Submission 48
Urbyte Pty Ltd
Submitter group: Companies.
An AI model identified the submitter group and wrote the summary, key points and recommendation summaries. Counts do not measure importance or merit.
Summary
Read the original submission for the full text.
Argues AI governance must cover evidence, persistent state, execution authority and verified outcomes, with proportionate, technology-neutral controls. Describes Urbyte's Keelo system and synthetic proofs, including later 100,000-command tests reporting recovery and denial across distinct routes. States these controlled results do not establish production readiness or universal security.
Key points
Quotes are the submitter’s own words. Text taken from documents, including scanned pages read by text-recognition software, can differ from the PDF. Check the original. Listed in the same order as on the submission’s page in the interactive guide.
Argues that governance of AI used in operational decisions must cover more than the model.
As artificial intelligence moves from assistance into operational decision making and execution, governance must extend beyond the model.
Argues that authority should be current, scoped to the organisation and action, and checked at the time of execution.
Authority must be current, scoped to the organisation and action, and checked when execution occurs.
Describes a synthetic 100,000-command proof reporting authorised execution, reconciliation, crash recovery and integrity checks, with matching reproduction.
Across 10,000 synthetic agents and 100,000 commands, the test produced 10,000 authorised effective external executions, zero unauthorised effective executions, zero duplicate effective executions, complete reconciliation, successful recovery from three forced runtime interruptions, and detection of all ten defined integrity attacks. A clean detached reproduction matched the deterministic evidence.
Argues that a provider receipt shows an action was processed but not necessarily that the intended result occurred.
A successful request or provider receipt proves that an execution attempt was accepted or processed. It does not necessarily prove the intended operational result.
Argues that controls should scale with consequence, reversibility, uncertainty and quality of evidence.
The degree of control should scale with consequence, reversibility, uncertainty and the quality of available evidence.
Describes a later proof reporting denial enforcement across materially distinct execution routes under the tested configuration.
Keelo enforced the same denied governed objective across materially distinct execution routes under the tested configuration. Separate routes converged on the same execution-authority boundary, and no prohibited route produced an unauthorised effective execution.
Recommendations
Recommendations are shown beside the submission’s own words. Listed in the same order as on its page in the interactive guide.
Require assurance of evidence, state, authority, execution and outcomes wherever AI influences or initiates consequential operational action.
Require assurance frameworks for evidence, state, authority, execution 1 Extend assurance beyond models and outcomes where AI influences or initiates consequential operational action.
Require evidence provenance, integrity references, observation time, organisational scope and applicable method or policy version for material decisions.
Require provenance, integrity references, observation time, 2 Require attributable evidence organisational scope and applicable method or policy version for material decisions.
Retain the rule used to derive a decision-relevant fact alongside source identity, integrity information, time and scope.
Evidence must retain source identity, integrity information, time, scope and the rule used to derive a decision relevant fact.
Require state to survive restarts, retain prior and resulting versions, and support deterministic reconstruction from durable events to identify the evidence, authority and policy applying to an action.
Operational governance cannot depend on transient conversational context. State must survive restarts, retain prior and resulting versions, and support deterministic reconstruction from durable events. This allows an organisation to determine which evidence, authority and policy applied when an action occurred.
Require current authority checks, accountable ownership, optimistic concurrency or equivalent stale-state controls, idempotency and duplicate safety.
Require current authority checks, accountable ownership, optimistic 3 Require execution controls concurrency or an equivalent stale state control, idempotency and duplicate safety.
Bind actions to the organisation, actor, accountable owner, permitted operation and current authorisation; fail closed if authority or version requirements are unmet.
Approval captured earlier in a process is insufficient if authority changes before execution. A governed system should bind the action to the organisation, actor, accountable owner, permitted operation and current authorisation. It should fail closed when authority or version requirements are not satisfied.
Require execution to produce a durable receipt linked to the governing decision and source evidence.
Execution must produce a durable receipt linked to the governing decision and source evidence.
Require durable execution receipts, independent outcome evidence where proportionate, reconciliation and explicit handling of unresolved obligations.
Require durable execution receipts, independent outcome evidence 4 Require outcome closure where proportionate, reconciliation and explicit unresolved obligation handling.
Define required operational properties and evidence rather than mandating a specific model, vendor or software category (technology neutrality).
Define required operational properties and evidence rather than 6 Preserve technology neutrality mandating a specific model, vendor or software category.
Permit staged technical proofs and pilots testing failure modes before production deployment, without treating synthetic proof as production readiness.
Permit staged technical proofs and pilots that test failure modes before 5 Use controlled proving environments production deployment, without treating a synthetic proof as proof of production readiness.
Scale controls with consequence/reversibility: provenance and review for low-consequence recommendations; authority, version checks, duplicate safety, receipts and closure for consequential/external actions; potentially independent verification, separated duties and stronger recovery for higher consequences.
The controls should scale with the consequence and reversibility of the action. A low consequence recommendation may need provenance and human review. A consequential or externally executed action should also require current authority, state version checks, duplicate safety, durable receipts and outcome closure. Higher consequence uses may require independent verification, separation of duties and stronger recovery controls.