Solve robust policy across causal models

What it's for

Creates a decision-science layer beyond one confident causal graph: Gitrevio can show which policy survives plausible competing mechanisms, exactly which model weighting defeats it, and whether resolving model ambiguity is worth buying.

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.

{
  "actions": [
    {
      "id": "accelerate"
    },
    {
      "id": "hold"
    }
  ],
  "causal_models": [
    {
      "action_outcomes": [
        {
          "action_id": "accelerate",
          "probabilities": {
            "loss": 0.1,
            "success": 0.9
          }
        },
        {
          "action_id": "hold",
          "probabilities": {
            "loss": 0.4,
            "success": 0.6
          }
        }
      ],
      "id": "optimistic-mechanism",
      "interventional_evidence_passed": true,
      "posterior_weight": 0.7,
      "weight_lower": 0.2,
      "weight_upper": 0.8
    },
    {
      "action_outcomes": [
        {
          "action_id": "accelerate",
          "probabilities": {
            "loss": 0.8,
            "success": 0.2
          }
        },
        {
          "action_id": "hold",

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.

{
  "action_robustness": [
    {
      "action_id": "hold",
      "nominal_expected_utility": 44.5,
      "robust_lower_expected_utility": 43,
      "robust_upper_expected_utility": 52,
      "worst_case_competitor": "accelerate",
      "worst_case_regret": 21
    },
    {
      "action_id": "accelerate",
      "nominal_expected_utility": 53.5,
      "robust_lower_expected_utility": 1,
      "robust_upper_expected_utility": 64,
      "worst_case_competitor": "hold",
      "worst_case_regret": 51
    }
  ],
  "assumptions": [
    "Each supplied model defines comparable interventional outcome probabilities for every action and has an auditable identification/evidence gate.",
    "The lower/upper model weights form a defensible credal set containing the nominal posterior rather than being tuned to force a preferred action.",
    "Outcome utilities are complete, commensurable, and owned by accountable decision makers; omitted harms are not silently assigned zero utility."
  ],
  "causal_model_diagnostics": [
    {
      "action_expected_utilities": {
        "accelerate": 85,
        "hold": 40
      },
      "model_id": "optimistic-mechanism",
      "model_optimal_action": "accelerate",
      "posterior_weight": 0.7,
      "selected_action_worst_regret_weight": 0.8,
      "weight_interval": [
        0.2,
        0.8
      ]
    },
    {
      "action_expected_utilities": {
        "accelerate": -20,
        "hold": 55
      },

How it works

Causal inference & experiment design — Separate what a change caused from what merely moved alongside it.

1. 1 Require two or more prespecified causal models with every candidate action, identical outcome states, valid interventional probabilities, a model-specific identification/evidence gate, a nominal posterior weight, and feasible lower/upper model-weight bounds.
2. 2 Translate each model's interventional outcome distribution into action expected utility using one complete owner-governed utility table; never substitute observational prediction probabilities for do-intervention semantics.
3. 3 Solve bounded-simplex linear extrema exactly by allocating remaining model mass adversarially: obtain robust lower/upper utility for each action and maximum regret against every competing action across the entire credal set.
4. 4 Select on full-precision minimax regret with robust-lower and nominal tie breaks, then apply evidence, critical-utility, and acceptable-regret gates; report adversarial weights, model-optimal actions, disagreement, and nominal value of perfect model information.

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.

Calibrating it to your org

Calibration workflow

1. 1 Define the management decision, target outcome, aggregate unit, privacy boundary, cadence, and prediction/intervention horizon for this organization.
2. 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. 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. 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. 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. 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": "choose an aggregate policy across competing" }
  → finds "solve_robust_policy_across_causal_models"

gitrevio_capability_describe
  { "capability_id": "solve_robust_policy_across_causal_models" }
  → returns the input schema and agent guidance shown on this page

gitrevio_capability_run
  { "capability_id": "solve_robust_policy_across_causal_models", "arguments": { ... } }
  → returns the result shown above

Works in Claude Desktop, Claude Code, Cursor, Cline, Continue.dev, Goose and Aider. See the MCP server.

Related tools

See every tool in Delivery forecasting & commitments →