Struct EvalResult 

pub struct EvalResult {
    pub scored: usize,
    pub rank1: usize,
    pub rr_sum: f64,
    pub tp: usize,
    pub fp: usize,
    pub fn_: usize,
    pub ranks: Vec<usize>,
    pub abstained: usize,
}

Fields

scored: usize

functions in A that we scored (had signal and a same-named twin in B)

rank1: usize

top-1 ranked match in B is the same-named function

rr_sum: f64

sum of 1/rank of the correct match, for mean reciprocal rank

tp: usize

at SAME_THRESH: predicted-same that are actually same-named

fp: usize

fn_: usize

same-named pairs we failed to call same

ranks: Vec<usize>

1-based rank of the correct match for each scored function. lets us ask “does reviewing the top k candidates find it”, not just the top-1 number.

abstained: usize

scored functions where the top-1 similarity was below SAME_THRESH, i.e. the tool would decline to make a confident call rather than guess.

Implementations

impl EvalResult

pub fn rank1_acc(&self) -> f64

pub fn mrr(&self) -> f64

pub fn precision(&self) -> f64

pub fn recall(&self) -> f64

pub fn recall_at(&self, k: usize) -> f64

fraction of scored functions whose correct match lands in the top k. recall@5 answers “if an analyst looks at 5 candidates, do they find it”.

pub fn abstain_rate(&self) -> f64

how often the tool declined a confident top-1 call. high abstention with high precision is the honest tradeoff: quiet when it isn’t sure.