From intent to commitment

A Business Truth describes an outcome, not a workflow

People and organizations start with what should be true: an outcome, its constraints, and the evidence needed before they can rely on a result. Axioms tooling helps express and check that description in Truth specs or Gherkin. The system can explore how to get there without asking the person to prescribe every step.

A Business Truth can exist before anyone makes a commitment. Validating its specification does not prove that the desired outcome has happened, and it does not authorize an action.

One problem can have several attempts

Organism helps choose formations for the problem, including Competition and Huddle. Alternative loops can run in parallel with different formations, resources, budgets, or approaches. Zero, one, or several may produce acceptable results.

If an attempt falls short, record what happened. A different budget, resource allocation, formation, or specification can support another attempt. Preserve the original conditions and outcome instead of rewriting the earlier attempt as a success.

A fixed point is not automatically a successful outcome

A fixed point means the context is stable under the loop’s defined rules. A favorable fixed point also meets the declared outcome and evidence requirements. Stability alone does not establish that a belief is true in the world.

An attempt can also stop on budget exhaustion, a policy block, or a need for human judgment. Those exits preserve useful evidence without claiming success. A solver that found no solution before time ran out has not necessarily proved that no solution exists.

Report what the evidence establishes

A heuristic, a feasible solution, a proven optimum within a formal model, searched evidence, and an independently checked proof make different claims. Keep those distinctions visible. The evidence requirement determines which kind is sufficient for this decision.

Two delivery plans might both meet the same capacity, budget, and service constraints. Either can be acceptable without being the same plan. Requiring a proven optimum is a separate business requirement, not something the solver or a confident model answer gets to assume.

A permitted exception may let an authorized person act with remaining uncertainty. That changes permission to act, not what the evidence proves. Some constraints cannot be waived by the current actor.

Selection is not yet a promise

A declared rule, another formation, or a person can select among acceptable results. Selection may be delegated under an explicit mandate. A selected recommendation does not grant itself the authority to bind the organization.

A commitment is an authorized undertaking tied to a specific version of a Business Truth. Others may make decisions in reliance on it. If selection and commitment happen together, the binding mandate must already exist.

The canonical commitment preserves the undertaking, its version, authority, constraints, evidence, and relevant uncertainty across explanations for different roles. A clearer explanation must not quietly become a different promise.

Keep the proposal that looked good until it did not

Shared context preserves the record of proposals, admitted facts, evidence, and decisions. Promotion admits a record under the context contract; it is not a universal proof of a statement about the world.

When new evidence changes an assessment, append the rejection, correction, or supersession and its reasons. Keep the original and what it influenced. A solver can use a filtered view without deleting excluded candidates from history.

A better idea does not release an existing commitment. Changes to an undertaking must be explicit because others may already have relied on it.

Reproducibility has a defined boundary

Converge is the low-level contract close to the mathematics: kernel, shared context, facts, authority boundaries, and promotions. Determinism and confluence claims require the complete inputs, versions, transformations, policy, and execution assumptions to be specified and satisfied.

A deterministic gate alone does not make every loop deterministic. The same human intent followed by fresh LLM calls need not produce identical proposals or results. Replaying recorded contributions is different from asking models to reason again.

Ferrox, Arbiter, Prism, and Soter contribute different forms of calculation, authorization, or assurance. Their claims must retain their own scope. Prediction and learned models can inform a proposal without establishing proof or granting authority.

Learn for the next decision without rewriting this one

The organizational adaptive cycle is Coordinate → Commit → Update Priors. Coordination builds a shared model from distributed evidence and disagreement. Commitment binds the organization. Comparing expected and actual outcomes can improve the model used in later attempts.

Storing or retrieving an episode is not sufficient: learning must change future reasoning or choices. Keep unsuccessful attempts and dissent alongside the selected results. Updating the shared model does not silently change a promise already made.

The parts of the system

Converge

The kernel contract for context, facts, authority boundaries, promotions, and bounded convergence.

Organism

Formations, alternative attempts, planning, review, and learning.

Axioms

Human-readable Truth specs/Gherkin tooling, validation, and checkable artifacts.

Helm

The operator surface for expressing outcomes, examining evidence, and exercising authority.

Runtime-Runway

Runtime capabilities such as messaging and storage around the components.

The combined environment can be understood as an operating-system model for intent. Converge’s kernel alone is not that whole environment. This is a responsibility map, not a claim that every described capability has been independently verified in a deployed system.