Struct ContextBundle 

pub struct ContextBundle {
    pub stage: String,
    pub budget_tokens: u32,
    pub used_tokens: u32,
    pub capsules: Vec<ContextCapsule>,
    pub excluded: Vec<ContextCapsule>,
    pub skipped: bool,
    pub top_score: f32,
    pub top_abs_evidence: f32,
    pub evidence_coverage: f32,
    pub uncovered_terms: Vec<String>,
    pub chronological: bool,
    pub known_fact_conflicts: Vec<String>,
}

Fields

stage: String

budget_tokens: u32

used_tokens: u32

capsules: Vec<ContextCapsule>

excluded: Vec<ContextCapsule>

skipped: bool

v0.6: true when the top capsule score was below min_score. All capsules are empty; no tokens were injected.

top_score: f32

v0.6: best composite score observed before the skip check. Useful for diagnostics (“why was the brain silent?”).

top_abs_evidence: f32

v2.7: best ABSOLUTE cosine evidence any memory candidate achieved. Unlike top_score this is not normalized, so “0.42” means the same thing on a 3-memory brain and a 3000-memory brain. -1.0 means no comparable cosine was available (lean builds or cross-model rows), so the cosine-calibrated gate has no verdict. Exposed for diagnostics and threshold sweeps.

evidence_coverage: f32

v2.6: what fraction of the query’s discriminating power the returned capsules cover, collectively, in [0, 1].

Kimetsu already abstains at the bundle level — nothing above min_score means an empty bundle and zero tokens. What it never did is say anything about a bundle it does return, so a reader handed three capsules that touch half the question has no way to tell that from three that answer it, and confabulates the rest. That is what BEAM’s abstention track measures, and where Kimetsu scores worst (45% / 30%).

IDF-weighted against the corpus, so a query term present in every memory contributes nothing and a rare one dominates — the same weighting the per-memory lexical floor uses, applied to the bundle as a whole. 1.0 when there is no discriminating term to measure against.

uncovered_terms: Vec<String>

v2.6: the discriminating query terms no returned capsule mentions.

The actionable half of evidence_coverage: a reader can be told precisely what memory does not know about, rather than being handed a number. Empty when coverage is complete or unmeasurable.

chronological: bool

v2.6: true when the query asked about order and the capsules were re-rendered chronologically, oldest first, each carrying its date.

The reader needs to be told, or a time-ordered bundle looks like a relevance-ranked one whose ranking has gone wrong. See crate::ordering for why ordering is rendered rather than retrieved.

known_fact_conflicts: Vec<String>

Conflicts observed in eligible evidence before delivery trimming.

Trait Implementations

impl Clone for ContextBundle

fn clone(&self) -> ContextBundle

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for ContextBundle

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.