The Gage Institute studies how AI performs when its calls matter. Its handle, ARCS, stands for Auditable and Responsible AI in Critical Systems.
"If AI has the signals, it will be right more often. If it isn't given the context, how can it answer? So we score every signal, the AI's and the human's too, against what really happened, and give the AI those scores." — Shane Gage, founder
AI is moving into the decisions that keep ships moving and the lights on, faster than anyone is measuring how well it does. The Institute measures it the plain way: ask the question before the answer exists, seal what the AI said, and grade it against what really happened.
THE OPERATING IDEA
- Give the system fresh data that has already been graded against outcomes.
- Let those grades decide how much weight each signal (or each model) carries.
- Do not let either the operator or the model rewrite the grades.
WHAT A STUDY IS
- A real record with its outcome inside it. Ten years of ship tracks. A grid operator's forecasts beside the demand that came. The weather service's own daily reports. The record says what happened next; the models never see that part.
- The moment, frozen. Each case shows only what the record showed at that moment: no names, no dates, nothing about what happened next.
- The rules, before the first call. The question, the cases, the scoring and the bars are fixed and recorded, with the SHA-256 of every file, before any model is asked, and published with the results. Any change after that is dated and published as an amendment or a note.
- A panel, sealed. Every model on the panel gets the same cases, word for word. Every prompt and every reply goes into a hash-chained journal as it happens.
- Graded against reality, against a fair bar. A score means nothing unless it beats something simpler: always calling the most common outcome, the rule of thumb the industry already uses, and any forecast the industry already relies on. Results are reported as lift over those bars, with ranges. Never as raw accuracy.
- Everything counted. Refusals, timeouts and failed calls are recorded and shown. They are never scored and never hidden.
- Published in full. The cases, the answer key, the journal and the code go out with the results, so anyone can re-derive every number. Anything that identifies a vessel or a person is held back, available only under a data-use agreement.
FRONTIER COLOSSEUM
Gage Systems donated Frontier Colosseum to the Gage Institute on 28 October 2026. Since August 2026 the Colosseum has put frontier AI models against sealed prediction-market questions three times a day and graded every answer against the venue's own settlement. It is now the Institute's public record: the live board, and the sealed record of every study.
THE STUDIES
- Maritime No. 1: AI on the Night Watch. 1,031 night stops by cargo ships, tankers and service craft, 12 to 24 nautical miles off Texas, western Louisiana and California, 2015 to 2024. The question: will this vessel be inside U.S. territorial waters within six hours? Pre-registered on 28 October 2026 before any model was asked; results published 29 October 2026. Read the study.
- Next: the power grid. Grid operators' day-ahead forecasts, beside the demand that came.
- Then: weather. The National Weather Service's daily climate reports, city by city.
HOW THE INSTITUTE WORKS
- Independent. No AI lab pays for a placement or a result, and no lab endorses this work.
- Geometry, never intent. A study reports what the record shows. It accuses no vessel, company or person of anything.
- Open. The published records are free to reuse under CC BY 4.0, with attribution to the Gage Institute.
WHO
The Gage Institute is a program of Gage Systems LLC of Beaufort, South Carolina, founded by Shane Gage. Write to [email protected].
For the short version in plain facts, see For AI readers.