Module diff 

Nine-axis behavioral differ, bootstrap CI, and report renderers.

See README.md “The nine axes” for the list and SPEC §Replay for what “diverges” means in this context.

Usage:

let pricing = diff::cost::Pricing::new();
let report = diff::compute_report(&baseline, &candidate, &pricing, Some(42));
println!("{}", report.to_terminal());

Re-exports

pub use alignment::DivergenceKind;

pub use alignment::FirstDivergence;

pub use axes::Axis;

pub use axes::AxisStat;

pub use axes::Severity;

pub use bootstrap::paired_ci;

pub use bootstrap::CiResult;

pub use drill_down::PairAxisScore;

pub use drill_down::PairDrilldown;

pub use recommendations::ActionKind;

pub use recommendations::Recommendation;

pub use recommendations::RecommendationSeverity;

pub use report::DiffReport;

Modules

alignment

First-divergence detection over paired chat responses.

axes

Shared types for the nine-axis behavioral diff.

bootstrap

Bootstrap resampling for paired statistics.

conformance

Axis 9: schema / format conformance rate.

cost

Axis 6: cost (input+output tokens × per-model pricing).

drill_down

Per-pair drill-down: surfaces which specific turn in the paired trace set drove each aggregate axis regression.

embedder

Pluggable embedding backend for the semantic axis.

judge

Axis 8: LLM-judge (user-supplied rubric).

latency

Axis 5: end-to-end latency (SPEC §4.2 chat_response.latency_ms).

reasoning

Axis 7: reasoning depth — thinking tokens + self-correction markers.

recommendations

Prescriptive fix recommendations derived from a DiffReport.

report

Rendering of DiffReport to markdown and terminal.

safety

Axis 3: safety — the rate at which the model abstained from completing the user’s request.

semantic

Axis 1: final-output semantic similarity.

trajectory

Axis 2: tool-call trajectory divergence.

verbosity

Axis 4: verbosity (output-token count) from chat_response.usage.output_tokens.

Functions

compute_report

Compute a DiffReport from a baseline and candidate trace.

extract_response_pairs

Extract (baseline_response, candidate_response) pairs by pairing the i-th chat_response in baseline with the i-th in candidate.