Review proposed by local AI
Ask a reviewer to compare source observation timestamps and clock settings.
Rule-linked evidence IDs: OMS-1001, CAR-1001
Human approval requiredBolt · Internal research · 11 October 2026
Every AI suggestion needs evidence. And a human decision.
A small technical proof built around fictional order-management, warehouse and carrier records. Real local-model inference chooses a restricted next review step; independent code checks prevent ungrounded event references from being accepted.
This is synthetic research, not an enterprise deployment. Every order, identifier, source and result below is fictional. The demonstration cannot send messages, change logistics systems, approve transactions or prove business savings.
63Local Python tests passed
4Fictional orders examined
3Model outputs accepted by evidence checks
0External system actions
Traceable example results
Read the review trail.
The model is Qwen 2.5 1.5B, run locally with Ollama. The rule engine detected three conditions, then the model selected a permitted reviewer step for each. A fourth order needed no model review.
Review proposed by local AI
Ask a reviewer to compare source observation timestamps and clock settings.
Rule-linked evidence IDs: OMS-1002
Human approval requiredNo alert matched
No predefined exception rule matched this fictional order; no model call was needed.
No automated actionReview proposed by local AI
Ask a reviewer to compare source observation timestamps and clock settings.
Rule-linked evidence IDs: OMS-1004
Human approval requiredWhat these results do — and do not — show
Bounded, not autonomous.
The local language model makes an advisory choice among predefined review steps, and the program validates its event citations against the triggering rule. The first unconstrained response in development referred to an irrelevant event ID and was rejected. The final three successful responses all suggested comparing timestamps, so this is not yet evidence of sophisticated differentiated AI diagnosis.
Independent customer measurements, production data connections, operator authentication, protected audit storage and commercial ROI remain unverified. The tests measure only the behavior of the supplied examples. The separate proprietary Bolt macOS application is not published as part of this illustration.
Evaluation proposals begin with a written description of a bounded operational problem and synthetic/non-sensitive examples. No production records, account access or confidential documents are needed for an initial enquiry.
Omar Baró · Unfire · Spain · Initial communication by email only.