FEATURES / PRODUCTION STATISTICS

Real statistics, not just dashboards

DORA tells you cycle time dropped by 14% last quarter. A single number rarely holds up in a review. Gitrevio ships a suite of production statistical methods — survival analysis, Bayesian change-point detection, doubly-robust cohort lift with confidence intervals, and axiom-verified attribution — all mounted, dependency-free, and calibrated against your own history.

The methods, named

BOCPD (Adams-MacKay 2007)
Bayesian online change-point detection with calibrated run-length posteriors. Catches regime shifts without thresholds you'd have to hand-tune. Mounted and dependency-free.
Cox proportional hazards + Kaplan-Meier (1958)
Survival analysis on time-to-merge and time-to-resolve. Per-team hazard ratios with proportional-hazards assumption tests. Shipped as a report surface.
Doubly-robust AIPW cohort lift
Cohort-lift modelling with propensity-score matching (Rosenbaum-Rubin 1983) and AIPW, with bootstrap confidence intervals. Robust to misspecification of either the outcome or the propensity model.
Axiom-verified Shapley attribution
When a composite score moves, the engine attributes the move across its factors using a Shapley library whose efficiency, symmetry, dummy-player, and linearity axioms are verified by tests — not an ad-hoc weighting.
Kalman filter (Joseph-form)
Smoothed metric series with MLE-tuned noise covariance. Joseph-form for numerical stability on long horizons.
Lognormal cycle-time forecasting
Lognormal MLE per team yields p50/p75/p90 completion dates, calibrated against your own history, with a risk-factor decomposition.
Bayesian network (in progress)
A joint posterior over engineering outcomes — delivery, quality, attrition — for querying conditional probabilities. Implementation complete; the surface is being wired in.
LinUCB policy search
Contextual bandit over process recommendations. Recommendations are tracked over time so the policy improves with feedback.
Anomaly → root-cause traceback
Anomaly detection chains through the metric dependency graph to surface the upstream signal behind the change.

A worked example

Lead time dropped 18% in March. BOCPD locates the change point, Shapley attribution partitions the drop across its factors, and survival analysis reports the shift in time-to-merge hazard.

# change-point + Shapley attribution on lead_time_p50
metric: lead_time_p50 window: 2026-03-01..2026-03-31
→ baseline: 4.2d observed: 3.4d delta: -18%
→ change-point: 2026-03-09 (BOCPD run-length posterior 0.94)
→ Shapley-attributed factors (sum = -18.1%):
- reviewer_assignment_latency: -9.2%
- pr_size_distribution: -5.1%
- ci_pipeline_p95: -2.4%
- residual: -1.4%
→ time-to-merge hazard ratio: 1.31 (Cox PH, PH-assumption OK)

Decisions, not correlations

Correlation dashboards stall in the executive review. "Cycle time and team size both dropped" is not a finding. A survival hazard ratio, a change-point posterior, or a doubly-robust lift estimate with a confidence interval is.

Every attribution is axiom-verified. The Shapley library's efficiency, symmetry, dummy-player, and linearity properties are checked by tests, so the factor weights hold up when a reviewer challenges them.

Calibrated, not pseudo-Bayesian

Posterior intervals are calibrated against your own history. The Kalman filter's noise covariance is tuned by MLE on your data; BOCPD's hazard prior is re-fit each month.

Forecasts are lognormal MLE fits on your own delivery history — p50/p75/p90 completion dates, recalibrated as new data lands rather than pinned to a fixed distribution.

The tools behind it

Causal claims on this page are produced by named functions in the tool catalog. Each one publishes its assumptions, its identification strategy and the point at which it declines to give an answer.

Causal inference & experiment design Synthetic control impact Honest intervention policy tree Audit probabilistic forecasts All causal evidence tools →

Ready to See Your Engineering work clearly?

Get started free