Separate an agent execution log from independent verification evidence
An agent’s success message and an outside observation of the result are different kinds of evidence.
One log can appear to become two witnesses
An agent reports that it finished a task, and a summary also calls the task successful. That can look like two confirmations even when both descriptions come from the same execution log. Truthound can check a separately prepared evidence table for declared origins and states. It does not monitor an agent at runtime or create an independent witness to the action.
The discussion concerns protections outside a model
CSA and Google held an agent-security tabletop exercise on 14 October 2026; the outcome statement appeared the following day. It discusses protections outside the model, runtime monitoring and access controls, alongside independent verification and model diversity. This is not a measured effectiveness trial. An adopting organisation still needs to define which records it will accept as separately sourced evidence.
Put four states beside each other
A proposed evidence bundle distinguishes intended call, authorized action, actual execution and external observation. Link call identity, actor, authorization reference, execution state and observation origin. A new summary of the log does not count as a new observation. Even differently named observers may depend on the same original record; keep that dependency visible rather than inferring independence from the number of reports.
Choose a confirmation route by consequence
For a low-impact read, a person comparing the execution record and returned value may provide an adequate review route. An irreversible write may require a target-system change record or separate approval process. The responsible owner chooses the route. In either case, a single success field should not silently stand for intent, execution and verified outcome, since those states can disagree in meaningful ways.
Expose gaps in the proposed evidence table
Truthound Core Python SDK 3.3.4 (open-source SDK, Polars backend) documents ColumnPairInSetValidator for allowed combinations supplied by the caller. It could check state and observation-origin pairs, with separate NotNullValidator checks for required execution or observation references. The organisation supplies valid_pairs and null handling. A row labelled external is not itself proof that the observation is independent or that a purported verification step actually ran.
Keep validation separate from runtime security
Pandera 0.35.0 (MIT open-source Python DataFrame validation library; pandas, Polars and other supported backends) custom checks can express the same evidence contract. If such a contract is already maintained, inspect who verifies independence before considering a replacement with Truthound. Both routes remain separate from runtime permissions, hardened execution and target-system audit records. Comparing table-validation interfaces does not establish which overall agent-security system is better, or demonstrate that a model cannot perform an unauthorized action. The documented custom-check example uses a pandas Series; do not assume its predicate transfers unchanged to another backend.
Pause when independence has no evidence
Start with one tool-call type and distinguish records recycled from the same source from separately observed results. Where outside confirmation is unavailable, retain an unverified state and reconsider the permitted action scope. Do not expand autonomy just because the evidence table is complete. The decision between approaches should turn on the available confirmation route, rather than on how many model-generated success messages have accumulated.
Conditions
Define evidence origins and approved execution-observation combinations in the audit table.
Alternatives
Keep an equivalent Pandera 0.35.0 evidence contract separate from runtime monitoring.
Limitations
An external label does not prove independence or that a verification step ran.
Evidence
- Event checked
- Subject checked
- Modified
