Substrate
Second Thought does not read a model’s private state. Token Factory returns observable inference. The calibration gap is computed from the packet, the probes, and the disagreement in front of us. Metacognition here means changing the reasoning strategy before the error is allowed to become the answer.
30B MoE · 3B active
Everyday cognition. The correct call when uncertainty, evidence, and complexity are already settled.
120B MoE · 12B active
Middle tier. Stronger reasoning when the calibration gap is opening, before Ultra is justified.
550B MoE · 55B active
Expensive cognition. Invoked only after the problem is framed and still unresolved.
Epistemic acquisition
Epistemic acquisition. Called only when surviving hypotheses cannot be separated on the current packet.
Contradiction between evidence, support coverage, untested assumptions, instability under leave-one-out probes, evidence age, hypothesis competition, task risk, failed verification, and reasoning that repeats without information gain. If a signal cannot be computed from the packet and the trace, it is not in the gap.
OpenAI-compatible chat completions at api.tokenfactory.nebius.com. Status: checking.
Search is a regulatory action, not a sidebar. Status: checking.
The evaluation lab is the job: the same battery, the same metrics, run off the request path when the trial count or the live model calls no longer belong in a browser. Lightning when the state is already calm. Super when the gap is opening. Ultra when contradiction, incomplete evidence, unresolved competition, and consequence line up. Then back down. That allocation is the feedback the stack is there to measure.