Verified Decision API

Approve. Deny.Or safelyrefuse.

One API call returns an accountable outcome grounded in your rules and facts—with cited proof, explicit uncertainty, and no path from an ungrounded model answer to approval.

Decision runtime
Three outcomes with safe refusalCited proof for every verdictBatch + idempotency for production

Endpoint response

POST /v1/decisions

Audit-ready

Deterministic payload

{
  "decision_id": "dec_9f2c4c1e…",
  "decision": "approved",
  "confidence": 0.85,
  "rationale": "Within return window",
  "proof": {
    "cited_rules": ["may_refund"],
    "symbolic_steps": 2,
    "neural_steps": 0,
    "proof_score": 0.90,
    "proof_status": "grounded"
  },
  "action": {
    "name": "issue_refund",
    "ready": true,
    "rule_name": "may_refund"
  },
  "grounding": { "status": "grounded" },
  "usage": { "billable": true, "elapsed_ms": 12 }
}

Outcomes

3

Default mode

Safe

Proof

Cited rules

Idempotent replay

24h

Classify

Returns approved, denied, or refused. Safe mode is the default, and ambiguity resolves to refusal instead of a guessed authorization.

Three outcomes • safe by default • human-review signal

Prove

Returns cited rules, symbolic and neural step counts, proof strength, grounding status, and an optional full trace for forensic review.

Cited rules • proof score • grounding • optional trace

Operate at scale

Score batches with per-item evidence, retry safely with idempotency keys, inspect recent decisions and stats, and meter each outcome explicitly.

Batch • idempotency • analytics • usage metering

Reference Proof Harness

Safety is measured, not implied.

The repository harness runs nine labeled commerce, access-control, supply-chain, and lending cases through the real backward-chaining engine. It exits non-zero on an unsafe approval or missed quality floor.

cargo run --example verified_decision_harness

100%

Decision accuracy

7 / 7 grounded cases

100%

Refusal correctness

2 / 2 under-grounded cases

100%

Proof validity

4 / 4 audited approvals

0

Unsafe approvals

CI failure threshold

Why Verified Decisions

Accountability is part of the response shape

This is not another classifier wrapped around an LLM. The API is built so a consequential outcome must be grounded, provable, and safely refusable.

Refuses Safely

Refused is a first-class outcome when the engine cannot ground a decision well enough to be accountable.

Ungrounded, ambiguous, failed, or human-review cases stop instead of becoming confident approvals.

Proves Every Verdict

Every accountable outcome carries the rules, proof strength, grounding status, and reasoning step counts behind it.

Operators can inspect cited rules and see exactly how much symbolic, neural, and planner work contributed.

Reasons Recursively

Backward chaining resolves policies that depend on other policies instead of flattening them into prompt text.

Transitive access, chained eligibility, and multi-level product or supply-chain logic remain explainable as proof paths.

Contains the Neural Layer

Models read messy language and propose candidates. Explicit rules and facts remain the source of authority.

The response exposes neural step counts, making model involvement visible instead of hiding it behind a fluent rationale.

Measures the Outcome

Decision IDs, latency, billing, proof depth, and cited rules make the API operational as well as explainable.

Recent-decision and aggregate-stats endpoints support monitoring, while response headers expose cost and quota inline.

Decision Pipeline

From natural language to an accountable verdict

Every call follows the same five-stage path: frame the goal, ground it in policy, prove it recursively, verify the evidence, and classify the outcome.

Turn a request into an accountable goal.

1. Frame

The API frames the natural-language query and optional action as a decision that must be authorized, not as a prompt that merely needs an answer.

  • Accepts a query plus an optional action such as issue_refund.
  • Defaults to safe mode for the most thorough verification path.
  • Carries caller session context and bounded solver overrides when needed.

A consequential request enters as a decision goal, not an invitation to improvise.

Control Flow
1
Natural-language query
2
Action under authorization
3
Decision goal
Decision Workloads

Put a provable authorization step in front of consequential actions

Use the same three-outcome API across CX, subscriptions, fulfillment, and agent tool use—one decision at a time or in metered batches.

Refunds and Replacements

Merchant problem

Teams need to move fast on damaged orders and wrong-item tickets without opening the door to leakage, abuse, or inconsistent policy handling.

How StateSet solves it

Send the request, action, order facts, and refund rules to the Decisions API. It returns approved, denied, or refused with the exact policy evidence needed before money or inventory moves.

Outcome profile

  • Auto-approve low-risk cases with explicit boundaries.
  • Block or escalate suspicious cases instead of hallucinating customer service gestures.
  • Give operators a clear audit trail when exceptions happen.
Why the Contract Matters

A fluent answer is not an authorization

Prompts and retrieval can improve language. The Decisions API adds the outcome contract, proof, refusal path, and production controls consequential actions require.

What defines correct
A plausible model response shaped by prompts and retrieval.
A derivation over explicit rules, facts, and grounded entities.
What happens when evidence is weak
The model still tends to produce a confident answer.
The API returns refused with a reason and human-review signal.
How multi-hop policy works
Longer prompts, larger contexts, and implicit model reasoning.
Recursive backward chaining with inspectable sub-proofs.
What survives audit
A generated rationale after the model has answered.
Decision ID, cited rules, proof score, grounding, and optional trace.
How production retries work
Caller-specific deduplication and uncertain duplicate cost.
24-hour idempotent replay plus explicit usage and quota headers.
StateSet Thesis

Let models propose. Make consequential actions prove they are allowed.

StateSet's Verified Decision API adds approved, denied, and safely refused outcomes—with recursive policy proofs, explicit grounding, safe retries, batch processing, and outcome-level metering.

Approved

grounded authorization

Denied

explicit policy block

Refused

safe human review

What merchants get

Request-scoped facts and rulesBatch decisions24-hour idempotencyRecent decisions and stats