Start with the outcome

A Business Truth says what should be true, not how to get there.

Describe the desired outcome, constraints, and evidence you need. Axioms tooling makes that description explicit in Truth specs or Gherkin. Formation can explore it through Competition, Huddle, and alternative formations, including parallel attempts.

Zero, one, or several attempts may produce an acceptable result. A declared rule, another review, or a person can select among them. Selection does not itself authorize a promise: a commitment binds the organization to a specific version of the Business Truth.

The Nine Axioms

A mathematical contract with explicit conditions.

The nine axioms describe the kernel contract. Determinism and confluence require the stated conditions on inputs, transformations, policy, and execution. Fresh model calls from the same human intent do not promise the same answer.

1
Monotonicity
Preserve the admitted record
2
Determinism
Pin inputs and execution conditions
3
Idempotency
No duplicate admitted effect
4
Commutativity
Require compatible transformations
5
Termination
A bounded stop is not success
6
Consistency
Govern the active view
7
Starvation Freedom
State scheduler and budget assumptions
8
Confluence
Unique result under stated hypotheses
9
Observability
Keep evidence and decisions traceable
The Engine

Context reaches a fixed point, or the run says why it stopped.

Agents read shared context and propose facts; the engine decides what gets promoted. The loop ends when no new proposal remains, the budget is exhausted, policy blocks commitment, or human input is needed. A fixed point is stability under the loop's rules, not automatic business success. An unfavorable result can inform another recorded attempt with different resources, budget, or formation.

Convergence Loop

Cycle 1 ↻ Running
ComplianceCost
Proposals: 2 Facts: 0
Cycle 2 ↻ Running
RiskCompliance
Proposals: 3 Facts: 2
Cycle 3 ↻ Running
Synthesizer
Proposals: 4 Facts: 3
Cycle 4 ✓ Fixed point
No agents wake
Proposals: 4 Facts: 4

Context (append-only)

Seed: "Evaluate Acme AI as vendor"
Fact: SOC 2 compliant (conf: 0.92)
Fact: TCO €1.4M / 18mo payback (conf: 0.78)
Fact: Vendor concentration medium (conf: 0.71)
Fact: Recommend proceed, with conditions (conf: 0.83)
✓ Context converged — fixed point reached
Illustrative records, not measured confidence. Append corrections and superseding assessments; preserve prior evidence. A fixed point still needs acceptance and authorization.
Governance

A proposal becomes a fact only through a governed gate.

Agent output starts as a proposal. Admission checks authority, structure, and applicable evidence requirements. A promoted fact is a governed record, not automatic proof of a claim about the world. Binding commitments need their own explicit mandate.

Promotion Gate

○
Proposal Agent outputs a ProposedFact with confidence and provenance
○
Authority Check Cedar policy: does this agent have permission for this action?
○
Schema Check Does the proposal match the expected fact structure?
○
Evidence Check Does this proposal meet the declared evidence requirement? A score alone is not proof.
○
Promotion Admit a record with provenance. Binding commitment requires its own explicit authority.

Four Honest Exits

✓
Converged
Fixed point reached. Acceptance, selection, and commitment remain distinct.
⏱
Budget Exhausted
Time or token budget hit. Partial evidence surfaced.
⛔
Policy Blocked
Cedar rule prevents commitment. Decision logged.
✋
Escalated to Human
Judgment, evidence, or authority is missing. Context is handed off.
Illustrative admission checks and independent loop exits. Promotion alone does not mean convergence or commitment.

Keep a proposal that looked good until it did not. Append the rejection, new evidence, or superseding assessment without erasing the original. Learning changes the shared model for later decisions; it does not silently rewrite an existing promise.

The Stack

A kernel is one part of the environment.

Runtime-Runway supplies capabilities such as messaging and storage. The combined environment can be understood as an operating-system model for intent; Converge alone is the low-level contract.

Layer 4

Helm

Control surface

Express outcomes, inspect evidence, and exercise explicit authority.

helms.zone
Layer 3

Axiom

Specification tooling

Describe what should be true in Truth specs and check the specification.

axioms.zone
Layer 2

Formation

Intelligence layer

Explore alternative plans through formations, review, and recorded attempts.

formation.zone
Layer 1

Converge

Kernel contract

Shared context, facts, authority boundaries, and governed promotions.

converge.zone

"We don't believe the future belongs to louder AI.
It belongs to systems that know when to stop."