022 and the surrogate run, filed as eight todos

[syncytium2]

Tony asked for an independent plan of action on the label-free detector thread, read from outside it. Eight todos and one INDEX row. No code, and no contact with that thread — its branches, PR #530 and its darkroom run folder were read only, and nothing was written to any of them.

The one defect, filed first

docs/todo/2026-09-12-ci-installs-no-elephant-so-the-surrogate-tests-skip.md — ci.yml installs .[ui,dl], tests/test_surrogates.py opens with importorskip("elephant"), and the surrogates extra pinning elephant==1.2.1 exists only on the draft branch. So the adapter written to correct Elephant's silent failures is exercised on one laptop. ⚠ Closes with that branch landing, not before.

What the paper does and does not license

• …-stellas-criterion-is-a-significance-test-not-a-training-signal.md — their criterion is the false-positive rate of a significance test; a training contrast is strictly stronger, because a surrogate can be conservative for a test and still hand a discriminator a marginal feature. Their five reasons for preferring trial shifting are properties of a test procedure, and the trial structure it needs does not exist in continuous imaging. Their dead-time practice — per-neuron minimum ISI, capped at 4 ms, because at low rates the minimum stops being biology — independently supports this project's ruling that τ is the producer's to declare.
• …-count-preservation-after-encoding-is-the-gate-stella-actually-found.md — their mechanism for uniform dither's failure is that binarization deletes events, not that intervals get too short. This encoder binarizes and clips onsets onto the boundary frames, so both channels are present and one is worse here. A per-ROI count of onsets surviving encoding could not be found in surrogate_stats.py by name; establishing whether it exists is the first task.

Reading the run that already happened

• …-join-the-leak-results-to-the-destruction-results.md — both halves are in the run folder and the join has not been made. Do-nothing survives the leak detector in every stream, which is what proves that detector is a gate and not a score. Most other survivors sit at one frame of displacement, 0.1 s against seconds-scale coordination. Rigid shift is the only candidate surviving out to 1.6 s fast and 1.4 s slow.
• …-the-verdict-rule-needs-a-displacement-floor.md — derived from destruction, not from test power, which is the defect the second review round already found once in the control gate.
• …-rigid-shift-without-wrap-loses-onsets-at-the-window-edge.md — truncation at the window end is an event-count difference: Stella's failure mode arriving by another door. The generation margin may already absorb it, and that should be measured rather than assumed.
• …-joint-isi-dithering-is-unmeasured-not-failed.md — every cell was dropped as intractable. Under a no-pruning ruling, dropping a candidate on resource limits is a decision, and an empty row reads like a negative result.
• …-the-label-free-detector-thread-has-no-exit-criterion.md — written before the numbers are read, or it cannot tell "not yet" from "no".

Numbers quoted in the todos are aggregates read from the run's run_summary.json and the discriminator metadata on 2026-09-12, each with its provenance stated in the file. No per-recording data and no recording ids.

🤖 Generated with Claude Code