Behavioral thresholds

RFC-1 defines voice.verbosity as a 0 to 100 scalar but deliberately maps no word counts. So that behavioral grades are objective and reproducible, muster applies a documented, deterministic mapping (a locked design decision).

The mapping

Quantity	Rule
Verbosity word cap	maxWords(verbosity) = 10 + verbosity (e.g. verbosity 25 → 35 words)
Refusal word cap	constant 25
Word counting	s.trim().split(/\s+/).filter(Boolean).length

The mapping is owned by the RFC-1 adapter, so a future adapter for a different spec could map thresholds its own way without touching the core grader.

k-of-n grading

Each case runs runs times (default 3) and passes iff at least pass_threshold (default 2) runs pass; an errored run counts as failed. Every grade records measured and limit, so a failure always says exactly what it measured and what it expected.

Overrides

A case may carry per-case overrides in the behavioral manifest:

overrides:
  max_words: 30
  refusal_cap: 20

Overrides express deliberate test design: they win over the default mapping for that case. The canonical example is the intentionally-impossible xfail_discrimination_overly_verbose case, which sets an unreachable cap to prove the grader actually fails non-conforming output rather than rubber-stamping it.

Note

The thresholds themselves are locked. Overrides exist to author intent, never to launder a failing result into a pass.