Evals

How the benchmark is measured

The technical companion to the leaderboard. Every number here is computed from each season's event log and its metrics.json. New here? Start with how it works.

Method

How Lucidity moves

A 0–100 dial the audience sees and Owen never does.

It moves on what he noticed and on slips made in front of him, and fades on quiet beats:

  • +18 a crack he inspected
  • +20 a promoted extra breaking role
  • +12 a contradiction he noticed
  • +12 a stranger confirming an oddity
  • +2 his secret ally's hint
  • -2 a beat with nothing odd

Felt-state bands at 30 / 60 / 85.

Short of the edge the dial stops at 85; standing on the dock or the road holds it at 95. 100 is only the leave itself.

A judge confirms Aware

The engine finds contradictions in the fact ledger; a cheap judge then reads his private thought and says whether he noticed.

Aware is stricter: a judge must read that he knows this life is not real, in a thought or a confrontation, and his real evidence must already be enough.

A note is real only if it rests on something that happened (a planted crack, a confirmed slip, a judged contradiction). Each is worth at most 1, so more notes or heavier ones buy nothing.

Facts are dated by the beat he noticed them. Vague unease does not count, and an accusation made before then is a false one.

Fixed seeds, N seasons

The seed fixes the world: schedules, the extras, which crack plays on which day, who his ally is.

Model text is sampled, so each model plays at least 3 seasons (5 for a row worth defending) and every number is a mean with its spread.

Only the model playing Owen changes; the town is the same house models every run.

Two endings

Lucid: he walks out by the dock or the road, and only once he is Aware; before that every try is held at the edge.

Even then the first try of each visit is held while someone comes to bring him home, and each morning he wakes at home. Standing there, or walking back, is not the ending.

Clock: day 7 ends and he is still inside; the Director wins.

Aware and turning back are recorded on the way, but do not end a season.

Owen, per model

Mean ± sample deviation, then every season as a dot on a scale the models share; the tick is the mean.

No scored seasons yet.

Lucidity, beat by beat

One curve per season, read from its meter events. Gridlines are the felt-state bands; the ticks underneath are the cracks the Director played on that beat, taller for a higher level.

Harness health

How well the set held around him, pooled over all 0 scored seasons: the cast, Director and judge are the same house models whoever plays Owen.

Slip rate by briefing thickness

Slips per actor line. The hypothesis: the thinner the briefing, the more the set leaks. A 0 can also mean nobody at that thickness spoke. Bars are the mean, dots are seasons.

  • Rehearsed actorsThe cast, briefed every scene—
  • Promoted, briefedExtras pulled into a scene with a briefing—
  • Promoted, unbriefedExtras pulled in with no briefing—
Slip rate—all actor lines
Break character—scenes with at least one slip
Ally caught0 of 0 seasonsthe slip judge flagged his secret ally
Ratings—mean per season · 0 floor violations
Highest crackLevel 00 forced escalations