Every other module reads a number off the venue and applies a rule. This one asks gpt-5.6-luna what a question actually resolves at, without showing it the price, and bets only when the answer and the market disagree by 20% or more. Most of the time they agree, and agreement is not a trade — so most of this page is opinions the module formed and did not act on.
No market has been priced by the model yet. Every new binary market gets an estimate; this fills as they arrive.
Every market the model priced, with the argument it gave — including the ones it argued itself out of. A multi-choice market is priced one answer at a time and each answer is its own row, named after the question. On this module the reasoning is the strategy, so a decline without it is the one decline nobody can check.
| market | price | model | gap | outcome |
|---|---|---|---|---|
| Nothing priced yet. | ||||
Nothing judged yet.
| utc | version | market | bet | result | why |
|---|---|---|---|---|---|
| 08-24 12:36 | v14 | AI solves Millennium Prize Problem in September 2026? | NO 72 | settled -72 | model 0.02; edge -0.37; luna-medium price-blind says 0.02 vs market 0.39 (edge -0.37 >= 0.3): As of August 24, 2026, six Millennium Problems remain unsolved, while recent AI systems have solved smaller research problems and Anthropic’s Claude improved a Riemann-Hypothesis-related bound to 67.2% but explicitly did not prove |
1 trade, 72 staked, realised -72, open +0 if held / +0 if sold.
In the traces on hand the model priced 0
answers across 0 markets and the module acted on
0 of them (0%). The journal holds
1 trades for 72 mana. 1 of them settled for -72.
An estimate the model declines to give is treated as no opinion and never
as a cheap one.
Traces are diagnostic and rotate; the trading journal is the book above.
Built 2026-09-22 18:16Z.