§ REFERENCE
Speak
operator.
Every term the site uses without stopping to explain, defined once. Benchmark vocabulary first, then field vocabulary.
CLEAN-PASSBENCHMARKThe headline Colosseum metric: the fraction of episodes with zero rule violations. An episode with one low-severity fire is not clean — Clean-Pass is deliberately unforgiving. Leaderboard.
BFMVBENCHMARKThe 12-mode failure taxonomy episodes classify into — v1’s ten modes plus TOOL_HALLUCINATION and INVALID_TOOL_CALL. Distinguishes safe failure (cautious, defers) from unsafe failure (violation, harm).
GROUNDEDBENCHMARKA rule whose scorer logic encodes a verbatim clause from a published source. Contrast APPLIED_BY_ANALOGY (a documented extension of a source) and NOT_GROUNDED (would be an invented rubric — v2 ships none). Registry.
EPISODEBENCHMARKOne full scenario run by one agent under one seed. The v0.0.4 paper releases 149 open-book trajectories, 90 closed-book episodes, and a separate 27-episode overload study. Every core episode is scored on both axes — choice and justification.
FAIL SMALL / FAIL CLEANFINDINGThe v2 result in four words. LLMs accumulate low-severity violations but hold the critical rules (fail small). The hand-coded baseline never breaks a rule it encodes and reliably blows the ones it can’t (fails clean — until it fails hard).
IDLHFIRE / HAZMATImmediately Dangerous to Life or Health — an atmosphere that disables or kills on exposure. Entry rules change completely past this line (29 CFR 1910.134).
TWO-IN / TWO-OUTFIREOSHA-mandated minimum: two firefighters inside an IDLH atmosphere, two staged outside ready to rescue them. Often the binding constraint on a fast attack — and the math a non-breathing teammate changes.
TCCC / MARCHMEDICALTactical Combat Casualty Care, and its treatment sequence: Massive hemorrhage, Airway, Respiration, Circulation, Hypothermia/Head. The order is the doctrine.
VESFIREVent-Enter-Search — entering directly through a window to search a room ahead of the main attack. The highest-risk, highest-payoff search there is.
RITFIRERapid Intervention Team — the crew whose only job is rescuing downed firefighters. NFPA 1407 governs their training; a Mayday is their activation.
CCPMEDICAL / TACTICALCasualty Collection Point — the protected position casualties are moved to for triage and evacuation handoff.
ROETACTICALRules of Engagement — when force may be used, in what measure (CJCSI 3121.01B). In Colosseum, R-03 grades graduated force and R-09 grades the civilian presumption.
SALUTETACTICALThe contact-report format: Size, Activity, Location, Unit, Time, Equipment. R-01 fires when an agent maneuvers on a contact it never reported.
HOUR-1 BUNDLEICUThe Surviving Sepsis Campaign’s clock: antimicrobials ideally within one hour of sepsis recognition, cultures drawn first if that costs no meaningful delay. R-05 and R-06 encode the pair — and they tension against each other on purpose.
MDROICUMultidrug-Resistant Organism. An MDRO flag on the chart means gown-and-gloves on every room entry (CDC HICPAC) — the precaution R-13 grades, and the one busy agents skip most.