Consequence Before Action

Authorization answers whether a machine may attempt an action. It does not answer whether the organization is prepared for the consequence that action can create. As agents begin moving money, changing infrastructure, modifying access, binding contracts, or acting across enterprise boundaries, that second question becomes a control primitive of its own.

Permission is necessary but not sufficient

A perfectly authorized action can still create unacceptable exposure. A treasury agent may possess valid authority to execute a payment, yet the organization may already have exhausted the risk capacity allocated to that destination or settlement window. An infrastructure agent may be authorized to perform a change, yet the recovery resources required for a worst-case failure may not be available. A data agent may have permission to export a dataset, yet the privacy, contractual, or jurisdictional consequence capacity for that use may be unavailable.

Traditional controls often evaluate these constraints indirectly through policy, static limits, insurance, approvals, or after-the-fact incident response. Consequence Capital makes the pre-action capacity explicit. The question becomes: before this machine creates the consequence, what bounded capacity has been reserved to absorb, contain, recover from, compensate for, or settle the exposure?

That idea is deliberately separate from permission. Sufficient reserve cannot authorize an action that lacks delegated authority. Valid authority cannot waive a required reserve. The two conditions are independent gates because they answer different institutional questions.

Not every consequence is money

The word capital can tempt people toward a purely financial interpretation. StableMind uses it more broadly and more strictly. Consequence capacity can exist across financial, human, legal and regulatory, operational, data and privacy, security, environmental, resource, contractual, and irreversibility dimensions. Those dimensions are not assumed to be fungible.

That non-fungibility is essential. A large financial reserve does not cancel an unacceptable human consequence. Insurance coverage does not make an irreversible data disclosure acceptable. A low expected monetary loss does not neutralize a regulatory prohibition. Irreversibility cannot be discounted merely because the probability appears small. Cross-dimension netting would allow the system to hide the very exposures the reserve is supposed to surface.

The practical consequence is that a request can be authorized yet still unable to proceed because one required consequence dimension lacks capacity. That is not a failure of authorization. It is the architecture correctly refusing to treat permission as a universal solvent.

Reserve state should be conserved

A reserve model only works if capacity cannot be duplicated by delegation or concurrency. If a parent has a fixed consequence budget and delegates to several agents, those children cannot each behave as though the entire budget belongs to them. Shared capacity must be conserved, and locks must reflect concurrent commitments rather than optimistic snapshots.

StableMind models reserve state with explicit transitions such as AVAILABLE, DELEGATED, LOCKED, CONSUMED, DISPUTED, FROZEN, and RELEASABLE. The names matter less than the invariant: the system can account for where capacity went and prevent the same capacity from being promised to multiple consequential actions at once.

This makes consequence reserve closer to a control-plane resource than a dashboard metric. A score can be recomputed. Conserved capacity must be accounted for. A score can look green while two actors independently spend against the same exposure. A reserve system must prevent that double allocation or surface it as a violation.

Concurrency is where this moves from philosophy into engineering. Two individually permissible agent requests can become jointly impermissible if they contend for the same finite consequence capacity. The reserve plane therefore needs atomicity or equivalent conservation semantics at the point where capacity is locked. Otherwise a fast multi-agent system can create exposure between policy evaluations, with every individual decision appearing compliant while the institution as a whole exceeds the limit it intended to preserve.

Release belongs after evidence

If reserve exists before action, its release should not depend on the same executor that consumed it. The execution receipt may support the chain, but it should not automatically return capacity to AVAILABLE. The organization needs evidence that the relevant postconditions are satisfied and that any required settlement, recovery, dispute, or observation window has reached the governed state defined for release.

This is where Evidence and Proof of Consequence connect to Consequence Capital without collapsing into it. Independent observations can support a release gate. A VERIFIED outcome may be one input. But evidence does not itself become reserve, and a green proof interface does not have unilateral power to move conserved capacity. Release remains a governed transition with attributable policy and state.

The same separation protects against premature closure. An executor can report success while the reserve remains LOCKED because settlement has not completed. Evidence can be PARTIAL while capacity remains frozen. A mismatch can move exposure into DISPUTED rather than pretending the operation is finished. These states make uncertainty operationally visible.

The deeper reason is institutional

Consequence Capital is ultimately about preserving the relationship between machine speed and institutional responsibility. Agents can evaluate and act faster than human committees, but they should not be able to manufacture unbounded institutional exposure simply because the action path is technically available. The control plane must carry enough information about consequence capacity to keep acceleration inside accountable limits.

That requirement becomes more important in multi-agent and cross-enterprise systems. One agent can initiate a chain whose cost, liability, security exposure, operational disruption, or settlement obligation appears somewhere else. Without conserved consequence capacity, the first agent can externalize exposure to actors that never accepted it. A reserve boundary forces those consequences back into the authorization and execution architecture before the action becomes irreversible.

The useful design principle is therefore simple even when the implementation is not: no machine may create an unreserved consequence. Authority answers whether the machine may act. Consequence Capital asks whether the institution has explicitly made room for what that action can do. The future control plane needs both.

This also gives operators a cleaner failure vocabulary. A request can be denied because authority is absent, paused because consequence capacity is unavailable, frozen because exposure is disputed, or held because verification has not satisfied a release condition. Those are different institutional states, and treating them differently makes remediation, escalation, and audit far more precise than a single generic “agent blocked” event.

No machine may create an unreserved consequence.