Allow
Authorize the requested action, target and parameters for a limited time. Check the permission again at execution.
Evaluate each requested action against explicit permissions, current evidence and operating limits. Sordino returns allow, limit, hold or reject; a system-specific adapter applies the decision at the action boundary. Scope a pilot around one action and test the whole path.
Enforced at the tool gateway.
ALLOW
The exact request receives a short-lived permission.
Click a condition to break or restore it. Drag the model to look around.
This browser example never connects to a live system or executes an external action. Counts and conditions are illustrative; each integration needs tested rules and enforcement.
02 / A decision with a clear meaning
REQUEST → CHECK → ENFORCE → RECORDSordino evaluates a specific action against stated rules. The decision says what may happen next, why, and which conditions must remain true.
Authorize the requested action, target and parameters for a limited time. Check the permission again at execution.
Offer a smaller action within the rule. It receives no permission until the revised request has been checked.
Pause a new action while evidence, an approval or a required control is missing. Explain what needs to change.
Refuse an action that conflicts with a rule or exceeds the actor’s authority. A human review cannot erase a hard constraint.
03 / The rules behind a permission
SHARED WORKFLOW · SYSTEM-SPECIFIC CHECKSSyntony connects each control to a test, a reason it matters, and a named owner. When a relevant condition changes, the affected permission is checked again.
Check who is acting, what they may do, which resources they may use, and whether they may delegate. A retrieved document cannot grant extra permission.
Bind the permission to the tested version, configuration, intended use and environment. A model update or changed device condition can invalidate specific permissions.
Check source, age, units, reference frame and uncertainty. A recent timestamp alone does not make a quantum result or sensor reading reliable.
Track shared limits across agents, devices, retries and time. Reserve capacity before granting permission; commit or release it when the adapter reports the outcome.
Check that the required reviewer, controller and fallback are available in time. A robot’s local control must keep working when a remote network or AI component fails.
Link each finding and source correction to the decisions that depend on it. Hold affected new actions, assign the fix and check it before restoring permission.
A completed job is not automatically a valid result. Check whether an output may be released, reused or presented as evidence. If a source is withdrawn later, flag dependent decisions and record who has responded.
Find a failure → Define the rule → Test the control → Enforce the decision → Re-test after change
04 / Where the decision takes effect
ONE CONTRACT · DIFFERENT ADAPTERSThe policy gateway makes the decision. An adapter enforces it in the system and reports what happened. Evidence sources and controls are chosen for that particular integration.
| System | Enforcement point | Checks that matter |
|---|---|---|
| AI agents | Tool and API gateway | Delegation, target, action scope, source trust, shared spending or action limits. |
| Robotics & sensing | Command admission and local controller | Observation age, reference frames, operating limits and a tested local fallback. |
| Quantum computing | Job scheduler and result release | Backend, workload, device conditions, relevant calibration records and result metadata. |
| Telecom & energy | Configuration-change gateway | Shared dependencies, rollout limits, confirmation of the previous change and recovery options. |
| Labs, bio & materials | Experiment and analysis orchestration | Instrument state, source traceability, test conditions, permitted workflow and result validity. |
| Defense & space | Command approval and execution | Human decision rights, current state, command expiry, delayed links and local fallback behavior. |
| Neurotechnology | Measurement review and action gateway | Identity matching, signal quality, permitted use, review requirements and device-specific limits. |
For physical systems, fast protective functions stay in the local controller. If the gateway fails, it stops granting new permissions; an operation already in progress follows its approved fallback. The right response may be to hold, continue to a safe state, or transfer control.
05 / The decision and the outcome
A RECORD YOU CAN INSPECTKeep the exact request, policy version, supporting evidence, reasons, expiry and responsible roles. Then record the adapter’s response and the observed outcome separately.
A rejected request needs proof it was rejected at the action boundary. A permitted action needs confirmation of what actually ran. These records feed the next evaluation and any investigation.
BROWSER DEMONSTRATION · UNSIGNED
This record describes this example. It is not independent proof of a remote action.
06 / From a test to a deployed control
START WITH THE ACTION BOUNDARYStart with one workflow and a concrete failure. Establish what the adapter can enforce before expanding its permissions or connecting another system.
Run the proposed policy against recorded or mirrored requests. Inspect false blocks, missed failures and missing evidence with the people responsible for the system.
Put the adapter at the action boundary. Test bypass attempts, retries, replayed permissions, missing acknowledgements, timeouts and fallback behavior.
Recheck affected permissions when the system, policy or evidence changes. Test the original failure and the intended work the control must still allow.
Scope a Sordino pilot. Start with one action boundary, a named owner and an agreed baseline. Define the rule and integration demonstration, then measure false blocks, missed failures and response time. AI tool gateways are the initial pilot focus; other systems need a separately scoped adapter and test plan.
Request a pilotStart with one action boundary