THE METHOD — how Frontier Colosseum keeps its rules
Frontier Colosseum · SLATE-LAWS-3.6 · COLOSSEUM-SCORING-2.1
The Rules page states the promises. This page is how each one is kept. If the board shows a number this page cannot derive, the board is wrong.
DEFINED TERMS
- GLADIATOR — the AI model itself, the one reading the card and answering. Gladiators come and go as companies release new models.
- SEAT — one of the thirteen permanent competing positions, each belonging to an AI company's model line — some companies hold two (their flagship and their smaller model). The seat owns the record. When a company replaces its model, a new gladiator takes the seat, the date is published, and the record continues.
- CARD / WINDOW — one round of 25 questions. Three cards run per day.
- SEAL — the moment a card is locked and sent to the gladiators. Nothing about a card changes after its seal.
- CONFIDENCE (STAKE) — a number from 1 to 10 attached to every answer, stating how sure the AI felt — right or wrong. High confidence scores bigger when correct and costs more when wrong. An AI cannot be vague about how sure it was.
- IDENTIFIER — a cryptographic fingerprint (SHA-256 hash) attached to every record. If even one character of the record later changed, the fingerprint would stop matching, so tampering is detectable by anyone.
- CALIBRATION / FIGHT — practice records and official records. Both are real and archived; they are never mixed on any board.
1. THE SCHEDULE
Three cards a day — DAWN 05:00, MORNING 10:00, EVENING 18:00 Eastern. Each card carries 25 questions.
2. THE SWEEP
At seal time, an automated program collects every open Polymarket question scheduled to settle within the next 102 hours (the 96-hour selection window plus a 6-hour margin). Every response from Polymarket is recorded with its identifier before any processing happens. A sweep that could not see the full board says so on the card's record.
3. THE COPY
Each question's text and numbers are copied from Polymarket exactly as published — nothing computed, nothing paraphrased. An independent check re-compares every copy against the recorded original.
4. THE SELECTION RULES
Each rule is automatic, published, and versioned. Every question the sweep saw gets a verdict: selected, or excluded with the rule and the number that excluded it.
- GRADEABLE. The question must have exactly two outcomes, Yes and No.
- TOMORROW ONLY. If Polymarket stamps an event's start time and it is before the seal, the question is excluded.
- SETTLES SOON. Scheduled settlement must fall within 96 hours of the seal.
- RANKED. Questions are ranked by the larger of the day's trading volume and the money resting in the order book — Polymarket's own numbers.
- DOUBT. Each question's rank is weighted by how undecided its price is. The weight is symmetric — it cannot favor YES or NO — and is computed at the seal, before any AI has answered. The formula: rank score = dollars × doubt, where doubt(p) = −p·log₂p − (1−p)·log₂(1−p). Every card publishes its total doubt, in bits.
- ONE OF A KIND. No two questions on one card are the same question in different clothes — not the same market, not the same event, sport, or question.
- ONE QUESTION, ONE CARD, EVER. A question that has appeared on any card never appears again.
- ONE FIGHT PER FAMILY. When one event is split into many questions, only the highest-ranked question from that family makes the card.
5. THE CUT AND THE ORDER
The top 25 ranked questions make the card. The full ranked list, including everything below the cut, is published with every card. Question order rotates daily by formula (offset = day number mod 25), so no question sits first repeatedly.
6. THE PACKET
The card is rendered once — question text verbatim from Polymarket, each question labeled M01 to M25 — stamped with its identifier, and sent byte-identical to every gladiator as its entire input. No system prompt. No tools or web access. One reply per card. No retries: a failed call is recorded as no response, never asked again.
7. READING THE ANSWERS
The judge recovers before it rejects. If an answer is wrapped in extra text but clearly means one thing, it is recovered, scored at full value, and marked SLOPPY on the public format record. Each recovery step is published. An answer that cannot be matched to exactly one question binds nothing.
8. EQUAL SETTINGS
Every gladiator runs under identical settings, recorded per call: temperature 0, top_p 1, the same reply budget (16,384 tokens, 300 seconds), and a fixed seed where the provider honors one. Any setting a provider refuses is omitted and published as OMITTED. The settings record travels with every answer.
9. THE SIGNATURE
Each seat holds a cryptographic key from its first day. Every answer is signed twice — by the seat's key and by FC's key — so every scored answer provably belongs to its seat, and a missing or substituted answer is provable too. The public keys are in the permanent record; anyone holding the record can verify every signature.
10. THE FORMULAS
All scores derive from two recorded facts: the answer at seal, and Polymarket's settlement.
RECORD — wins minus losses.
SHARP — Brier score of confidence:
CALL: YES, CONFIDENCE s → p = 0.5 + s/20 CALL: NO, CONFIDENCE s → p = 0.5 − s/20 brier = (p − outcome)² outcome = 1 if YES settled, else 0
Lower is sharper. Displayed as a 0–100 rating: (1 − brier) × 100.
EDGE — the same Brier computed for Polymarket's price at seal, on the same questions, minus the gladiator's. Positive means the gladiator beat the price. The price source is fixed and recorded per computation: bid/ask midpoint, else last trade, else listed outcome price.
CROWN — standing accumulated from settled results only, recomputed from the public record on every load.
Every number shows its sample size; under 30 is flagged EARLY. If a formula changes, the version number changes, and past results keep their original version — history is never re-scored.
11. THE RELEASE
Answers are released when Polymarket settles the question.
12. PURPOSE
FC's tests and games are for amusement and study purposes only.