Every claimed source, re-verified at the task level

Run 2026-09-19. Code: gate/verify_sources.py. Output: gate/verify_sources.json. Control: PRIOS (ds004370) returns 361 stimulation trials in its sampled events file; had it returned zero, the counting method would be broken and the rest of this discarded.

Why this run exists

Two entries in docs/source-census.md were found wrong on 2026-09-19. Both had been written from what a deposit advertises — task labels, subject counts, the presence of the BIDS-iEEG stimulation schema — rather than from its events tables. And gate/prios_io.list_keys was found to truncate at 1,000 keys without following S3 continuation tokens, so any earlier sweep of a large deposit saw a partial view and reported it as complete.

Seven of the eleven deposits below had never been opened at all. The census said so plainly: "The BIDS argument says they should carry intensity too, but that is an inference, not a check." This is the check.

Result

What the inference got right, and wrong

Of the seven deposits never opened, four carry stimulation with per-trial current as the BIDS argument predicted — ds005448, ds004457, ds004977, ds003708. Three do not.

ds007703 was recorded as "n=2, SPES in white matter." It has 8,410 keys and its events files contain one row each. Whatever is in that deposit, the task events do not describe stimulation trials.

ds005558 was recorded as "n=9, closed-loop." Its only task is catFR3, a free-recall memory paradigm, with 1,320 event rows and no stimulation trial types. Closed-loop stimulation may well have occurred; it is not in the events tables, so it cannot be read.

ds004642 was recorded as "n=10, intraoperative." It has 65 keys, ten subjects, and zero events files. There is no table to read.

What the run found that nobody had looked for

ds003848 task-CHOCS1 carries 861 stimulation trials with current recorded, across four runs — more than its own SPESclin task and more than any single PRIOS run. CHOCS was not mentioned anywhere in the census, which discussed ds003848 only as SPESclin-plus-Sleep. It was found by enumerating tasks rather than by looking for the tasks already believed to be there.

Standing after this run

Eight deposits are confirmed to carry electrical stimulation with per-trial current recorded, spanning 51 human subjects. Seven of those deposits, and 44 of those subjects, have never been computed. The census's central claim — that human intracranial stimulation data is the register's most productive vein — was an inference when written and is now a verified count.

Three deposits leave the usable list. The register's supply is smaller than the census implied and better characterized than it was.

Limitations, stated

• One events file is sampled per task, not all of them. A task whose sampled file carries stimulation is assumed to carry it throughout; a task whose sampled file carries none is reported as carrying none. For the three failures above the sampled file was checked against the others by row count, but not row by row.
• Criteria 2 and 3 — channel count and sampling rate — are not tested here. A deposit passing criterion 1 still has to pass those before a reading is issued.
• has_current reports whether any row carries a current value, not whether every stimulation trial does.