Struct BenchmarkDescriptor 

pub struct BenchmarkDescriptor {
    pub id: &'static str,
    pub name: &'static str,
    pub summary: &'static str,
    pub detail: &'static str,
    pub duration_hint: &'static str,
    pub expected_secs: u64,
    pub updated: &'static str,
    pub needs_confirmation: bool,
    pub intended_for: Option<ModelExpectation>,
    pub threshold_params: &'static [(&'static str, &'static str)],
    pub sensitivity: Sensitivity,
    pub ctor: fn() -> Box<dyn DynBenchmark>,
}

2026-09-26: Static identity of a benchmark, and how to construct one. The registry (registry::ALL), the TUI list pane and the run directories (ArtifactStore::runs_dir) all read it.

Fields

id: &'static str

2026-09-26: Stable and filename-safe ([a-z0-9-], checked by registry::tests::ids_are_unique_and_filename_safe). Names the run directory <home>/runs/<id>/.

name: &'static str

summary: &'static str

2026-09-26: One line for the suite list.

detail: &'static str

2026-09-26: A paragraph for the detail pane: what it measures and what it costs.

duration_hint: &'static str

2026-09-26: Rough wall time at default parameters, for display, e.g. "~15 min".

expected_secs: u64

2026-09-26: Expected wall time at default parameters, in seconds: the certification planner’s estimate when no completed run of this id has been measured (bench_certify::plan::units). Zero only on descriptors whose duration_hint starts with unrunnable (registry::tests::every_runnable_benchmark_declares_an_expected_duration).

updated: &'static str

2026-09-26: When this benchmark’s measurement last changed (new thresholds, prompt set or scoring rule), as YYYY-MM-DD; not when its code was edited. It tells a reader whether two runs are comparable.

needs_confirmation: bool

2026-09-26: True when starting has a side effect beyond load on the endpoint. The TUI (tui::bench_keys) and bench certify (--yes) require an explicit confirmation for these.

intended_for: Option<ModelExpectation>

2026-09-26: The checkpoint families this benchmark is defined on, if any. The executor’s endpoint check (coherence::probe_for) warns on a mismatch and never refuses the run. None means the benchmark measures whatever it is pointed at.

threshold_params: &'static [(&'static str, &'static str)]

2026-09-26: Parameters whose run-time value comes from the selected model variant’s committed bound, as (param key, metric key) pairs, for benchmarks that compute their own verdict against a knob that is also a BENCH.toml threshold (the agentic gate’s wall_budget_s against sum_wall_s). A gate run (bench_resolve::apply_threshold_params) and the TUI derive the value from the variant’s bound; an explicit --param wins.

sensitivity: Sensitivity

2026-09-26: Whether this benchmark’s number is a speed number, and so corruptible by the state of the box that produced it. A required field of the descriptor, not a list in the hardware policy, so every new benchmark has to answer it. See crate::hardware::policy.

ctor: fn() -> Box<dyn DynBenchmark>

Implementations

impl BenchmarkDescriptor

pub fn build(&self) -> Box<dyn DynBenchmark>