Where Certainty sits in the stack
Certainty Labs is defining the layer between AI and agent capability and consequential enterprise action — the infrastructure that determines whether a proposed action is sufficiently evidenced, bounded, and authorized before it executes.
Foundation models and agent capability that can generate outputs and propose actions.
Workflow and tool-calling runtimes that sequence what an agent attempts to do.
Determines whether a proposed action is sufficiently evidenced, bounded, and authorized before it changes enterprise state — and preserves why.
Systems of record where an authorized decision takes effect as consequential action.
Architectural direction: we are progressively extracting these decision primitives from applied systems into reusable infrastructure.
What Certainty owns, and what it deliberately does not
Certainty Labs owns the decision layer itself. It does not own model capability, agent orchestration, or the systems of record on either side of it.
A decision as a connected record
The system must preserve the reasoning and authority surrounding a decision—not only its final output.
System behavior, not slogans
Claims earn confidence through traceable evidence, not persuasive output.
Unknowns and conflicts stay attached to the decision instead of disappearing behind a score.
Consequential action records who had the authority to approve it.
A dependable system can stop when evidence or authority is insufficient.
Reasoning, authorization, action, and outcome remain connected over time.
Shared foundations never silently absorb a product’s entities, context, or workflows.
Architecture as commitment
Accountability is represented as a system concern—not added later as marketing language.
readonly outcome: string;
readonly decisionStatement: string;
readonly justifyingClaimRefs: readonly Claim["draftId"][];
readonly uncertainty: Uncertainty;packages/agent-contracts/src/entities/recommendation.tsThe contract keeps an agent's outcome, decision statement, justifying claim references, and uncertainty attached to one proposal shape — evidence and claims must resolve to a recommendation before a human decision is made.