Evidence · Methodology reports
The same codebase — no retraining, no domain-specific rebuilds — tested across cybersecurity, education, medicine, robotics, marine and navigation. Every report, recorded case and repository is listed here.
How we test
A Software-in-the-Loop test is an industry-standard approach for verifying AI behavior in simulation before physical deployment. For each domain, we first ask: does an accepted SIL standard exist? If yes, we run it. If not, we define the criteria, document them publicly, and hold ourselves to them.
Raw trial data (JSON reports, Criterion benchmark outputs) is available for independent verification on request — we don't hide numbers that don't look perfect, we explain them.
What is being tested, stated before the run.
The harness, the criteria and what counts as a pass.
Every trial reported, not the best one.
What the result does not show.
The reports
Marine · New · September 2026
A full ten-phase structure inspection with no human in the loop: an 8.871 m sweep at 98.7% of commanded rate, refusals predicted to within 0.1 s, every number scored against ground truth. Over a hundred defects, and the rules each one earned, published in full.
Opens in the browser · no registration · 15 pages
Marine · August 2026
Jean's decision layer held a 6-DOF underwater vehicle on station for 363 s against a 120 s criterion, median drift 0.042 m, and refused all five impossible tasks for the right reason. The calibration journey is published in full.
Opens in the browser · no registration · 9 pages
Cybersecurity · BethanyShell
CALDERA nation-state emulation under SIL methodology: 92.3% APT-class block rate, zero false positives across 110 benign requests, ~1.07ns per check. Includes repo access and reproduction instructions.
Opens in DocSend · quick email registration
Robotics · 7-DOF arm
Standard targets threaded 11/11, impossible ones refused 4/4 with specific reasons, zero false successes across all runs. Fine-precision threading is a documented boundary, not a hidden one.
Opens in DocSend · quick email registration
Education · Parlier, CA
A +22 point student proficiency shift alongside 7–10 hours per week returned to the teacher. Student data never left the classroom. The testimonial is on record.
Opens in DocSend
Medicine · NEJM
Seven consecutive New England Journal of Medicine diagnostic cases, answered on local hardware at roughly 27 seconds and 2 cents of compute per case. Each case is recorded.
YouTube playlist
Navigation · Oakland
5,490 live route nodes evaluated across Oakland road networks with zero LLM calls in the cognitive path. The full write-up is on the Ground page; raw data on request.
Defense: the five validation runs are published on the Jean for Defense page, with the full methodology in the cybersecurity report above.
Verify anything
The BETHANY evaluation suite, methodology and SIL environment are open-source and reproducible — clone it and run the red-team emulation on your own machine. 8,906 tests, every one of them yours to run.
Open the repository$ git clone https://github.com/AyindeRudolph/bethany-sil $ cd bethany-sil $ cargo test $ cargo test --test sil_run
Go further
Independent verification
Tell us which domain and what you want to check. We send the raw data, not a summary of it.
Request raw dataSends to ir@myasolutions.org · we respond personally