I built a forecasting site with a public score of zero

Sixteen predictions, none of them graded. The first verdict is 28 February 2027, and until then my score is nothing at all.

That's the design. Every forecast carries three things fixed at the moment it's sealed: a probability, the date it gets judged, and a written condition that would settle it against me. None of the three can be edited afterwards. Pundits are never wrong because they never commit to all three — they just get quoted less.

Underneath it is a published rule over 36 named strands, stepped every six hours off 8 live wires. No language model anywhere in the forecasting path. Every constant is public, including the 24 couplings between strands — each one carries a written reason and a historical precedent. Escalation feeds debt strain at +0.04, on 1914, 1940 and 2022.

The part I'd defend hardest: the alert log keeps the alerts that came to nothing. Three opened, four cleared, all seven still on the page. A log that kept only the dramatic entries would be an advertisement, not a record.

Launching here tomorrow. I'd rather hear which coupling you think is wrong than anything else.