Infer revealed policy preferences maxent irl

Infer aggregate linear state-feature rewards and their implied stochastic policy from sequential demonstrations using finite-horizon maximum-causal-entropy inverse reinforcement learning.

Causal inference & experiment design Estimate CTOs Tech CEOs IT managers Team leads Investors

What it's for

Makes the priorities implicit in recorded team/project policy choices inspectable without pretending to infer personal intent.

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
actions array of objects (2 fields) ≥ 2 items Evidence Yes
convergence_tolerance number ≥ 1e-10, ≤ 0.1 Your calibration Optional
demonstrations array of objects (3 fields) ≥ 50 items Evidence Yes
discount_factor number ≥ 0.5, ≤ 1 Your calibration Optional
feature_names array of string ≥ 1 item Evidence Yes
iterations integer ≥ 10, ≤ 10000 Numerical control Optional
l2_penalty number ≥ 0, ≤ 100 Your calibration Optional
learning_rate number ≥ 0.00001, ≤ 10 Your calibration Optional
states array of objects (2 fields) ≥ 2 items Evidence Yes
temperature number ≥ 0.01, ≤ 100 Your calibration Optional

Each demonstrations record

Field Type Required
actions array of string (≥ 2 items) Yes
id string (non-empty) Yes
states array of string (≥ 3 items) Yes
Example input 
{
  "actions": [
    {
      "id": "wait",
      "transition_matrix": [
        [
          0.9,
          0.1
        ],
        [
          0.05,
          0.95
        ]
      ]
    },
    {
      "id": "repair",
      "transition_matrix": [
        [
          0.1,
          0.9
        ],
        [
          0.1,
          0.9
        ]
      ]
    }
  ],
  "demonstrations": [
    {
      "actions": [
        "wait",
        "wait",
        "wait"
      ],
      "id": "revealed-policy-0",
      "states": [
        "healthy",
        "healthy",
        "healthy",
        "healthy"
      ]
    },

Truncated for display — the full payload is 751 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.

Example output 
{
  "assumptions": [
    "The supplied aggregate state/action process is Markov-sufficient, transition probabilities are stable, and demonstrations represent the policy population.",
    "Observed choices are approximately maximum-causal-entropy responses to a linear reward over the declared state features; omitted features can reverse inferred weights.",
    "Reward scale and correlated features limit identification, so weights describe revealed model preferences rather than motives, intent, or normative priorities.",
    "Use only privacy-eligible team, project, service, or company demonstrations; never infer an individual's loyalty, character, or employment suitability."
  ],
  "decision": "revealed_preferences_weakly_identified",
  "fit": {
    "discounted_feature_mismatch_norm": 0.307151,
    "gradient_norm": 0.28103676,
    "iterations": 300,
    "l2_penalty": 0.01,
    "mean_demonstrated_action_log_probability": -0.518146,
    "temperature": 1
  },
  "implied_policy_at_start": [
    {
      "action_probabilities": {
        "repair": 0.99,
        "wait": 0.01
      },
      "most_likely_action": "repair",
      "state_id": "degraded"
    },
    {
      "action_probabilities": {
        "repair": 0.4287,
        "wait": 0.5713
      },
      "most_likely_action": "wait",
      "state_id": "healthy"
    }
  ],
  "method": "finite_horizon_maximum_causal_entropy_inverse_rl_v1",
  "revealed_preferences": [
    {
      "empirical_discounted_exposure": 1.8525,
      "feature": "health",
      "feature_mismatch": 0.307151,
      "model_discounted_exposure": 1.545349,
      "standardized_reward_weight": 2.620849
    }
  ],

Truncated for display — the full payload is 52 lines.

How it works

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

  1. 1 Infer aggregate linear state-feature rewards and their implied stochastic policy from sequential demonstrations using finite-horizon maximum-causal-entropy inverse reinforcement learning.
  2. 2 Evaluate the method-specific diagnostics and gates returned by the function, then abstain unless the declared decision clears them under locally governed thresholds.

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

  • Assignment or adoption timing, interference policy, overlap, attrition, and outcome availability match the estimand encoded by the method.
  • A causal label is warranted only when design and diagnostic requirements pass; otherwise treat the result as descriptive or abstain.

Minimum evidence

  • states: at least 2 rows/items
  • actions: at least 2 rows/items
  • demonstrations: at least 50 rows/items
  • feature_names: at least 1 rows/items

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

  • privacy-eligible state/action demonstration sequences
  • action-conditioned transition matrices
  • state reward feature matrix
  • state discretization and epoch
  • feature catalog
  • discount, entropy temperature, and regularization

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 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": "infer aggregate linear statefeature rewards and" }
  → finds "infer_revealed_policy_preferences_maxent_irl"

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

gitrevio_capability_run
  { "capability_id": "infer_revealed_policy_preferences_maxent_irl", "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

Audit causal claim negative controls

Gate a causal effect claim using prespecified negative outcome/exposure controls, Benjamini-Hochberg multiplicity control, and an omnibus chi-square falsification test.

Causal inference & experiment design

Audit cluster randomization integrity

Audit cluster-randomized experiments for practical baseline imbalance and differential outcome observation, with cluster-size-weighted standardized differences and assignment permutation diagnostics.

Causal inference & experiment design

Audit growth incrementality experiment integrity

Audit aggregate randomized growth experiments before anyone trusts channel incrementality: enforce unique experimental units, nondegenerate logged propensities, both arms, control-spend discipline, baseline balance, spillover and evidence gates; then estimate propensity-weighted baseline-adjusted contribution, cluster-unit bootstrap uncertainty and incremental return on spend.

Causal inference & experiment design

Audit joint outcome network integrity

Audit whether a company-specific Bayesian joint-outcome network is fit for reliance by validating point-in-time lineage, DAG and CPT completeness, effective support, protected-attribute exclusions, and strictly out-of-time outcome calibration against a baseline.

Sequential Bayesian & bandits

Audit staggered rollout identification

Audit staggered team-by-team adoption with not-yet-treated controls and require every simultaneous pre-period interval to fit inside a governed equivalence margin.

Causal inference & experiment design

Design balanced stepped wedge rollout

Assign teams or other aggregate clusters to capacity-constrained rollout waves with pair-exchange optimization of cumulative causal balance, represented population, and rollout risk.

Causal inference & experiment design

See every tool in Causal evidence & experiments →

Ready to See Your Engineering work clearly?

Request a free demo