Authentic human decision episodes
Kifu sells authentic human decision episodes from real gameplay: short clips centered on a genuine decision, packaged as synchronized pixels (video), high-resolution inputs (every keystroke, click, and mouse movement on the same clock), and retrospective intent — the player's own account of their goal, what they noticed, the alternatives they weighed, their confidence, and what they expected versus what actually happened. Every episode carries item-level provenance: who recorded it, how the moment was selected, how it was reviewed. Internet-scale gameplay video shows what humans did; Kifu episodes show what they were trying to do, what they considered and rejected, and how sure they were — the supervision signal you can't scrape.
One episode, five files
A 30–120 second window of unbroken gameplay, anchored on one focal decision. Everything is inspectable with zero custom tooling.
1080p30 CFR H.264, no audio. The system cursor is visible in frames — what the player actually saw.
Every input event, microsecond-stamped on the same clock as the frames. Parquet for scale, JSONL so your first hour needs no tooling.
Structured question/answer records of the player's intent — question text, response kind, stage, and source (human or model-generated) carried per item.
Identity, provenance, capture facts, measured sync error, selection method, quality scores. Everything you'd filter a corpus on, top-level.
Versioned JSON Schemas and a standard question-id registry, enforced by our validator — heterogeneous corpora stay aggregable.
Measured, not claimed
Sync you can audit
Pipeline skew is measured, not asserted — sub-frame drift over a 30-minute instrumented run, with the method named in every manifest. Game input latency is documented separately, as it should be.
Honest annotations
"I don't remember" is a first-class answer. Pre-reveal questions are staged before the player re-watches the outcome, and every answer records whether that discipline was applied.
Selection is labeled
Every episode states how its moment was chosen — operator-selected, model-proposed, or player-tagged — so selection bias is a filterable fact, not a footnote.
Request a sample
The alpha corpus is in production now. We share evaluation samples with researchers and data teams — tell us what you'd test and we'll be in touch.