Enum MetricProvenance 

pub enum MetricProvenance {
    Euclidean,
    OutputFisher {
        rank: usize,
    },
    OutputFisherDownstream {
        rank: usize,
    },
    WhitenedStructured {
        factor_rank: usize,
    },
}

Where the per-row metric came from — the provenance that makes “likelihood-metric ≠ gauge-metric” diagnosable instead of silent.

Object 4 (the gauge object) reads this to certify which inner product the fit actually used; #974 fills MetricProvenance::WhitenedStructured with a factor-analytic residual-covariance whitening.

Variants

Euclidean

M_n = I_p for every row. The likelihood is isotropic and the gauge pullback reduces to the bare J_nᵀ J_n. This is the default and is bit-for-bit the historical isotropic-φ̂ path.

OutputFisher

M_n = U_n U_nᵀ (+ solver-only δI) from supplied per-row output-Fisher factors U_n ∈ ℝ^{p × rank}. The canonical “one unit of latent motion ↦ one unit of behavioral change” metric: residuals are whitened in the output-Fisher inner product and the gauge pulls back through the same factors. The rank is carried in the provenance so a consumer (Object 4) can certify the factor rank that produced the inner product.

Fields

rank: usize

OutputFisherDownstream

M_n = U_n U_nᵀ from per-row output-Fisher factors that aggregate the downstream influence of position n over future positions through the KV path, rather than the same-position logits of MetricProvenance::OutputFisher (#980, mechanism 2).

The same-position pullback ∂logits_t/∂x_t can be ≈ 0 for a feature whose entire causal effect lands many tokens later (information carried forward through attention); a gauge built on it is blind to exactly that content. This provenance is the forward-looking alternative: each row’s factor U_n is the top-rank factorization of the aggregated output Fisher Σ_{t ≥ n} (∂logits_t/∂x_n)ᵀ F_t (∂logits_t/∂x_n) over future positions the residual stream at n reaches. It is provenance-generic: it whitens nothing ([Self::whitens_likelihood] is false, like MetricProvenance::OutputFisher) and drives the gauge / lens / enrichment unchanged ([Self::is_output_fisher_like]). The lens/gauge machinery consumes it identically; only the scientific reading changes — dormant-feature detection becomes forward-looking (a feature driving far-future tokens now registers behavioral coupling that the same-position metric reported as ≈ 0).

Fields

rank: usize

WhitenedStructured

Structured-residual whitening: M_n = Σ_n^{-1} from the estimated factor-analytic residual covariance Σ_n = Λ c(z_n) Λᵀ + D (#974), with factor_rank the selected factor count. Produced by Structured-residual producers materialize this provenance when they fit a residual-covariance whitening model; the only provenance for which whitens_likelihood is true. It carries the same low-rank factor layout as MetricProvenance::OutputFisher.

Fields

factor_rank: usize

Trait Implementations

impl Clone for MetricProvenance

fn clone(&self) -> MetricProvenance

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Copy for MetricProvenance

impl Debug for MetricProvenance

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

Formats the value using the given formatter.

impl Eq for MetricProvenance

impl PartialEq for MetricProvenance

fn eq(&self, other: &MetricProvenance) -> bool

Tests for self and other values to be equal, and is used by ==.

fn ne(&self, other: &Rhs) -> bool

Tests for !=. The default implementation is almost always sufficient, and should not be overridden without very good reason.

impl StructuralPartialEq for MetricProvenance