What the policy sees
Actor-observable structured state: visible balls and projectiles, visible shooter/global state and level geometry. Tunnel-hidden balls and privileged debug state are excluded. Rendered pixels never enter V1.
aispeedrun.ai · evidence release 001
AlphaZuma V1
Fourteen individual-level clears from learned policies, selected only from frozen blind evaluations and re-executed deterministically for the videos below.
VERIFIED · 2026-08-13
Category boundary: these are structured-state, deterministic-simulator, native-tick records. They are not original-client RTA runs and are not claims against the official Speedrun.com PC leaderboard. Human times are shown only as a cross-domain numerical reference.
verified IL runs
14
watchable videos
14
numerical leads*
4
official-client WR claims
0
Run conditions
The conditions are part of the record, not fine print. Every video repeats the essential category boundary in-frame.
What the policy does
Deterministic argmax over wait, fire or swap plus one of 180 aim bins. The policy is an entity 1D CNN with attention and an MLP head.
Human-level input envelope
120 ms reaction delay, 1080°/s maximum aim speed, 18,000°/s² maximum acceleration and a 50 ms minimum button interval.
How time is measured
Native simulator ticks at 100 Hz, starting at environment reset and ending at the terminal win state. This clock does not share the original executable's RTA start/end boundaries.
How a run was selected
For 13 levels: fastest win among 16 unseen paired seeds for each of two frozen checkpoints (512 total attempts across 16 tested levels). Adv 2 used a separately preregistered 11-checkpoint × 32-seed blind matrix.
How video was accepted
Each selected attempt was rerun without rendering and again with rendering. Outcome, tick, score, trajectory hash, decision count and action counts all had to match the frozen blind receipt.
Cross-domain numerical leads
*The simulator number is lower than the currently displayed human IL number. This is interesting context, not an official head-to-head record.
Loading evidence…
All 14 records
Every row binds a checkpoint hash, blind seed, terminal tick, score, trajectory hash, accepted video and poster.
| level | AI time | human reference | blind result | model | seed | evidence |
|---|---|---|---|---|---|---|
| Loading evidence… | ||||||
Record videos
Simulator visualizations at 50 fps. Videos use preload="none" so the full gallery does not download until requested.
Loading videos…
Audit trail
The public JSON is canonical for the values displayed on this page.
Morning paired blind
- 2 frozen models × 16 levels × 16 paired unseen seeds
- 512 completed attempts; all aggregate validation checks passed
- 13 published fastest winning attempts
- Aggregate SHA-256:
b82e0175…8621c8
Independent Adv 2 blind
- 11 frozen checkpoints × 32 unseen seeds = 352 attempts
- 11.10 s candidate selected by the preregistered ranking
- Three additional candidate replays were identical
- Evaluation SHA-256:
8af20964…405e9
Download the complete public evidence manifest →
Manifest SHA-256 sidecar →