ruvector-coherence

Quantitative coherence metrics for comparing attention mechanisms — measure what gating costs and what it preserves.

Overview

When replacing softmax attention with a gated alternative (such as min-cut gating), the central question is: does the output stay coherent? This crate provides a suite of metrics, comparison utilities, quality guardrails, and batched evaluation tools to answer that question quantitatively.

"Coherence" here means the degree to which gated attention outputs preserve the semantic and structural properties of baseline softmax outputs. The crate measures this through vector similarity, contradiction detection, mask overlap analysis, and statistical aggregation with confidence intervals.

Modules

Metrics Explained

contradiction_rate

Measures the fraction of output pairs where the dot product between prediction and reference vectors is negative. A high contradiction rate signals that gating has inverted the semantic direction of outputs.

use ruvector_coherence::contradiction_rate;

let predictions = vec![vec![1.0, 2.0], vec![3.0, 4.0]];
let references  = vec![vec![1.0, 1.0], vec![-1.0, -1.0]];

let rate = contradiction_rate(&predictions, &references);
// rate = 0.5 (second pair contradicts)

entailment_consistency

Computes mean pairwise cosine similarity between consecutive output vectors. High values (close to 1.0) indicate that adjacent outputs remain aligned -- useful for detecting whether gating introduces erratic token-to-token swings.

use ruvector_coherence::entailment_consistency;

let outputs = vec![vec![1.0, 0.0], vec![0.9, 0.1], vec![0.8, 0.2]];
let consistency = entailment_consistency(&outputs);
// consistency close to 1.0 (outputs smoothly evolve)

delta_behavior (DeltaMetric)

Compares baseline and gated attention outputs element-by-element, returning:

use ruvector_coherence::delta_behavior;

let baseline = vec![1.0, 2.0, 3.0];
let gated    = vec![1.1, 1.9, 3.1];

let delta = delta_behavior(&baseline, &gated);
println!("Coherence delta: {:.6}", delta.coherence_delta);
println!("Decision flips:  {}", delta.decision_flips);
println!("Path change:     {:.6}", delta.path_length_change);

Mask Comparison

compare_attention_masks (ComparisonResult)

Provides a full comparison between two boolean attention masks:

use ruvector_coherence::compare_attention_masks;

let baseline = vec![true, true, false, false, true];
let gated    = vec![true, false, false, true, true];

let cmp = compare_attention_masks(&baseline, &gated);
println!("Jaccard:      {:.3}", cmp.jaccard);       // 0.500
println!("Edge flips:   {}", cmp.edge_flips);        // 2
println!("Sparsity ratio: {:.3}", cmp.sparsity_ratio);

Standalone helpers jaccard_similarity and edge_flip_count are also available for use outside of the full comparison struct.

Quality Guardrails

quality_check (QualityResult)

A pass/fail gate that checks whether gated output stays close enough to baseline output. The check passes when cosine similarity meets or exceeds a configurable threshold.

use ruvector_coherence::quality_check;

let baseline_out = vec![1.0, 2.0, 3.0];
let gated_out    = vec![1.1, 2.1, 3.1];

let result = quality_check(&baseline_out, &gated_out, 0.99);
println!("Cosine sim:  {:.4}", result.cosine_sim);
println!("L2 distance: {:.4}", result.l2_dist);
println!("Passes:      {}", result.passes_threshold);

Batch Evaluation

evaluate_batch (BatchResult)

Runs delta_behavior and quality_check across an array of sample pairs, aggregating results with standard statistics.

use ruvector_coherence::evaluate_batch;

let baselines = vec![vec![1.0, 2.0, 3.0]; 100];
let gated     = vec![vec![1.05, 1.95, 3.05]; 100];

let batch = evaluate_batch(&baselines, &gated, 0.99);

println!("Samples:    {}", batch.n_samples);
println!("Mean delta: {:.6}", batch.mean_coherence_delta);
println!("95% CI:     [{:.6}, {:.6}]", batch.ci_95_lower, batch.ci_95_upper);
println!("Pass rate:  {:.1}%", batch.pass_rate * 100.0);

Typical Workflow

1. Run attn_softmax()  --> baseline outputs
2. Run attn_mincut()   --> gated outputs + keep_mask
3. quality_check()     --> per-sample pass/fail
4. compare_attention_masks() --> mask overlap analysis
5. evaluate_batch()    --> aggregate stats with 95% CI
6. Export via ruvector-profiler CSV emitters

use ruvector_attn_mincut::{attn_softmax, attn_mincut};

let (seq_len, d) = (32, 64);
let q = vec![0.1f32; seq_len * d];
let k = vec![0.1f32; seq_len * d];
let v = vec![1.0f32; seq_len * d];

let baseline = attn_softmax(&q, &k, &v, d, seq_len);
let gated = attn_mincut(&q, &k, &v, d, seq_len, 0.5, 2, 0.01);

use ruvector_coherence::*;

let delta = delta_behavior(&baseline.output, &gated.output);
println!("Coherence delta: {:.6}", delta.coherence_delta);
println!("Decision flips:  {}", delta.decision_flips);

let quality = quality_check(&baseline.output, &gated.output, 0.99);
println!("Passes: {} (cosine={:.4})", quality.passes_threshold, quality.cosine_sim);

let baselines = vec![baseline.output.clone(); 100];
let gateds = vec![gated.output.clone(); 100];

let batch = evaluate_batch(&baselines, &gateds, 0.99);
println!("Mean delta: {:.6} +/- {:.6}", batch.mean_coherence_delta, batch.std_coherence_delta);
println!("95% CI: [{:.6}, {:.6}]", batch.ci_95_lower, batch.ci_95_upper);
println!("Pass rate: {:.1}%", batch.pass_rate * 100.0);

Related Crates

License

Licensed under the MIT License.