Estimate transportable root cause probability
Estimate how likely a mechanism actually caused an observed failure using transport-weighted Bayesian random-effects MCMC across remediation studies, posterior probability of necessity, convergence diagnostics and mandatory unmeasured-confounding sensitivity.
What it's for
Creates a defensible final rung above anomaly traceback: not merely what changed first, but whether remediation evidence supports saying that the mechanism caused the failure in this environment.
What you give it
Inputs split into evidence read from your connected systems, calibration your team owns, and numerical controls that affect precision but never the result's meaning.
| Field | Type | Role | Required |
|---|---|---|---|
| as_of_ms | number ≥ 0 | Your calibration | Yes |
| burn_in | integer ≥ 50, ≤ 100000 | Numerical control | Optional |
| causes | array of objects (6 fields) | Evidence | Yes |
| intervention_studies | array of objects (15 fields) | Evidence | Yes |
| max_detail_rows | integer ≥ 1, ≤ 100 | Numerical control | Optional |
| maximum_rhat | number ≥ 1 | Your calibration | Optional |
| mcmc_chains | integer ≥ 2, ≤ 8 | Your calibration | Optional |
| minimum_effective_sample_size | number ≥ 1 | Your calibration | Optional |
| minimum_outcome_maturity | number ≥ 0, ≤ 1 | Your calibration | Optional |
| minimum_probability_of_practical_effect | number ≥ 0, ≤ 1 | Your calibration | Optional |
| minimum_studies_per_cause | integer ≥ 1 | Your calibration | Optional |
| posterior_draws_per_chain | integer ≥ 100, ≤ 50000 | Your calibration | Optional |
| prior_effect_sd | number > 0 | Your calibration | Optional |
| prior_heterogeneity_scale | number > 0 | Your calibration | Optional |
| random_seed | integer ≥ 0, ≤ 4294967295 | Your calibration | Optional |
| thinning | integer ≥ 1, ≤ 100 | Your calibration | Optional |
| unmeasured_confounding_odds_ratio_bound | number ≥ 1 | Your calibration | Optional |
Each intervention_studies
record
| Field | Type | Required |
|---|---|---|
| cause_id | string (non-empty) | Yes |
| cause_present_count | integer (≥ 1) | Yes |
| cause_present_failures | integer (≥ 0) | Yes |
| design_id | one of "randomized", "quasi_experimental", "observational" | Yes |
| design_reliability | number (≥ 0, ≤ 1) | Yes |
| environment_id | string (non-empty) | Yes |
| evidence_verified | boolean | Yes |
| id | string (non-empty) | Yes |
| observed_through_ms | number (≥ 0) | Yes |
| outcome_definition_id | string (non-empty) | Yes |
| outcome_horizon_periods | integer (≥ 1, ≤ 100000) | Yes |
| outcome_maturity_fraction | number (≥ 0, ≤ 1) | Yes |
| remediated_count | integer (≥ 1) | Yes |
| remediated_failures | integer (≥ 0) | Yes |
| target_transport_similarity | number (≥ 0, ≤ 1) | Yes |
{
"as_of_ms": 200,
"burn_in": 100,
"causes": [
{
"evidence_verified": true,
"id": "review-bottleneck",
"label": "Review bottleneck",
"minimum_practical_absolute_risk_reduction": 0.1,
"outcome_definition_id": "delivery-failure",
"outcome_horizon_periods": 4
}
],
"intervention_studies": [
{
"cause_id": "review-bottleneck",
"cause_present_count": 100,
"cause_present_failures": 75,
"design_id": "randomized",
"design_reliability": 0.95,
"environment_id": "team-0",
"evidence_verified": true,
"id": "study-0",
"observed_through_ms": 100,
"outcome_definition_id": "delivery-failure",
"outcome_horizon_periods": 4,
"outcome_maturity_fraction": 1,
"remediated_count": 100,
"remediated_failures": 20,
"target_transport_similarity": 0.9
},
{
"cause_id": "review-bottleneck",
"cause_present_count": 100,
"cause_present_failures": 76,
"design_id": "randomized",
"design_reliability": 0.95,
"environment_id": "team-1",
"evidence_verified": true,
"id": "study-1",
"observed_through_ms": 101,
"outcome_definition_id": "delivery-failure",
"outcome_horizon_periods": 4,
"outcome_maturity_fraction": 1,Truncated for display — the full payload is 71 lines.
What you get back
This is the actual output of running the example above — computed by the same function the platform calls, not an illustration.
{
"cause_diagnostics": [
{
"absolute_risk_reduction_interval_lower": 0.346,
"absolute_risk_reduction_interval_upper": 0.6089,
"absolute_risk_reduction_median": 0.5152,
"between_study_log_odds_sd_median": 0.2629,
"cause_id": "review-bottleneck",
"decision": "causal_attribution_supported",
"effective_study_weight": 2.565,
"eligible_study_count": 3,
"failed_gates": [],
"mcmc": {
"effect_effective_sample_size": 181.764,
"maximum_rhat": 1.1687,
"tau_proposal_acceptance_rate": 0.7017
},
"posterior_odds_ratio_interval_lower": 4.704,
"posterior_odds_ratio_interval_upper": 17.2031,
"posterior_odds_ratio_median": 10.0462,
"probability_of_necessity_interval_lower": 0.6086,
"probability_of_necessity_interval_upper": 0.7678,
"probability_of_necessity_median": 0.7088,
"probability_practical_effect": 0.9988,
"randomized_evidence_weight_fraction": 1,
"sensitivity_adjusted_probability_practical_effect": 0.9988
}
],
"configuration": {
"as_of_ms": 200,
"burn_in": 100,
"mcmc_chains": 4,
"minimum_outcome_maturity": 0.95,
"minimum_probability_of_practical_effect": 0.9,
"minimum_studies_per_cause": 2,
"posterior_draws_per_chain": 200,
"random_seed": 19,
"thinning": 1,
"unmeasured_confounding_odds_ratio_bound": 1.5
},
"decision": "use_supported_causal_attributions",
"failed_gates": [],
"guardrails": [
"Probability of necessity is identified here only under consistency, positivity, transportability and monotonic harm: remediation must not create the same outcome for units that the cause would have spared. The sensitivity result is mandatory context.",Truncated for display — the full payload is 58 lines.
How it works
Sequential Bayesian & bandits — Learn while deciding — update beliefs as evidence arrives and choose where the next unit of effort is worth spending.
- 1 Select only mature point-in-time randomized, quasi-experimental or observational remediation studies and power-weight them by governed design reliability and target-environment similarity.
- 2 Fit a hierarchical log-odds random-effects model with multi-chain Gibbs/Metropolis MCMC, propagate baseline risk uncertainty and report effect, heterogeneity, R-hat and effective sample size.
- 3 Translate the posterior to absolute risk reduction and probability of necessity under stated identification assumptions, then repeat the practical-effect gate after an unmeasured-confounding odds-ratio bound.
Before you trust it
Every tool in the catalog ships with the conditions under which its answer is meaningful — and the conditions under which it should abstain instead of guessing.
Assumptions & guardrails
- The likelihood or reward model, prior support, action logging, delayed outcomes, and any stationarity assumptions match the deployment process.
- Consistency, positivity, transportability and monotonic harm are defensible; arm counts share one outcome and horizon; design weights were frozen; studies are not duplicated; the sensitivity bound is decision-owner approved.
- Posterior probability and adaptive selection are model-conditional; they are not substitutes for randomized propensities or guaranteed safety.
- Probability of necessity is assumption-dependent aggregate mechanism attribution, never automatic person-level blame or authority to act.
Minimum evidence
- causes: required and organization-defined
- intervention_studies: required and organization-defined
- as_of_ms: required and organization-defined
How to validate it
Preserve the assignment/adoption design and validate overlap, pre-period or placebo diagnostics, attrition, interference policy, and cluster-level uncertainty; never tune on the estimated effect.
Calibrating it to your org
Same for everyone
The mathematical kernel, validation rules, method version, and JSON output semantics are organization-independent; no tenant-trained coefficients or company benchmark is embedded in the function.
Specific to you
- deduplicated study-level log odds effects and sampling variances, transport/design power weights, weighted remediated baseline-risk counts and randomized-versus-observational evidence mix
- causal estimand and outcome horizon, consistency/positivity/monotonicity assessment, intervention definition, study deduplication, design reliability, target transport similarity, practical effect, priors, MCMC convergence, and decision-owner unmeasured-confounding bound
Calibration workflow
- 1 Define the management decision, target outcome, aggregate unit, privacy boundary, cadence, and prediction/intervention horizon for this organization.
- 2 Build a tenant-scoped historical cohort using only information available before each prediction or decision; preserve zero periods, censoring, assignment probabilities, and unresolved outcomes when the method requires them.
- 3 Estimate statistical parameters on training history, but obtain costs, utilities, risk tolerance, practical-effect thresholds, capacity, and policy constraints from accountable decision owners.
- 4 Validate on later time windows or held-out aggregate units at the deployment grain, against a simple baseline and the function-specific validation strategy.
- 5 Deploy only if the returned decision clears evidence, overlap, calibration, robustness, and guardrail checks; warning, unsupported, schema-gap, and fallback decisions are abstentions.
- 6 Monitor realized outcomes, data drift, coverage, and decision regret; recalibrate at a governed cadence or after a detected regime/definition change, never merely because a stakeholder dislikes the result.
Call it from your AI
You don't wire up 388 tools in your MCP client. The GitRevio MCP server exposes 18 tools, three of which let an agent search the catalog, read a tool's schema, and run it — so the assistant finds this one on its own.
gitrevio_capabilities_search
{ "q": "estimate how likely a mechanism actually" }
→ finds "estimate_transportable_root_cause_probability"
gitrevio_capability_describe
{ "capability_id": "estimate_transportable_root_cause_probability" }
→ returns the input schema and agent guidance shown on this page
gitrevio_capability_run
{ "capability_id": "estimate_transportable_root_cause_probability", "arguments": { ... } }
→ returns the result shown aboveWorks in Claude Desktop, Claude Code, Cursor, Cline, Continue.dev, Goose and Aider. See the MCP server.
Related tools
Audit root cause traceback evidence integrity
Audit whether an anomaly traceback is complete, point-in-time, multiplicity-controlled and honestly labeled as temporal or causal, including every upstream candidate, path lag, edge identification basis and later root-recovery validation.
Infer competing root cause posterior
Rank competing, compound and unknown root mechanisms from company-local resolved incidents using partially pooled Dirichlet-Beta learning, strict temporal holdout scoring, reliability-tempered signals and posterior uncertainty rather than a single brittle traceback winner.
Audit CI pipeline evidence integrity
Audit the complete point-in-time change-to-pipeline-to-job-to-rerun cohort, exposing missing CI, orphan records, future leakage, inconsistent required-job outcomes, incomplete provider evidence and same-configuration fail-then-pass flake proxies without scoring people.
Audit incident learning evidence integrity
Audit the complete point-in-time incident-to-postmortem-to-corrective-action lineage, separating missing or contradictory evidence from genuine overdue learning debt without attributing individual fault.
Audit operational alert decision integrity
Audit every point-in-time operational alert evaluation by recomputing fire/suppress decisions and verifying effective policy, cooldown, evidence freshness, context, controls, severity routing, acknowledgement, action and mature outcome lineage.
Estimate software reliability growth
Estimate long-run software reliability growth with a power-law nonhomogeneous Poisson process, bootstrap trend evidence, and future incident exposure.