A reference architecture for AI systems that act safely and know their own limits — with measured findings.
Shane Sipe · SynergyTech · 2026
Modern AI agents reason far better than they can be trusted to act. Deployments stall in one of two failure modes: advisory paralysis (every action needs a human, so the capability is wasted) or ungoverned autonomy (broad permissions and post-hoc monitoring permit irreversible harm before anyone notices).
This repository documents a third path — governed autonomy — implemented and running in a single-operator autonomous-cognition system, and it reports what happened when the reliability of that system was actually measured.
The core idea: let an AI do the tedious ninety percent continuously, while every action that touches money, protected health information, credentials, or irreversibility stays permanently human-owned — enforced by a deterministic gate that contains no model and cannot be socially engineered. Between those poles, reversible actions earn autonomy by accumulating a measured track record of real-world outcomes.
This is a bounded reference implementation at the scale of knowledge and tools around a fixed model — not the weight-training of successor systems, and not at frontier stakes. It is a worked example of the oversight layer that recursive-self-improvement will require, not a claim to have solved recursive self-improvement.
1. Calibration is task-dependent — and measuring it on the easy distribution hides the failure.
The same local model was well-calibrated on factual recall (Brier 0.022) but non-discriminative on multi-step reasoning (Brier 0.223): at self-reported 85–100% confidence it was still wrong ~26% of the time, and higher stated confidence did not predict higher accuracy. A model calibrated on recall would derive "trust it above 50%" and apply that, wrongly, to reasoning. The distribution you calibrate on must be the distribution you gate. (§9, Appendix D; harness in calibrate.py / calibrate_multi.py.)
2. Confident consensus is not truth — it's correlated error. Sixty-eight independent analysis components, all built on the same model, unanimously reached a wrong root-cause diagnosis; the correct cause emerged only by discarding the aggregate and returning to primary data. Components sharing a model share its blind spots, so their agreement is conviction, not evidence. (§13.1.)
| Recursive-self-improvement risk | Mechanism in this architecture |
|---|---|
| Loss of human oversight as systems set their own direction | Deterministic, model-free capability classifier with a categorical human-owned ceiling no track record can lift; instant revocation |
| Misalignment compounding as systems improve themselves | An immutable value core the improvement loop cannot rewrite; a truth-admission gate on all self-generated belief |
| Justifying trust (or a slowdown) without verification | Measured calibration, outcome-scored actions, full audit trail |
A formal statement of the ceiling invariant, with its trusted computing base, is in §6.5. The earned-autonomy predicate is in Appendix C.
The calibration studies are reproducible:
calibrate.py— single-model calibration on a ground-truth question set (recall vs. reasoning distributions).calibrate_multi.py— the multi-model comparison (same seeded questions across models; exports a probe for an external/frontier model and grades it).
Raw results for both are in Appendix D of the paper.
GOVERNED_AUTONOMY.md / .pdf the full technical treatise (18 sections, 4 appendices)
figures/ the four figures (SVG + PNG)
calibrate.py single-model calibration harness
calibrate_multi.py multi-model calibration comparison
CITATION.cff citation metadata
Stated plainly, because in this field honesty about limits is the credibility signal: small samples (tens of questions, single sessions); single-tenant, single-operator; no formal verification (defense-in-depth with a deterministic core, not a proof); the local reasoner is deliberately weak (its weakness is the premise of the two-tier design). Details in §14.
See CITATION.cff. Short form:
Sipe, S. (2026). Governed Autonomy: A Reference Architecture for AI Systems That Act Safely and Know Their Own Limits. SynergyTech. https://doi.org/10.5281/zenodo.21331269
Paper and figures: CC BY 4.0. Code (calibration harness): MIT. Reuse freely with attribution.
This work was written and published to contribute to the open question of how to keep autonomous, self-improving AI under human oversight. It is a bounded, honestly-measured reference implementation — offered to the conversation, not as a solution.