Authority Leases and Concurrent Evidence Failure in Safe Agentic Execution
Separating Evidence Qualification from Action-Time Authorization. LatentAtlas treats a queued action as a bounded authority contract that must still be valid at execution time.
The action-time authority problem.
Agentic systems often decide before they act. A model, worker, or automation can inspect evidence, queue an action, and execute later through a scheduler, retry loop, delegated worker, or tool call. Between the decision and the action, the world can change.
Concurrent Evidence Failure names the related failure family: true observations can fail action-time validity when identity, freshness, authority, visibility, or materialization timing no longer align.
Frozen review outcomes.
The unresolved bucket is not discarded. It becomes a measured state of the review instrument: the packet needs more evidence before it can become an outcome label.
| Outcome | Count | Meaning |
|---|---|---|
| Packet-supported block | 99 | The conservative block or revalidation decision was supported by packet-visible evidence. |
| Packet-unsupported block | 22 | The packet-visible evidence indicated the conservative block was not warranted. |
| Insufficient evidence | 25 | The packet did not contain enough evidence to judge correctness or falseness. |
Findings.
These are descriptive findings from a frozen masked review set, useful because they turn action-time uncertainty into reviewable evidence states.
False blocks are measurable
22 of 146 reviewed packets were packet-unsupported block judgments: 15.1% of all reviewed packets and 18.2% of resolved packets.
Evidence insufficiency is a result
25 reviewed packets remained insufficient-evidence judgments, or 17.1% of all reviewed packets.
Temporal authority matters
14 of 48 latest product-detail-page temporal-authority packets were packet-unsupported block judgments.
Identity conflict was block-grade
4 of 4 masked identity-conflict packets were packet-supported block judgments in the reviewed set.
Blocked PDP was not outcome evidence
7 of 7 access-blocked product-detail-page packets remained insufficient-evidence judgments.
Deduplication stayed stable
146 event-level units, 146 subject clusters, and 0 duplicate subject clusters.
Concurrency is a bounded research track
6 local failure families are carried forward as testable research tracks, not broader population rates.
Authority leases as a control surface.
A lease ties a prior decision to the actor, action, target, evidence, permission, policy, impact ceiling, and expiry state that made it valid. At action time, the system checks whether the lease still holds.
Decision-time retrieval, proof, and state are captured before the action is queued.
The action is bound to actor, target, permission, source, impact ceiling, and expiry.
At execution time, the system checks whether identity, state, policy, and permission still match.
The worker executes, revalidates, blocks, or sends the packet to manual review.