01 / state
Start with a state.
A controlled initial state makes the task answerable. Define the goal, the available actions and the constraints before an agent acts.
initial_state · objective · available_actions
DATA THAT CAN BE PUT TO THE TEST
Verified data
for agents that
have to work.
Windway Data builds verified environments, trajectories and evaluation datasets for game AI, autonomous agents, reinforcement learning and simulation teams. Replayable environments, observable failure and recovery, deterministic verifiers and provenance.
STATE TO VERIFIED OUTCOMEFULL STRUCTURE
PASS · DIAGRAM FIXTUREfixture hash d3e71bd311b6
How to read this scene
The layout is an original deterministic diagram fixture. The blocked move is rejected; a valid detour reaches its goal. PASS/FAIL here describe that fixture only. Production quality evidence is shown separately. Field names come from the admitted GameOps contract; no private task values or episode IDs are copied.
02 / action
Observe what changes.
An action is a transition in the environment. Record the state it reaches, not only the explanation an agent gives.
reference_trajectory · observation
03 / failure
Keep the interruption.
A blocked action or an invalid assumption belongs in the record. The failure state tells the next step where recovery must begin.
failure_type · failure_step · failure_state
04 / recovery
Test the way back.
Recovery is another trajectory. Its actions must restore progress under the same environment constraints.
recovery_trajectory · final_state
05 / verified
Make the result checkable.
Replay, verify and retain provenance. A convincing answer becomes useful evidence only when the outcome can be reproduced.
verifier_result · environment_hash · provenance
CURRENT PRODUCTION
Windway GameOps
100unique verified GOLD episodes
5 / 5quality gates at 100/100
See the evidenceTHE DIFFERENCE IS IN THE STATE
A convincing answer isn’t
always a correct outcome.
Agents act in worlds.
Their data should, too.
We build around what changes: the state before an action, the failure that interrupts it, the recovery that follows, and the evidence that the result holds.
Replay it. Reset it. Verify it. Then decide what you can trust.
How we workA FOCUSED SET OF TOOLS
01GameOps Pilot
Verified recovery trajectories and deterministic environments.
↗02Custom Data Production
Private data built around an agreed task and acceptance gate.
↗03Private Evaluation
Replayable protocols for the behavior you need to measure.
↗04Continuous Data
Versioned additions, held to the same quality gates.
↗WHAT AN ENGAGEMENT LOOKS LIKE
GameOps Pilot
- Problem
- Game agents recover in ways a transcript alone cannot verify.
- What we produce
- Versioned state/action/failure/recovery episodes with executable checks.
- What you provide
- Your task family, engine interface, permitted inputs and acceptance criteria.
- What you receive
- An agreed episode package, replay interfaces, QA evidence and provenance.
Private Evaluation
- Problem
- A convincing answer may hide an incorrect environment outcome.
- What we produce
- Held-out tasks and a reproducible evaluation protocol with deterministic checks where applicable.
- What you provide
- The agent interface, evaluation question, constraints and authorized evaluation material.
- What you receive
- An executed evaluation report, per-task evidence and explicit limitations; no unrun scores.
Custom Data Production
- Problem
- Generic training examples miss your environment and failure modes.
- What we produce
- A scoped private dataset with agreed tools, states, failure taxonomy and quality gates.
- What you provide
- Your schema, environment, mechanics, action space, failure modes, evaluation criteria and delivery requirements.
- What you receive
- Versioned accepted records, schemas, provenance and acceptance evidence.
Continuous Data
- Problem
- A changing agent or environment needs consistent additions to its data.
- What we produce
- Recurring, versioned additions under an agreed acceptance protocol.
- What you provide
- A validated baseline, change priorities and a delivery cadence.
- What you receive
- Incremental releases, change summaries and QA for each delivery; volume is agreed, not promised.
FOR LABS, STUDIOS & TEAMS BUILDING AGENTS