Verifiable edge-AI primitives. No cloud. No mocks. Run it yourself.
Most AI projects ask you to trust their claims. This one ships the tests and dares you to check. Every headline number is computed live, on-device, from stdlib + numpy — and a regression suite proves each claim holds.
python3 sovereign_demo.py # live pipeline, ~40s
python3 sovereign_ops/evals/eval_harness.py # the proofs: 5/5 evals
Author: Chad Edward Holland · Runs on a phone (Termux, Python 3, numpy) · 2026
Three correct implementations of hard problems, each with a test that proves it:
-
Causal inference — Pearl's backdoor criterion and d-separation, from scratch. Handles the collider trap that trips up most implementations: it refuses to adjust for a collider, because doing so manufactures a correlation that isn't there (Berkson's paradox). Deterministic across runs.
-
Conformal prediction — distribution-free prediction intervals. A 90% interval contains the truth ~90% of the time, verified under both Gaussian and non-Gaussian noise. Nonconformity compares a model's output to the true label, never a label to itself (a subtle bug this suite guards against).
-
Thermal-gated compute — compute budget shrinks under device heat. The point that matters: when the budget is too small to find a valid answer, the system returns "no valid answer, escalate" rather than silently guessing wrong. Honest failure under load is the feature.
eval_harness.py is a guard, not a demo. Every eval is something that was
actually verified true-or-false during development, including a regression
test that catches the conformal self-comparison bug if anyone reintroduces it.
Nothing here is marked "done" without a test that runs. Current: 5/5 passing.
These are correct implementations of established methods (Pearl 2009; Vovk et al. conformal prediction), not novel algorithms. The value is that they're dependency-free, run on a phone, are honestly scoped, and are backed by tests you can execute. No external-dataset validation, no production claims, no benchmarks against other systems. Prototype, self-reported, single device — and every claim is reproducible with one command. See STATUS.md for the full verified/unverified breakdown.
Vincit Omnia Veritas — truth conquers, but only when checked.