FINDINGS

Across 234 Apparatus I plates with an ordinal EXAMINER I reading, the mean signed drift between the requested and the observed phase was +1.60: 176 plates read later than requested, 39 at the requested phase, 19 earlier. The apparatus photographs later moments than it is asked for. The remaining 30 readings were abstentions and are excluded from the arithmetic.

METHOD: Signed drift is index(observed) minus index(requested) across the seven ordered phases; the earliest complete EXAMINER I examination per plate; MULTIPLE-STATES and INDETERMINATE counted as abstentions.

EXAMINER I reported FRAGMENTATION on 156 of 264 Apparatus I plates, 59 percent, at a mean certainty of 96.

METHOD: All Apparatus I plates with a complete EXAMINER I examination; the earliest examination per plate; certainty as reported, unweighted.

Thirty Apparatus I calibration plates, imagination 0.00 to 0.39, were hand-labelled on 21 August 2026. The hand label matched the requested phase on 7 of 30 plates, 23 percent.

METHOD: Thirty calibration plates hand-labelled by the operator against the seven-phase vocabulary plus MULTIPLE-STATES and INDETERMINATE; agreement counted as hand label equal to requested phase. First disclosed in the Colophon, 21 August 2026.

Visitor calls on Apparatus I plates, binned by the archive's three imagination bands: OBSERVED, 9 of 85 calls ruled COULD NOT HAVE HAPPENED, 11 percent; THRESHOLD, 18 of 45, 40 percent; IMAGINED, 56 of 87, 64 percent. The share ruled impossible rises with the control.

METHOD: Call-weighted; the denominator is visitor plausibility calls, not plates. Bands are the archive's public filter bands over imaginedness. Bands under 25 calls would report INSUFFICIENT; none fell below it.

Thirty Apparatus I calibration plates, imagination 0.00 to 0.39, hand-labelled by the operator on 21 August 2026, were compared against two readings of the same plates. The examiner reading stored in the archive at that date, made by blind-plate-v2, matched the hand label on 6 of the 28 plates carrying one, 21 percent. The revised examiner, blind-plate-v3, matched it on 13 of 30, 43 percent, and on 12 of the same 28, 43 percent. Agreement with the human reading doubled. The concentration did not disperse, it moved: blind-plate-v2 reported FRAGMENTATION on 24 of 28 plates, 21 of them where the hand label did not, and blind-plate-v3 reported FRAGMENTATION on none of the 30 while reporting CORONET on 20, where the hand label read CORONET on 7. FINDING 002 reports EXAMINER I reading FRAGMENTATION on 59 percent of Apparatus I plates. The stored readings measured here are 28 of the readings behind that figure, all of them made by blind-plate-v2.

METHOD: Hand labels are the operator's own blind readings against the seven-phase vocabulary plus MULTIPLE-STATES and INDETERMINATE, recorded before either machine reading was consulted. The blind-plate-v3 readings come from a read-only probe run and are not stored examinations; they do not appear in the archive's examination history and no plate was re-examined to produce them. Two of the thirty plates carried no stored examiner reading and are excluded from the blind-plate-v2 figure; the shared-28 figures are reported alongside so the two examiners are compared on identical plates. Agreement is the hand label equal to the machine reading. All figures recomputed from the per-plate record at calibration-archive/2026-08-21-phase-calibration.json. The generator baseline of 23 percent on this same sample is FINDING 003.

Twenty-nine Apparatus II plates, imagination 0.00 to 0.39, hand-labelled by the operator on 25 August 2026, were compared against the phase each was asked to photograph. The hand label matched the requested phase on 19 of 29 plates, 66 percent, against 7 of 30, 23 percent, for the Apparatus I calibration plates reported in FINDING 003. The two sets differ in composition. Pre-impact is 10 of 29 plates here and 5 of 30 there, and it matched on all 10 against 2 of 5. Excluding pre-impact from both, the figures are 9 of 19, 47 percent, against 5 of 25, 20 percent. Emergence, the phase the generator revision was written to correct, matched on 1 of 6; five of the six were read as coronet, which is the overshoot the revision was intended to remove. Eleven of the thirty Apparatus I plates were hand-labelled MULTIPLE-STATES or INDETERMINATE. None of the twenty-nine Apparatus II plates were.

METHOD: The twenty-nine plates are every public Apparatus II exposure at imagination 0.40 or below at the time of computation, not a sample drawn from them: the sampler reported 29 eligible against a target of 30 and returned all of them. Hand labels are the operator's own blind readings against the seven-phase vocabulary plus MULTIPLE-STATES and INDETERMINATE, recorded without sight of the requested phase or the laboratory reading. Agreement is the hand label equal to the requested phase. An abstention can never equal a requested phase and is counted as a non-match, as it was in FINDING 003. The Apparatus I figures are recomputed here from the same per-plate record that produced FINDING 003 rather than carried over. Both eras were labelled by one person, four days apart, the second in knowledge that a generator revision had shipped. All figures recomputed from calibration-archive/2026-08-21-phase-calibration.json and calibration-archive/2026-08-25-phase-calibration-apparatus-ii.json.

Across 49 Apparatus II plates with an ordinal laboratory reading, the mean signed drift between the requested and the observed phase was +0.61: 21 plates read later than requested, 21 at the requested phase, 7 earlier. Recomputed on the same day across 236 Apparatus I plates, the figure is +1.59. The apparatus still photographs later moments than it is asked for, by less than half as many phases, and the single most common outcome is now no drift at all. A further 7 Apparatus II readings were abstentions and are excluded from the arithmetic.

METHOD: Signed drift is index(observed) minus index(requested) across the seven ordered phases, positive when the plate shows a later moment than was asked for; MULTIPLE-STATES and INDETERMINATE are abstentions and never enter the arithmetic. Values are those published in the archive export, recomputed from the record on every read. The generator and the examiner were both revised between the two eras, so this measures the apparatus as a whole and cannot separate their contributions; FINDING 006 measures the generator against hand labels and FINDING 005 the examiner. The Apparatus I figure here is recomputed from the current record and is not the figure published as FINDING 001, which stated the archive as it stood on 23 August 2026 and stands as recorded.