reliax

Per-decision reliability envelopes for AI systems.

Status: placeholder release. This version reserves the package name and contains no client API yet. Watch reliax-python for the SDK.

Reliax wraps any scoring model and assesses each prediction at inference time, without touching the model's weights or retraining it. Every decision carries a formal coverage guarantee, a calibrated risk bracket, a drift verdict, a routing decision under a policy you own, and a hash-chained audit record an auditor can replay.

What is guaranteed

Three components are theorems, and routing runs on them:

What is not

Two components are signals, labelled as signals, and never enter a routing rule: the triage rank, and the per-cell calibration opinion. They are monitors, not routers.

Coverage is a property of the procedure over exchangeable data, not a probability about any individual prediction. Distribution-free per-instance conditional coverage is provably unattainable, and we do not claim it.

Evidence, including the negative results

Measured on public credit data and reproducible end to end at reliax-evaluation. Every published number is regenerated by a runner script in that repository; nothing is hand-typed.

The unflattering results are in the same README as the good ones. Under severe distribution shift our fusion score performs worse than random referral, and it says so there rather than in a footnote.

Licence

Apache-2.0. The client SDK, the certificate specification and the verifier are permissively licensed, permanently — an auditor must be able to verify a Reliax certificate without anyone's permission.

The reliability engine is licensed separately and is free for non-production use. See reliax.io.