19% · live — The highest single ISO week of OpenRouter aggregate token volume in September 20

Probability history

Called at 40% on August 29, 2026. Not yet repriced. 19 days past deadline, awaiting adjudication.

How the number was built

Final method: stacked-fixed-weights (resolved n=74 < 80). Each rung is an independent estimate; the final is a stacked combination.

Stated (issue date)40%

Reference class (n=25)65%

Recalibrated49%

Feature model29%

Council8%

Final19%

Drivers

Reference class qtype:threshold runs 0.65 (n=25) vs stated 0.40.
Hierarchical calibration moves stated 0.40 to 0.49 (lens software, kappa_stated=80.0, kappa_lens=200.0).
Feature model at 0.29: conjuncts=1, named disclosure=False, horizon=32d.
Council of 5 at 0.08 with dissent.

Cautions

all 5 votes came from one vendor (openai); extremization disabled, persona diversity only

Council

5 personas on one model (openai/gpt-5.6-sol); other vendors were unavailable this run, so read this as one model's persona-averaged view, aggregated by log odds mean without extremization. Dissent: geopolitics-regulation-analyst at 0.32 vs grid-and-power-economist at 0.02: The supplied trigger omits the numerical bar, and no September 30 vintage of OpenRouter’s weekly volume series is provided. / The supplied prediction does not state a numeric token-volume bar, so the required comparison cannot be reproduced from OpenRouter’s September weekly data.

grid and power economist · openai/gpt-5.6-sol

The supplied prediction does not state a numeric token-volume bar, so the required comparison cannot be reproduced from OpenRouter’s September weekly data. Under the literal resolution rule and resolves-against-forecaster policy, that omission should produce a miss unless the archived weekly prediction contains an unambiguous written threshold. The forecaster’s 0.4 and historical threshold base rate do not overcome this document-level defect, especially after the deadline.

18%

semiconductor supply chain analyst · openai/gpt-5.6-sol

The deadline has passed, but neither the numeric bar nor OpenRouter’s 2026-09-30 historical vintage is supplied. The model-release digest is only an indirect demand signal and cannot establish weekly aggregate token volume. Although the forecaster’s threshold reference class has performed well, the relevant 0.4–0.5 calibration bin has only two observations. Because the written trigger cannot be tested from the provided record and ambiguity explicitly resolves against the forecaster, I assign a materially lower probability than the stated 0.4.

capital markets credit analyst · openai/gpt-5.6-sol

The written trigger supplies no numerical bar, so the September 2026 peak ISO-week token volume cannot be compared with the required threshold. Because scoring is explicitly against the written trigger and ambiguity resolves against the forecaster, this should fail regardless of the underlying OpenRouter series, absent omitted specification elsewhere in the controlling record.

superforecaster statistician · openai/gpt-5.6-sol

I use the lens:software base rate of 0.446. The decisive downward adjustment is that no numerical bar appears in the supplied question, while resolution explicitly depends on the forecast’s written trigger and ambiguity resolves against the forecaster. The digest also provides no September OpenRouter volume observations from the required 2026-09-30 vintage, so there is no verifiable evidence that every relevant ISO week remained below an identifiable threshold. The forecaster’s 0.4–0.5 calibration bin has only two cases and cannot overcome this specification problem.

32%

geopolitics regulation analyst · openai/gpt-5.6-sol

The supplied trigger omits the numerical bar, and no September 30 vintage of OpenRouter’s weekly volume series is provided. Under literal scoring and ambiguity against the forecaster, that materially reduces the chance of a successful resolution. A maximum-over-weeks condition also creates more upside-tail exposure than an average-volume test, while September’s open-model release activity could have produced a usage spike. The software-ledger base rate is only 0.446, and the favorable low-confidence calibration sample has just two observations.