Benchmark

Every metric is measured one of three ways: arm comparison (each question is assigned an answer policy at random, so the policy is the only thing that differs between comparable tasks; live: resolve), judge rubric, or self-report. A few numbers are derived from the beliefs you state.

Arm comparison

Scope: this person. The cross-person comparison is the offline runner's job.

arm	n	turns to settled	settled rate	satisfaction	follow	verdict	surprise	resolution	\|pred−stated\|	intent click	clarification	probe
no turns yet

Judge rubric

—next question predicted, judge score (n=0)

—suggested follow-up clicked (rate)

—|P predicted − P stated| (n=0)

—understanding: delayed probe, judge 0–100 (n=0)

Self-report

Legibility, understanding and verdict are a ladder: you could follow it, you understood it, you could judge it. Each only means something on top of the one below.

—legibility: could you follow it? (1–5)

—verdict: could you judge it? (1–5)

—satisfaction (1–5)

—surprise (1–5)

—confidence: mean stated sureness

—'don't know' rate

Derived from your stated beliefs

The martingale metrics follow the paper: R² of a linear fit of update on prior, sign predictability from logistic direction fits, both read with informativeness, because never updating also looks like a perfect martingale. Needs ≥ 8 stated pairs.

—resolution: sureness after − before (n=0)

—re-ask drift (n=0)

—martingale R² of update on prior (batch, n=0; 0 = martingale)

—martingale sign-predictability (0 = coin)

—slope of update on prior

—informativeness: mean |update|

Behaviour

—turns to settled (n=0)

—project: turns to resolved

—follow-up was a clarification (rate)

—answer length, words

—answer reading grade

—"I checked a claim" (rate)

—reading-list vote balance

0search turns

0return visits

Scope: no person selected. Unmeasurable values show as —, never 0.

Unveil, an enlightenment learning engine. MVP. Spec on the hub; code in meno-sh/Unveil.