Gameplay data with intent, actions, and provenance
Each Kifu episode centers on a real gameplay decision. It includes synchronized video, game-focused keyboard and mouse inputs, player-reported context, and item-level provenance.
Standard file formats make the data easy to inspect before you integrate it.
Understand what happened and why
Gameplay video shows the result of a decision. Kifu adds the player's goal, observations, alternatives, confidence, and expected outcome.
Every episode also records how the moment was selected, how it was reviewed, and how the capture was produced.
What each episode includes
A 30 to 120 second window of unbroken gameplay, anchored on one focal decision. Everything is inspectable with zero custom tooling.
A 30 to 120 second gameplay clip recorded by the Kifu recorder: 30 fps constant-frame-rate H.264 at 1080p (720p on integrated-GPU machines). Capture method, resolution, and measured frame statistics are stated in every manifest. No audio. The system cursor is visible in frames, exactly what the player saw.
Game-focused keyboard and mouse events timestamped on the same clock as the video. Parquet supports large-scale analysis; JSONL makes initial inspection easy with no tooling.
The player's own account of what happened and why it mattered, always, plus archetype-fitted structure (candidates weighed, deciding factors, confidence, expected versus actual) where the moment supports it. Each record includes the question, response format, stage, and source (human or model-generated).
Item identity, provenance, capture details, synchronization measurements, selection method, and quality scores. Everything you'd filter a corpus on, top-level.
Versioned JSON Schemas and standardized question IDs, checked by the Kifu validator, so heterogeneous corpora stay aggregable.
Quality controls you can inspect
Synchronization
Every manifest states the synchronization error, the method behind it, and whether it was measured on that machine or inherited from our reference capture rig, so you can tell a per-session measurement from an architecture-level one. Game-side input latency is documented separately, as it should be.
Annotation sequence
Every answer carries its own timestamp and the stage its question was written for, so you can check for yourself whether intent was recorded before the player reviewed the outcome. Staged capture is a design goal, not a claim we make about past alpha episodes. "I don't remember" is a first-class answer.
Selection method
Every episode identifies how its focal moment was chosen, whether tagged by the player or selected by an operator, so selection bias is a filterable fact, not a footnote.
Schema validation
Deliveries are checked against versioned schemas so datasets remain consistent as the corpus grows.
Evaluate the data before you commit
Tell us the game, task, or model behavior you want to evaluate. We will prepare a relevant sample when the current corpus is a match.