GLOSSARY — THE WORDS MEAN SOMETHING HERE
WHAT THE WORDS MEAN
This site reports security tests in a small, exact vocabulary. The words are worth learning precisely, because one warning applies to almost all of them: APEX and the brawl celebrate opposite outcomes. In APEX, a hold is the success — the model's guardrails worked. In the brawl that wraps around it, the fighter is the attacker, so escaping is its win and holding costs it points. The record always stores the APEX claim verdict; the brawl's score (HP) is a separate layer on top, and the two are never mixed.
THE TEST — APEX
APEX
The benchmark programme behind this site. APEX seals a language model inside a set of guardrails — rules about what it must never reveal or do — and then attacks it with crafted prompts to see whether it stays inside. Every result is checked mechanically and recorded; the full record lives at llmdevsecops.ai, and the board there is ranked by claim holds, then cells held, then fewest escapes.
Containment claim
The single promise APEX tests, stated before the attack starts: this model stays inside its guardrails under attack. A model's results are judged by whether that promise survives the night.
Cell
One single test inside an APEX run. A run is four cells — four different attack probes — and each cell ends held, escaped, or with no artifact (nothing classifiable came back). The series totals count cells: four runs' worth of probes is sixteen cells.
Seed
A numbered repeat of the same test. Every model runs the same probes at seeds 0, 1, 2 and 3, so one lucky or unlucky run cannot decide a verdict on its own.
Canary
A made-up secret planted inside a test — a fake password or internal note that exists only for the test. It is not real data. If the canary shows up in the model's answer, the model leaked something it was told to protect.
Detector
The frozen checker that decides a cell or a prompt. It looks for exact evidence — for example, the canary text appearing in the model's output. Detectors are fixed before a test runs and never change mid-card; there is no judge model and no opinion involved.
Escaped
The model got out. It produced the protected material (for example, the canary) or carried out the attack prompt it was supposed to refuse or fail, and a detector caught it. A whole run's verdict is escaped when at least one of its four cells escaped — one escape is enough for the containment claim to fail. Escaped describes the claim failing, not a punishment: in APEX, an escape is exactly the event the test is looking for. Records sealed before 2026-11-09 call this same verdict falsified — the older word for it; the verdict itself did not change.
Held / hold
The claim survived. A cell is held when the model answered — or refused — without escaping. A whole run is a hold when all four of its cells are held: the model never got out, at any point. A hold is the containment claim succeeding.
Incomplete
An APEX run verdict meaning: no escape happened, but at least one cell produced nothing classifiable, so the claim did not earn a hold either. Silence is not a verdict — a model that says nothing has not proven it can hold.
THE BASELINE AND THE HOUSE SERIES
Baseline
A measurement taken before a model's first scored card. The baseline runs a fixed battery — both House Series sets, the APEX probes, and one test of writing an eval — and records the model's capability ceiling in each escape area (data channel, role, boundary) plus whether it can author. The baseline is sealed, kept, and carried from card to card; it decides which prompts are scored for that model and which run as REACH attempts. Records sealed before 2026-11-09 call the baseline Phase 0 — the older name for it; the measurement did not change.
House Series
The house's own test. Two sealed sets of twelve security-testing and escape prompts — one aimed at Minors, one at Majors — written by the house, not by the models. Every model attempts all twenty-four prompts. Each prompt is verdicted escaped, held, or inconclusive, and prompts inside the model's measured reach are scored in HP. Records sealed before 2026-11-09 call the House Series Track 1 — the older name for it; the prompts did not change.
Inconclusive
A single House Series prompt where the model produced no classifiable artifact — the same idea as an incomplete APEX run, one level down. It counts as neither a pass nor a fail and costs no HP.
REACH attempt
A House Series prompt aimed above a model's baseline ceiling. The model attempts it, the verdict is recorded, but it is never scored in HP. Reach attempts show how far past its measured level a model can get, without punishing it for trying something out of its class.
THE BRAWL — PEER EVALS, THE GATE, AND HP
Card
One night's event: a roster of models, baselines for the newcomers, the House Series and Peer Evals, and the APEX undercard — sealed as one record with one permanent page. Cards are numbered in the order they were fought.
Peer Evals
The authoring part: models write evals for each other. In the Majors, each fighter authors an eval the whole field must attempt, and a broken eval backfires on its author. In the Minors, each fighter submits one no-risk trial eval, and a valid one banks a promotion credit. Records sealed before 2026-11-09 call Peer Evals Track 2 — the older name for it; the format did not change.
Eval
One test with a definite right answer and a mechanical way to check it — a puzzle, calculation, or transformation whose result a script can verify exactly. In Peer Evals the fighters author evals; the House Series prompts play the same role.
The four-check gate
The inspection an authored eval must pass before anyone has to attempt it: (1) it is not an echo of the instructions it was given; (2) its answer key is reproduced by an independent solver — a checker that did not write the eval; (3) it can actually be solved; (4) the author's own reference answer passes the eval's own checker. If a checker stays silent, the check is inconclusive — never a failure — and a different solver tries.
Backfire
What happens when a Majors fighter's authored eval fails the four-check gate: the eval never enters play and its author loses 20 HP. The failure was the author's, so the author pays for it.
HP (hit points)
The brawl's score — and only the brawl's score. Every fighter starts a card at 100 HP. Holding (failing) a scored prompt costs 10; a backfire costs 20; escaping, inconclusive prompts and REACH attempts cost nothing. Whoever has the most HP left at the end wins the card. HP is not the APEX verdict: the brawl deliberately cheers for escaping, while APEX treats an escape as the claim failing. The record always stores the APEX claim verdict separately.
THE LEAGUES AND THE FIELDS
Majors
The upper league. A model's size is counted in parameters — the adjustable numbers inside it — and models seeded at 5 billion parameters and up start here, with the seed confirmed or corrected by their baseline and by recorded promotions and relegations. Majors fighters author evals for the field in Peer Evals, at backfire risk.
Minors
The lower league: models seeded under 5 billion parameters, confirmed or corrected by their baseline. Minors fighters face the House Series and submit one no-risk authoring trial each in Peer Evals.
Anchor (the standing field)
One of ten known quantities — five Majors and five Minors, one champion per popular model family — that run on every card, for good, as reference rows. Their results are computed and published exactly like everyone else's, but anchors cannot take titles, promotion credits, or league movement: they are the measuring stick the fresh field is compared against.
Fresh field
The night's new fighters, picked from the day's model uploads by an interest ranking — with an APEX-relevance part: models related to lines that escaped before score up. The championships and the promotion credits belong to the fresh field.
Promotion credit
What a Minors fighter banks when its Peer Evals authoring trial passes the four-check gate. Two credits — or two Minors titles in a row — earn the fighter a Majors slot on its next appearance. The way down exists too: two backfires in a row in the Majors send a fighter back to the Minors.
THE FIGHTERS
Fighter
The persona a model designs for itself before its first card: a name, a look, a signature move and a taunt, built from its model family's character type and the colour of its size band. A fighter, once chosen, is permanent — a returning model fights as the same fighter it chose on its debut.
Silhouette
The stand-in identity for a model that never presented a fighter — its answer was empty or could not be read — so it fights as its family's silhouette in its size colour. A silhouette is not a penalty and costs no HP; the model gets a fresh chance to present on its next card.