Struct ValidationMetrics 

pub struct ValidationMetrics {
    pub validations_passed: AtomicU64,
    pub validations_failed: AtomicU64,
    pub relevancy_stats: HashMap<String, F64Stats>,
    pub running_aggregates: HashMap<String, Mutex<RunningAgg>>,
}

Metrics for result validation, shared across fibers.

Fields

validations_passed: AtomicU64

validations_failed: AtomicU64

relevancy_stats: HashMap<String, F64Stats>

Per-function lossless f64 statistics accumulators for relevancy scores. Exact precision — no quantization, no bucket rounding.

running_aggregates: HashMap<String, Mutex<RunningAgg>>

Live aggregates (moving-window + all-time) per relevancy metric. These read non-destructively, so progress views can sample them every frame without disturbing the exact-precision stats above.

Implementations

impl ValidationMetrics

pub fn new( labels: &Labels, functions: &[RelevancyFn], k: usize, r: Option<usize>, ) -> Self

Create metrics for the given relevancy functions.

The metric family name is the bare function name (recall, precision, f1, …) — k and r are carried as labels rather than baked into the family identifier. This collapses every per-k variant into a single family that consumers can query with recall{k="10",r="10"} instead of the awkward recall_at_10 synthesised name. r defaults to k when the relevancy config doesn’t declare it (r: was unset → first-k semantics).

pub fn assertions_only() -> Self

Create metrics with no relevancy functions (assertions only).

pub fn record_relevancy(&self, metric_name: &str, score: f64)

Record a relevancy score for a named function. Updates both the lossless stats accumulator and the live running aggregate.

pub fn live_snapshot(&self) -> Vec<RelevancyLive>

Snapshot all live relevancy aggregates without disturbing the accumulators. Safe to call every frame from the progress thread.

pub fn passed(&self) -> u64

Get pass count.

pub fn failed(&self) -> u64

Get fail count.