Laboratory
A long-running argument
with our own results.
A standing laboratory, a growing family of models, and a habit of treating our own measurements as the first thing to attack.
Active lines
Explainability at machine speed
A decision taken in microseconds that a human can still audit a year later, with the signal, the weight and the threshold intact.
Adversarial resilience
Models measured against opponents who have been told exactly how the model works. Anything that only survives obscurity is not counted as a result.
Decision provenance
The chain from raw signal to enforced outcome, kept whole and tamper-evident, so an incident can be reconstructed rather than recounted.
False-positive economics
The cost of being wrong, measured on named corpora of legitimate software — backup agents, compilers, encryption tools — rather than estimated.
Bounded autonomy
How far a system may act before a human is required, and how that boundary is enforced in code instead of described in policy.
Sovereign inference
Training and inference that stay inside a perimeter, including for institutions whose networks will never reach the internet.
Standard of proof
A number, on a named corpus, or nothing.
A model score is not a result. A result is a measurement someone else could repeat, on a corpus we are willing to name, with the failures counted alongside the successes.