Recommend safe contextual bandit action
Recommend contextual aggregate interventions with Bayesian reward learning only inside a posterior logistic harm constraint, explicitly falling back to a governed baseline when no arm is safe enough.
What it's for
Adds a safety layer to adaptive intervention search so a high expected uplift cannot outrank an unacceptable incident or guardrail risk.
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 |
|---|---|---|---|
| arms | array of objects (1 field) ≥ 2 items | Evidence | Yes |
| baseline_arm_id | string non-empty | Your calibration | Yes |
| current_context | array of number ≥ 1 item | Evidence | Yes |
| history | array of objects (5 fields) ≥ 100 items | Evidence | Yes |
| max_detail_rows | integer ≥ 1, ≤ 500 | Numerical control | Optional |
| maximum_harm_probability | number ≥ 0.001, ≤ 0.5 | Your calibration | Optional |
| maximum_iterations | integer ≥ 10, ≤ 1000 | Numerical control | Optional |
| minimum_safety_probability | number ≥ 0.5, ≤ 0.999 | Your calibration | Optional |
| posterior_draws | integer ≥ 200, ≤ 20000 | Numerical control | Optional |
| prior_precision | number ≥ 0.000001, ≤ 1000 | Your calibration | Optional |
| reward_noise | number ≥ 0.000001, ≤ 1000000 | Your calibration | Optional |
| seed | integer | Numerical control | Optional |
Each history
record
| Field | Type | Required |
|---|---|---|
| arm_id | string (non-empty) | Yes |
| context | array of number (≥ 1 item) | Yes |
| harm | integer (≥ 0, ≤ 1) | Yes |
| id | string (non-empty) | Yes |
| reward | number | Yes |
{
"arms": [
{
"id": "baseline"
},
{
"id": "safe_improvement"
},
{
"id": "risky_acceleration"
}
],
"baseline_arm_id": "baseline",
"current_context": [
0
],
"history": [
{
"arm_id": "baseline",
"context": [
-0.5
],
"harm": 1,
"id": "safe-bandit-0",
"reward": 0.9
},
{
"arm_id": "safe_improvement",
"context": [
-0.13366336633663367
],
"harm": 1,
"id": "safe-bandit-1",
"reward": 1.9732673267326732
},
{
"arm_id": "risky_acceleration",
"context": [
0.23267326732673266
],
"harm": 1,
"id": "safe-bandit-2",
"reward": 4.046534653465346
},Truncated for display — the full payload is 941 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.
{
"arms": [
{
"arm_id": "safe_improvement",
"historical_actions": 34,
"posterior_safety_probability": 0.935,
"predicted_harm_probability": {
"mean": 0.121863,
"p90": 0.181361
},
"predicted_reward": {
"mean": 1.946758,
"p10": 1.755583,
"p90": 2.152475
},
"probability_reward_exceeds_baseline": 1,
"safety_eligible": true,
"thompson_selection_probability": 1
},
{
"arm_id": "baseline",
"historical_actions": 34,
"posterior_safety_probability": 0.925,
"predicted_harm_probability": {
"mean": 0.120166,
"p90": 0.193824
},
"predicted_reward": {
"mean": 0.977374,
"p10": 0.765619,
"p90": 1.176239
},
"probability_reward_exceeds_baseline": 0,
"safety_eligible": true,
"thompson_selection_probability": 0
},
{
"arm_id": "risky_acceleration",
"historical_actions": 34,
"posterior_safety_probability": 0,
"predicted_harm_probability": {
"mean": 0.492152,
"p90": 0.597429
},Truncated for display — the full payload is 81 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 Recommend contextual aggregate interventions with Bayesian reward learning only inside a posterior logistic harm constraint, explicitly falling back to a governed baseline when no arm is safe enough.
- 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
- The likelihood or reward model, prior support, action logging, delayed outcomes, and any stationarity assumptions match the deployment process.
- Posterior probability and adaptive selection are model-conditional; they are not substitutes for randomized propensities or guaranteed safety.
Minimum evidence
- arms: at least 2 rows/items
- history: at least 100 rows/items
- current_context: at least 1 rows/items
- baseline_arm_id: required and organization-defined
How to validate it
Validate on future periods or held-out aggregate units, compare with a simple baseline, and require stability across plausible metric definitions and decision thresholds.
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
- binary harm within a fixed post-deployment window
- complete context vector with stable feature ordering
- current aggregate context snapshot
- baseline arm
- harm definition and maximum probability
- minimum posterior safety probability
- reward utility scale
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": "recommend contextual aggregate interventions with bayesian" }
→ finds "recommend_safe_contextual_bandit_action"
gitrevio_capability_describe
{ "capability_id": "recommend_safe_contextual_bandit_action" }
→ returns the input schema and agent guidance shown on this page
gitrevio_capability_run
{ "capability_id": "recommend_safe_contextual_bandit_action", "arguments": { ... } }
→ returns the result shown aboveWorks in Claude Desktop, Claude Code, Cursor, Cline, Continue.dev, Goose and Aider. See the MCP server.
Related tools
Aggregate risk register copula
Aggregate risk-register occurrence and lognormal severity marginals through a validated Gaussian copula into expected loss, VaR, CVaR, dependence amplification, and tail shares.
Allocate restless bandit interventions
Allocate scarce recurring interventions across evolving Markov units with Whittle indices, explicit indexability checks, and paired policy simulation.
Attribute commercial dependency tail loss
Calculate expected loss, VaR and CVaR for commercial value concentrated in shared technical components, then allocate every modeled tail-loss dollar exactly once across components with normalized negative-log survival hazard rather than overlapping leave-one-out sensitivities.
Audit commercial resilience claim integrity
Audit resilience ROI claims against a unique commercial-source to technical-component graph: recompute each action's avoided loss under joint failure scenarios, cap support at graph-derived value, detect duplicate effects, probability drift and weak evidence, and prevent overlapping component benefits from being sold twice.
Audit decision rank robustness smaa
Measure rank acceptability, regret, pairwise dominance, and central winning weights under uncertain criterion scores and bounded stakeholder weights.
Audit extreme metric tail dependence
Detect extreme metric co-exceedances beyond independence with empirical tail coefficients, permutation inference, practical magnitude gates, and FDR control.