ALLOW / DENY / ESCALATE
Only ALLOW can create a bounded grant. The grant stays tied to the exact action, material, state, scope and expiry.
Control the consequence, not the intelligence.
SnapSpace governs the path from policy and approval through authority, execution, STOP, effect, reconciliation and evidence.
Policies become runtime controls. Approvals actually gate actions. Authority stays bounded during execution. Active work can be STOPped. Uncertain effects are reconciled. Evidence records what was authorised, attempted, applied and resolved.
AI is moving from answers to actions.
Agents can call tools, move money, change systems, coordinate infrastructure and operate autonomous machines. Written policy, model alignment and observability are not enough once software can cause an external effect.
The control problem is no longer only what the model says. It is what the system is allowed to do.THE CONTROL PATH
Policy becomes enforceable control at runtime.
POLICY→RUNTIME CONTROL→DECIDE→BOUNDED AUTHORITY→EXECUTE→EFFECT→EVIDENCE
STOP is separate.
STOP interrupts active consequential work, revokes remaining authority in scope and contains further effects where technically possible.
Unknown stays unknown.
Acknowledgement is not effect. If the outcome is uncertain, SnapSpace reconciles the original attempt instead of blindly retrying or inventing success.
PROOF
Qualification across scale, failure modes and environments.
SnapSpace does not lead with a small pass counter. The programme evidence spans software, distributed software/SIL, real GPU and physical-simulator environments with repeated qualification, adversarial testing, deterministic replay and sealed evidence.
exact software/SIL control-equivalence stress point
pinned bounded multi-domain software composition
two independent runs with the same semantic sequence
frozen offline qualification; receiver acceptance remains external
byte-identical repeat followed by STOP-enhanced regression
closed-loop simulator qualification with identical semantic result
Each figure keeps its own claim ceiling. Repeated runs and inherited regressions are not added together to create a synthetic grand total.
Inspect the proof hierarchy →OPERATING BOUNDARIES
The same control model has been exercised across very different systems.
AI agents
Runtime authority, exact-use execution, STOP and receipts
Distributed compute
Admission, state identity, multi-kernel control and real-GPU reference
Institutional systems
Release revalidation, receiver admission and responsibility truth
Autonomous systems
Closed-loop SITL, sensor response, endurance and interruption
Sovereign domains
Bounded projection, independent authority, partition and replay
Transactions
Attempt identity, uncertain effects and reconciliation
The domain is allowed to change. The control discipline is not: explicit authority, bounded execution, STOP, truthful effect state and evidence.
See qualification by domain →Governance does not stop when policy is written.
SnapSpace is designed for the complete consequence path: policy becomes a runtime control; approval gates action; authority remains bounded; execution can be interrupted; uncertainty is preserved and reconciled; evidence records what actually happened.
CONTROL AT RUNTIMEBOUND AUTHORITYSTOP ACTIVE WORKRECONCILE UNKNOWN EFFECTSPROVE THE OUTCOME
INDEPENDENT DOMAINS
See the federation modelShare state, not authority.
SnapSpace allows independently controlled domains to exchange bounded state without treating visibility, transport or shared context as permission to act. The receiving domain keeps its own admission and consequential authority.
Independent engineering for consequential AI control.
SnapSpace is built as an evidence-led control programme: qualified results keep their environment and limits, negative results remain visible, and open questions stay research until they pass the relevant gate.