ReBlink

Early ReBlink research on using V-JEPA 2 surprise to find temporal glitches in gameplay footage without labeled bug training, with VideoGlitchBench as the next evaluation.

A design exploration of how language could open Total War's campaign and battle controls without lowering its strategic bar, then support policies, standing orders, doctrines, command capacity, and counterplay.

A design exploration of how language could open Civilization's control surface without lowering its strategic bar, then support policies, governors, commanders, and counterplay.

A design exploration of how language could open the control surface in Age of Empires IV while preserving difficulty and creating new strategy through policies, standing orders, and costed Overseers.

Language control can expand who can play a hard game, which devices can support it, and how a studio should assess its potential return.

ReBlink Matter combines studio material library management with search, generation, refinement, and engine-native shader generation.

Playthrough captures reasoning traces with the play so a designer or QA lead can tell whether a surprising result is a real rule change, a bad line, unusual agent reasoning, or a contaminated run.

World-intelligence companies treat game engines as a rich labeled world, but telemetry is not automatically a training set. Train on accepted, attributable transitions from an authoritative engine.

A practical setup for training a smaller model on actions the real game engine has already accepted.

What a studio has to expose in the real game before Playthrough can start: a legal-action door in the engine, why we do not start with pixels and clicks, and what that interface opens later.