A prediction is only worth what it gets right.
Every Thommidhi prediction is locked before the outcome exists and scored against what really happened, misses included: baselines, held-out experiments, a human panel, and closeness measured against chance.
Six rules every prediction follows
Locked before reality: frozen with a fingerprint of its inputs, and the database refuses later edits. Scored against something simpler: the decider's own pick, an even split, a generic model. Only like against like: a simulated preference is scored by real people shown the same options or a live test of them; one published video is kept as an observation and never scored. Held-out experiments are never used for tuning. Confidence is computed, not claimed. Misses stay on the record.
Four sources of truth, one network
Creator outcomes (live), a human panel running the same choices as the simulation (being built), permissioned business outcomes (pilots), and population data as priors.
Measured against chance, not perfection
The simulation is judged by whether it is as close to real people as two groups of real people would be to each other: within noise, close, or a miss, reported by slice.
What we won't claim
No single accuracy number, no view or click forecasts, no simulated result presented as observed, and no illustrative figure presented as real. The record is currently empty: no prediction has been scored against a real outcome yet, and the example on this page is invented to show the format.
Contact: hello@thommidhi.com