TIA Run: a game driven by TIA

1. The road moves at a fixed rate and never waits for the API. A segment with no answer yet means the car holds its lane.
2. So the browser asks ahead: each call asks about K segments the car reaches only after the measured round-trip p95 plus 0.5 s. Every answer is tied to its own segment; one that lands late is dropped and counted.
3. TIA Run has its own call lane (30 calls a minute per IP), and each call carries 9 or 20 segments, inside the 2,000-token limit per call. Only these two K values reached at least 95% safe lanes on the held-out test, so the game sends no other. Speed rises over 8 levels of 24 segments each, up to 0.32 s per segment; the end of level 8 is the finish line, and the page then points to the table of the race.
4. Each anti-bot check gives 500 calls. When they run out the road stops (no car drives blind) and asks “Keep racing?”; press it to pass the check again and go on. Restarting costs no check while calls are left. Each IP has 2,000 calls a day, the whole site 20,000.
5. On the road, the faded cars stand ahead of or behind TIA by their own count of segments safe and on time in this race, in the same order as the standings strip; that is only how it is drawn, every car covers the same distance. Each faded car takes the lane its right answer points to, and crashes or slips when its answer came late, as the numbers in the table say.

The approach follows “Real-Time Reasoning Agents in Evolving Environments” (Wen, Ye, Zhang, Yang, Zhu; arXiv 2511.04898): the world keeps moving while the model thinks, and a step with no action in time takes a default action.

• No decision is faked. The car follows only TIA’s answers: no reflex, no pathfinding, no corrected answers.
• Near-equal probabilities mean the model is unsure. We print the numbers exactly as returned, with no rounding for looks and no edits.
• The road is random, but every segment always has at least one lane with no obstacle.
• The demo calls the currently served TIA build through the System One API. TIA 1.5 remains a local experimental candidate.

TIA: real answers, the round trip measured in your browser, K segments per call. Every other car is a simulation: it is handed the right answer, one question per request, one request after another, the way JevBench v1.4.2.2 measures; each request’s delay is drawn to match the published p50 and p95. Every car sees the same road ahead (30 seconds of driving).