A field step generates the events, it does not coincide with them

[syncytium2]

Answers the one question interface2 asked with the PRE_ARTIFACT_KILLER export (handoff §4 item 2, "the one we would most want your eyes on"). Their two readings of slow baseline at 7.28% against fast baseline at 2.33% — a step is arithmetically likelier to land inside a broad slow event, or a step is itself slow-shaped and is generating slow calls — have different consequences and their data could not separate them. Under the first the artifact work removes a coincidence; under the second there is no coincidence to remove, because the events are the artifact.

Reading A is refuted without any data

Onset inside a window is a point-in-window test, so under uniform onsets the in-window fraction is the window's duty cycle — identical in both streams whatever their rate and whatever their event width. Width cannot enter a test on a point.

Measured, redrawing only the step's position inside the region that holds it (2,000 draws per slice, keeping that recording's own rate, roster and region length):

within	fast	chance	slow	chance
±1 s	8.84%	0.16%	13.70%	0.15%
±2 s	11.01%	0.30%	14.36%	0.28%
±10 s	12.21%	1.44%	15.71%	1.46%

Flat in the stream at every half-width, exactly as predicted. What that reading describes is an extent-overlap test, not the onset test that was run.

What is at the step is the whole field at once

20250904_211: 45 distinct ROIs onsetting in slow within ±2 s of its step against a chance expectation of 0.27. 20260122_259: 36 against 0.51. Fast stands as far off its own chance line — 39 against 1.38.

And it is one waveform rather than many: the spread of what the producer wrote into width_sec collapses at the step to a CV of 0.046, against 0.3–0.7 for the rest of the same region. Each stream is compared only against itself, and the declared width_def is read from the folder and quoted rather than interpreted.

SAP013 blocked the first draft of this, correctly — it explained what the duration rule meant, which is not this repo's to say. Reading the declared string instead of asserting it is also the better tool: nothing here hardcodes an assumption about the producer's rule, and a stream declaring two would make that panel meaningless, which the folder check already refuses.

Two of nine do nothing in slow

20240723_22 has no slow onset at its step at all; 20250808_186 has four against a chance 2.4 (p = 0.15) — while both show the fast recruitment plainly. So a confirmed step can generate a field-wide slow call, in seven of nine here, not that it always does. Worth having before the flag is built: a per-event on_field_step that assumes every step recruits will be wrong twice in nine, and whatever separates those two is not in the exported columns.

Why this changes what the artifact work has to do

It is not a slow-stream problem — the 7.28-versus-2.33 contrast is about which windows the nine steps happened to fall in, not about slow being selectively affected.

And it is not confined to baseline: the largest slow recruitments sit in wash and TTX regions. That is why this got a figure rather than a paragraph — coordination under TTX is a live finding of this lab's, measured on these six detectors (FOUNDATIONS §9; syncytium2/foundations §15.1b), and a field step inside a TTX window manufactures precisely the signature that finding is made of.

Scope

tools/make_field_step_figure.py (re-runnable, --dataset selectable), the rendered figure in docs/learned/, and the measurement as JSON. The darkroom copy is the one a person opens. This decides nothing about the detections — whether a coordinated event built from step-coincident members is itself false is a call on bugarach's output, and interface2 explicitly left it here rather than writing rules for it. The nine step times are Tony's adjudications quoted from the folder's README; 282 further candidates are unadjudicated, so the effect can only grow.

sapper clear · test_index_resolves and test_site_coherence green · no source under test changed.

🤖 Generated with Claude Code

[syncytium2]

⚠ Hold: this has not been murderboarded, and by this repo's own rule it should be before it lands.

CLAUDE.md names "a figure or its caption" as a document deliverable and says to run /murderboard <artifact> on the built artifact rather than hand over a first draft. This is a first draft. It is a figure whose header carries the argument, it answers another team's question, and it makes a claim about TTX windows that touches FOUNDATIONS §9 — which is exactly the class the gate exists for.

The measurement itself is re-runnable and the numbers came from the export rather than from prose, so the risk is in the wording, not the arithmetic. Three things a review would go at first:

1. "generates them" is a causal claim from an observational coincidence structure. What is measured is simultaneity plus waveform uniformity at nine adjudicated times. That is strong, and it is not the same sentence.
2. the fast width *widening* at the step is stated and not explained. I deliberately routed it to the producer rather than guessing, but a reader may take the silence as a gap.
3. n = 9 slices, 7 of which show the slow effect, on one person's adjudications, with 282 candidates unadjudicated. The footer says so; the headline does not.

Merging is Tony's call. The figure is already in the darkroom and the finding has been sent to interface2, so nothing is blocked by holding this.