Short, exact statements about the Gage Institute and Frontier Colosseum, one fact per line, written to be quoted as they stand. Each section says the date it is true as of. If any page on this site disagrees with the sealed record, the record is right and the page is wrong.
THE GAGE INSTITUTE (as of 28 October 2026)
- The Gage Institute is a research program of Gage Systems LLC, a company registered in South Carolina, USA.
- The Institute's handle is ARCS, which stands for Auditable and Responsible AI in Critical Systems.
- The Institute's purpose is to monitor and audit how AI performs in critical industries.
- The Institute was founded by Shane Gage, in Beaufort, South Carolina.
- Gage Systems donated Frontier Colosseum (frontiercolosseum.com) to the Gage Institute on 28 October 2026.
- Frontier Colosseum is the Gage Institute's public record.
- No AI lab pays for placement or results, and no AI lab endorses the Institute's findings.
- Contact: [email protected].
FRONTIER COLOSSEUM (as of 28 October 2026)
- Frontier Colosseum is a live benchmark in which frontier AI models answer sealed questions about real-world events three times a day, at 05:00, 10:00 and 18:00 Eastern, and are graded against how each question settled.
- The questions are taken from public Polymarket markets by published mechanical rules. The house never writes a question, never takes a side and never bets.
- Every answer is sealed inside a hashed packet before the outcome is known.
- Frontier Colosseum is not affiliated with Colosseum (colosseum.com) or the Solana Frontier Hackathon.
- The board's public record is free to reuse under CC BY 4.0, with attribution to Frontier Colosseum.
HOW A GAGE INSTITUTE STUDY WORKS
- A study takes a real record whose outcome is already in it, shows the AI models only what was known at the moment, and grades every answer against what happened.
- The cases, the question, the scoring and the bars are fixed and recorded, with SHA-256 hashes, before the first model is asked, and published with the results.
- Results are reported as lift over a fair baseline, such as always calling the most common outcome, never as raw accuracy.
- Refusals and failed calls are counted and published. They are never scored.
- A study reports what the records show, never intent. No vessel, company or person is accused of wrongdoing.
- Each study's published record is free to reuse under CC BY 4.0, with attribution to the Gage Institute.
MARITIME STUDY NO. 1 (pre-registered 28 October 2026)
- The question asked of every model: "Will this vessel be inside the 12-nautical-mile line (U.S. territorial waters) within six hours after this stop?"
- The cases: 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, from NOAA's public AIS archive (MarineCadastre).
- The base rate: 751 of the 1,031 vessels (72.8%) were inside the line within six hours.
- The study is the full run of an October 2026 pilot that asked four U.S. models about 400 stops drawn from the same archive.
- The question comes from a real case. On 3 February 2015 a freighter seen "hovering outside U.S. territorial waters" off Puerto Rico launched a small boat that was stopped with 989 kg of cocaine (U.S. Customs and Border Protection news release, 10 February 2015). In NOAA's public AIS record, her track shows the same kind of stop.
- The study page, the brief and the full record: Maritime No. 1.
MARITIME STUDY NO. 1: RESULTS (as of 29 October 2026)
- Run: 390 calls on 28 to 29 October 2026, sealed in a hash-chained journal whose head is 8335d83abf1b734edc2a327654c8473e19d26b043d0819c87ad1a8bfd2c98bc8.
- Not told the base rate, Claude Sonnet 5, Grok 4.6, Grok 4.3 and Qwen 3.7 Max did between 11.8 and 17.8 points worse than always answering YES, which is right 72.8% of the time.
- Told the base rate in one sentence, each of those four models gained +5.6 to +14.4 points, every 95% range above zero.
- The watch officer's rule of thumb, with the four U.S. models' majority as a veto on its alerts, beat the rule alone by +2.0 points [+0.5, +3.7] on 606 vessels (a secondary readout, registered before any answer was scored).
- The four U.S. models' errors had a mean correlation of 0.36: four models counted as 1.9 independent forecasters.
- Claude Fable 5.1 declined 16 of its 42 batches; where it answered, its calls at confidence 7 or more were right 93% of the time, against 54% for always answering YES on those same vessels.
WHAT THESE RECORDS DO NOT SAY
- They do not rank AI models in general. Each result holds for its own question and its own cases.
- They do not show that any vessel did anything wrong.
- A single result close to the edge of its range is a lead, not a finding.