Within-Cell Null Control (2026-08-21)

Answers Findings Log item 105(d), and replaces the timestamp-shuffle control that items 8 and 93 were read from.

Reproduce with ./scripts/evaluate-italy-spatial-shift-controls.sh.

The shuffle was never a null

permute-spatial-targets shuffles whole time slices: it keeps each anchor's 19-cell label vector intact and reassigns it to a random anchor. That looks like a temporal null, and against independent rows it would be one.

The rows are not independent. Anchors are 30 minutes apart and the target horizon is 1–7 days, so consecutive target windows overlap by more than 99% and each cell's label sequence is a handful of long runs. Shuffling the order of those rows converts a few long runs into many short ones, and every run boundary is a label transition — the quantity Target Design shows is the real sample size.

Measured on the item-104 tables:

	Real order	Shuffled controls
`era_0` held-out transitions	1	94, 98, 102, 111, 119
`era_3` held-out transitions	19	315, 317, 322, 336, 342

The control carried up to a hundred times the evidence of the run it was meant to null. It was not the same task made harder; it was a different, far better-resourced task, and no comparison between the two means anything.

This is why item 8's reading — "all five controls beat real order, so destroying temporal structure makes the task easier, which is the opposite of a predictive signal" — does not survive. The controls did not beat real order by being easier to fit. They beat it by being scored on more data.

A circular shift is the null the shuffle was meant to be

shift-spatial-targets moves the whole labeled matrix in time by a single offset, wrapping at the end. Within each cell that preserves:

the sequence length, exactly;
the positive count, exactly;
the run-length structure, and therefore the transition count, apart from the one wrap seam;
the spatial pattern at each anchor, since every cell moves together.

What it destroys is the alignment between features and labels, which is the only place a temporal signal could live. Unlabeled rows keep their empty labels and do not enter the shift — the evaluator drops them anyway, and moving them would change which anchors are scoreable and reintroduce the asymmetry the control exists to remove.

The offset is drawn away from both ends (--min-shift-fraction, default 0.1) so a control cannot come back as a near-copy of the real ordering.

The preservation check

Every run prints it, and it is the check that the control is usable:

shift: 349 anchors across 19 cells
transitions: 74 before, 76 after

On era_3, per cell:

Cell	Length	Positives before → after	Transitions before → after
`cell_09_43.25_10.75`	520	226 → 226	16 → 16
`cell_11_43.25_13.75`	520	257 → 257	10 → 10
`cell_14_44.75_10.75`	520	156 → 156	9 → 10
`cell_10_43.25_12.25`	520	104 → 104	7 → 8
`cell_07_41.75_13.75`	520	134 → 134	8 → 8

Lengths and positive counts are identical everywhere. Two cells gain one transition each, where the wrap seam happens to fall inside a run. Total 74 → 76, against the shuffle's 74 → 322.

A before/after difference of more than two or three means the shift is not preserving structure and the control should not be used.

What the valid null shows on `era_0`

era_0, all_features, 600 epochs both sides, stratified on target_cell_id:

Run	Pooled	Stratified	Strata scored	Transitions
real order	`0.392915`	`0.576923`	1 of 19	1
shift 101	`0.429952`	`0.500000`	4 of 19	6
shift 202	`0.515816`	`0.611111`	4 of 19	5
shift 303	`0.494048`	`0.488990`	3 of 19	4
shift 404	`0.624008`	`0.525442`	3 of 19	4
shift 505	`0.554688`	`0.449642`	2 of 19	3

The evidence gap is closed to within a factor of a few: 3–6 transitions against the real run's 1, where the shuffle gave 94–119. The controls behave like a null should, landing on a stratified mean of 0.515 across a 0.450–0.611 band centred on chance.

A residual asymmetry remains, and it is not the shift misbehaving. era_0's held-out window is genuinely unrepresentative of era_0: it holds 6.2% of the era's transitions in 20.2% of its anchors. Any control that draws its held-out labels from elsewhere in the record lands on a busier stretch. era_3 is closer to fair at 25.7% of transitions in 20% of anchors. This is worth stating plainly — the real run's test partition is the quietest fifth of the era, which is a further reason not to read its score.

The item-8 pattern splits, and neither half is a result

Item 8 reported that all five shuffled controls beat real chronological order, and read it as evidence against a temporal signal. Under the valid null:

Pooled, 5 of 5 controls still beat real order. But era_0's real pooled score is 0.392915, well below chance, and any null sits near 0.5. This says the real run is sub-chance pooled — the item-93 cross-era artifact — not that shuffling helps.
Stratified, 1 of 5 beats real order. The real run is near the top of the null band rather than below it, which is the opposite of the item-8 reading.

Both halves rest on 1 held-out transition against 3–6, so neither is interpretable. The honest conclusion is not that item 8 was backwards; it is that the comparison item 8 drew cannot be made on this record at all, and the appearance of a clean directional result came from a control scored on a hundred times the evidence.

What the valid null shows on `era_3`

era_3 is the era worth reading: 19 held-out transitions against era_0's 1, 8 scoreable cells against 1, and a held-out window that is close to representative of the era.

The null is evidence-matched here. Held-out transitions, five seeds:

	Real	Shift control	Shuffle control
`era_0`	1	3, 4, 4, 5, 6	94, 98, 102, 111, 119
`era_3`	19	10, 13, 17, 17, 22	315, 317, 322, 336, 342

The shift controls bracket the real run. That is what the comparison needed and never had.

Stratified balanced accuracy, 600 epochs both sides:

Ablation	Real	Control mean	Control range	Controls beating real
`seismic_only`	`0.552836`	`0.461116`	`0.400764`–`0.552511`	0 of 5
`seismic_vlf`	`0.539445`	`0.483559`	`0.404473`–`0.585987`	1 of 5
`vlf_only`	`0.532717`	`0.532322`	`0.476887`–`0.624882`	2 of 5
`all_features`	`0.404832`	`0.454386`	`0.374233`–`0.500000`	3 of 5

The item-8 pattern does not survive a matched null

Under the shuffle, 5 of 5 controls beat real order for every ablation, in both eras. Under the shift the count is 0, 1, 2, or 3 of 5 depending on the ablation, with no systematic direction. The clean directional result that item 8 reported — and that item 93 re-derived within eras — was a property of the control, not of the data. It should not be cited again.

`seismic_only` clears its null by a hair, and that is not a result

seismic_only is the one ablation no control beats. The margin is 0.552836 against a best control of 0.552511: a gap of 0.0003, on 19 label transitions, with the control spread running 0.40–0.55. Real order ties its best control. Counting "0 of 5" and reading it as separation would be exactly the error Target Design was written to stop.

Adding modalities makes it worse, consistently

Reading down the ablation table, every feature family added to seismic history lowers the stratified score: seismic_only 0.5528 → seismic_vlf 0.5394 → all_features 0.4048. vlf_only sits at 0.532717 against a control mean of 0.532322 — level with its own null to four decimal places.

This is consistent with items 102 and 104, and it is still not evidence of absence. At 19 transitions the interval on every one of these numbers spans the whole table. What can be said is narrower and worth saying plainly: with a null that is finally matched on evidence, no modality separates from chance, and nothing improves on seismic history alone.

What this settles

Settled:

A usable null exists. Sequence length and positive count preserved exactly per cell, run structure preserved apart from the wrap seam, held-out transitions bracketing the real run instead of exceeding it by two orders of magnitude.
The timestamp-shuffle results in items 8 and 93 are retired. Not reversed — retired. The comparison they drew cannot be made with that control.
No ablation separates from its null on era_3. The closest is seismic_only, by 0.0003.

Not settled:

Whether any modality carries signal. Nineteen held-out transitions cannot answer that, and this document does not claim to. The blocker recorded in item 104(d) is unchanged: calendar time.
Whether the wrap seam matters at longer records. It adds at most a couple of transitions here; on a record with many more runs it would be worth replacing the single circular shift with a block bootstrap.

Nothing here supports any claim about earthquake prediction.