Audit probability policy subgroup equity
Audit an aggregate probability-driven policy across governed groups using weighted selection, error-rate, predictive-value, Brier, and calibration disparities; within-group bootstrap uncertainty; practical tolerances; privacy/support abstention; and Benjamini-Hochberg false-discovery control.
What it's for
Makes trusted AI and management decisions auditable by showing where an aggregate policy is differently calibrated or differently wrong across governed groups—while enforcing support, privacy, uncertainty, and multiple-testing gates.
What you give it
Every tool is a pure function: it computes from what you pass it and reads nothing on its own. Inputs split into evidence you supply from your connected systems, calibration your team owns, and numerical controls that affect precision but never the result's meaning.
| Field | Type | Role | Required |
|---|---|---|---|
| bootstrap_draws | integer ≥ 200, ≤ 20000 | Numerical control | Optional |
| confidence_level | number ≥ 0.8, ≤ 0.99 | Your calibration | Optional |
| false_discovery_rate | number ≥ 0.001, ≤ 0.25 | Your calibration | Optional |
| max_detail_rows | integer ≥ 1, ≤ 500 | Numerical control | Optional |
| maximum_allowed_disparities | object | Evidence | Yes |
| minimum_conditional_rows | integer ≥ 5, ≤ 5000 | Your calibration | Optional |
| minimum_group_rows | integer ≥ 20, ≤ 10000 | Your calibration | Optional |
| observations | array of objects (6 fields) ≥ 60 items | Evidence | Yes |
| reference_group_id | string non-empty | Your calibration | Optional |
| seed | integer | Numerical control | Optional |
Each maximum_allowed_disparities
record
| Field | Type | Required |
|---|---|---|
| brier_score | number (≥ 0, ≤ 1) | Yes |
| calibration_gap | number (≥ 0, ≤ 1) | Yes |
| false_positive_rate | number (≥ 0, ≤ 1) | Yes |
| positive_predictive_value | number (≥ 0, ≤ 1) | Yes |
| selection_rate | number (≥ 0, ≤ 1) | Yes |
| true_positive_rate | number (≥ 0, ≤ 1) | Yes |
{
"bootstrap_draws": 200,
"maximum_allowed_disparities": {
"brier_score": 0.05,
"calibration_gap": 0.05,
"false_positive_rate": 0.1,
"positive_predictive_value": 0.1,
"selection_rate": 0.1,
"true_positive_rate": 0.1
},
"observations": [
{
"decision": 0,
"group_id": "reference",
"id": "equity-reference-0",
"outcome": 0,
"predicted_probability": 0.25
},
{
"decision": 1,
"group_id": "reference",
"id": "equity-reference-1",
"outcome": 1,
"predicted_probability": 0.75
},
{
"decision": 0,
"group_id": "reference",
"id": "equity-reference-2",
"outcome": 0,
"predicted_probability": 0.25
},
{
"decision": 1,
"group_id": "reference",
"id": "equity-reference-3",
"outcome": 1,
"predicted_probability": 0.75
},
{
"decision": 0,
"group_id": "reference",
"id": "equity-reference-4",
"outcome": 0,Truncated for display — the full payload is 434 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.
{
"assumptions": [
"Groups, reference, metric definitions, decision threshold, disparity tolerances, and privacy minimums are governed before outcomes are inspected.",
"Rows are independent at the declared aggregate decision unit or have already been clustered to that unit; bootstrap resampling stays within groups.",
"Observed outcome differences may reflect need, access, history, measurement, or policy and do not by themselves identify discrimination or mechanism."
],
"comparisons": [
{
"absolute_disparity": 0,
"confidence_interval": [
-0.2,
0.2
],
"disparity": 0,
"group_id": "comparison",
"group_value": 0.5,
"material_supported_violation": false,
"maximum_allowed_disparity": 0.1,
"metric": "selection_rate",
"p_value": 1,
"q_value": 1,
"reference_value": 0.5
},
{
"absolute_disparity": 0,
"confidence_interval": [
0,
0
],
"disparity": 0,
"group_id": "comparison",
"group_value": 1,
"material_supported_violation": false,
"maximum_allowed_disparity": 0.1,
"metric": "true_positive_rate",
"p_value": 1,
"q_value": 1,
"reference_value": 1
},
{
"absolute_disparity": 0,
"confidence_interval": [
0,
0Truncated for display — the full payload is 157 lines.
How it works
Constrained optimization — Pick the best feasible option under real limits — budget, headcount, dependencies, capacity — rather than ranking a list and hoping it fits.
- 1 Freeze the decision unit, eligible population, group vocabulary, reference policy, outcome horizon, six metric definitions, practical disparity tolerances, and privacy minimums before inspecting group results.
- 2 Calculate weighted selection rate, true-positive rate, false-positive rate, positive predictive value, Brier score, and signed calibration gap for every privacy-eligible group and the governed reference.
- 3 Resample observations within each group, recompute every disparity against the same explicit or overall reference, derive percentile intervals, and form null-centered bootstrap significance values.
- 4 Apply Benjamini-Hochberg correction across every group-metric comparison and flag only disparities that exceed the practical tolerance, exclude zero at the governed confidence level, and survive the false-discovery threshold.
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
- Objectives use commensurable locally governed value units, constraints reflect real feasibility, and uncertainty covers plausible adverse inputs.
- The policy threshold, outcome definition and maturity horizon are identical across groups, weights represent the intended eligible population, and rows are independent at the declared resampling unit.
- Group labels are lawfully governed, access-controlled, privacy-eligible, and suitable for an aggregate equity audit; missing group membership has an explicit policy rather than silent deletion.
- The recommendation is optimal only for its stated objective, feasible set, evidence, and solver guarantee; it is not a universal management optimum.
- A supported disparity is an investigation and policy-design trigger, not proof of discrimination, a causal explanation, a legal conclusion, or evidence about the merit or intent of any person.
- Do not optimize a policy to parity on these six summaries alone; evaluate utility, access, measurement quality, base rates, uncertainty, and lawful context with accountable review.
Minimum evidence
- observations: at least 60 rows/items
- maximum_allowed_disparities: required and organization-defined
How to validate it
Use chronological train/calibration/test windows, compare proper scores and decision value with a simple baseline, and recalibrate only from outcomes resolved after prediction time.
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
- one row per eligible aggregate decision unit with time-consistent group membership
- explicit missing-group cohort and clustered resampling unit when rows are dependent
- lawful group vocabulary and access policy
- reference group or overall-reference policy
- six practical disparity tolerances, outcome horizon, privacy minima, confidence, and false-discovery rate
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 20 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": "audit an aggregate probabilitydriven policy across" }
→ finds "audit_probability_policy_subgroup_equity"
gitrevio_capability_describe
{ "capability_id": "audit_probability_policy_subgroup_equity" }
→ returns the input schema and agent guidance shown on this page
gitrevio_capability_run
{ "capability_id": "audit_probability_policy_subgroup_equity", "arguments": { ... } }
→ returns the result shown aboveWorks in Claude Desktop, Claude Code, Cursor, Cline, Continue.dev, Goose and Aider. See the MCP server.