Decisions
you can
defend.
The reasoning layer for autonomous agents. Every conclusion ships with the chain that produced it: the memories cited, the tools queried, the alternatives considered, the policy that gated the answer. No black box. No vibes. Grounds in Recall, gates Plan.
Chain-of-thought is just a step.
Reason is a five-stage pipeline. A question becomes a reasoned conclusion, grounded in facts, gated by policy.
Parse
Question → typed Intent. What is being asked? What's the success criterion? Ambiguous questions are rejected.
Ground
Pull relevant facts from Recall — prior decisions, entity context, relevant history. Every answer stands on evidence.
Reason
Derive the answer step-by-step. Each step cites a memory or prior conclusion. Nothing appears ex nihilo.
Verify
Is the answer consistent with the policy? Does it contradict any prior fact? Policy gates approve or reject.
Bind
Emit the answer + the chain that produced it. Every claim is auditable. Every step is replayable.
The receipt is the answer.
Not the conclusion alone — the entire chain of reasoning, the facts cited, the alternatives considered, the policy that approved it. Auditable. Replayable. Defensible.
Receipt {
id: "reason_adf2",
question: "Approve Q3 budget?",
answer: "Yes",
confidence: 0.87,
evidence: [
{ type: "fact", id: "f_123",
content: "Q2 spend $1.4M" },
{ type: "rule", id: "r_456",
content: "8% increase OK" },
{ type: "prior", id: "p_789",
content: "CFO approved similar" }
],
chain: [ /* 4-step reasoning */ ],
policy: "approval_gate_v2"
}GROUNDS
Every claim is backed by a fact from Recall, a rule, or a prior decision.
CHAIN
The reasoning steps that connected the facts to the conclusion. Each step is logged.
POLICY
Which policy gate approved this answer. Decisions are policy-compliant by construction.
REPLAYABLE
Given the same facts, the same reasoning re-runs and produces the same answer. Deterministic.
Every claim is a node.
Reason builds a claim graph. Each node is a derived fact. Edges represent logical dependency. The graph is versioned and persistent — the provenance of every decision.
Contradictions are caught.
Policy gates verify consistency before committing. Does the answer contradict a prior fact? Does it violate a rule? The receipt declares what policy approved it.
Factual
Answer must not contradict Recall facts.
Policy
Answer must satisfy all active policies.
Temporal
Timeline of facts must be coherent.
Scope
Answer stays within stated assumptions.
The numbers we're chasing.
Reason v0.1 targets below. Testing on 400 diverse reasoning tasks: decisions, approvals, analyses.
ANSWER ACCURACY
POLICY COMPLIANCE
Five traces. One answer.
Reason generates multiple reasoning chains in parallel and synthesizes them. Consensus is high-confidence. Disagreement surfaces faster than in a single chain.
The numbers in production.
Questions/day
Avg chain length
Policy compliance
Median latency
Composes with Plan + Recall.
Reason grounds in Recall for facts. Plan uses Reason to gate policy decisions. Together they form the cognitive stack.
Facts, entities, history. Reason queries it for grounding.
Derives answers. Gates decisions. Generates receipts.
Executes on reasoning gates. Prevents unsafe actions.
Reason vs. the alternatives.
| Feature | Reason | Chain-of-thought | Baseline LLM |
|---|---|---|---|
| Fact grounding | Recall | BYO | None |
| Policy gates | Built-in | Custom | None |
| Receipt generation | Typed | None | None |
| Consistency checking | Automatic | Manual | None |
| Multi-chain reasoning | 5 traces | 1 chain | 1 chain |
Join the waitlist.
Reason v0.1 ships Q1 2027. Research preview for design partners.
Get early access →