Trace the evidence path before trusting an agent decision
A simulator can produce the needed value while an agent answers from a different file. A code warning can sound convincing while its cited program relationship is wrong. A safety monitor can run while a detected result fails to reach its blocking decision.
The common task is to follow evidence to the decision it must support. Each study tests a specific connection and bounds its conclusion to the evidence it observed.
What you will understand
How can changing the delivery of a tool report change an agent’s answer when its computation is held fixed?
What evidence should a code verdict cite, and how can a framework test whether safety evidence reaches a decision?
What can an auditor establish from recorded runtime events, and where does that evidence stop?
The route to an answer changes what gets read
Imagine asking an agent to run a chemistry simulation and choose the answer printed in its report. The simulator, PHREEQC, produces a formatted report whose numbers match the precision of the answer options.
The agent can also write a shorter output file with more digits. If the report appears directly in the tool response, the relevant number is already in view. If the tool instead returns a table of contents, the agent must open a report section to see it.
The delivery choice mattered in a paired replay with GPT-5.2. Researchers fixed the simulation files and the visible history for 131 numeric questions, then continued each case three times per delivery condition.
They calculated each question’s average difference in correct-answer rate and averaged those differences across questions, counting missing answers as incorrect. Delivering a table of contents reduced accuracy by 9.4 percentage points relative to delivering the report inline.
The comparison concerns those frozen continuations, where the pending tool result was the changed input.
That result poses a broader question: where can evidence be lost or checked between a tool or program and a consequential decision? An answer, a code-security verdict, and a decision to block an unsafe action depend on different kinds of evidence.
Their failures can look similar from the outside: the system finishes and presents a decision. To understand the decision, a reader needs to identify what evidence was available, what the system used, and whether the claimed connection can be inspected.
Where the lesson goes next
Check the code behind a verdict
Follow evidence to the safety decision
Replay what the runtime recorded
Test the point where evidence matters