GuildBuild Inc.
← All insights

KN-11 · Agent claims and observations

What a model assumes when you leave a gap

An agent can fill a gap with a plausible assumption and report it as something it observed. GuildBuild recorded two examples of the pattern, a session without tools and a seat that supplied times without a fresh clock reading.

4 min readGuildBuild Team
AI agentsVerificationAgent reliabilityHarness engineering
In this article

Give claims a source

A transcript can show whether a tool call happened. A fresh clock reading can establish the time. The two cases below show what a report can get wrong when those observations are missing.

Two starts without tools

The team made two fresh starts on a bare Claude Code seat with no project instruction file and zero tool definitions. It sent one word in each start.

  1. In the first start, the model printed 13 shell commands and then said its commands had returned no output. The transcript recorded zero tool calls.
  2. When asked whether it had run anything, the model said it had not. It also said it had invented the filenames.
  3. The second start ended in a malformed tool call.

The harness’s own prompt still mentioned tools. That leaves at least two possible influences, model training and the prompt. The study could not separate them.

Three records without a fresh clock reading

  1. One morning, a lead seat wrote clock times for overnight events. When a clock was checked, three of the timestamps were still in the future.
  2. About eight hours later, two brief headers said 14:20 and 14:50. Clock readings were about 14:11 and 14:18. The morning’s instruction named notes and reports, leaving brief headers outside its wording.
  3. Twelve days later, the seat used 13:10, 13:30 and 13:50 on three briefs. The next clock reading was 13:28.

The seat explained that it had extrapolated from one clock reading and its sense of elapsed time. The reported times repeatedly ran ahead of observations. In a sequence of agent handoffs, a guessed header can make a recipient appear late even when the underlying record shows otherwise.

Attach the observation to the claim

A written rule about notes and reports did not stop the later header errors. The team moved the check to the moment it needed the fact. A timestamp needed a fresh clock reading or a recorded commit timestamp from the same call. When neither was available, it used event order. A claim that a command ran needed a matching tool call in the transcript.

Figure 01Evidence comparison

Which record can support the claim?

Command execution

Reported claimThe first start printed 13 commands and claimed they returned no output.

Observation to checkThe transcript showed zero tool calls. Printed text established only what the agent wrote.

Clock time

Reported claimTwo brief headers said 14:20 and 14:50.

Observation to checkClock readings were about 14:11 and 14:18. Take a fresh reading when writing, or preserve event order when time is unknown.

The tool case used two fresh starts. The time case covered three records from one seat. These examples describe how claims were checked. Frequency remains unknown.

Check what the setup leaves unsaid

  1. List facts the agent needs to report, such as time, tool availability, working folder or the person receiving a handoff.
  2. Choose the evidence source for each fact. Use it in the turn that writes the report.
  3. When a source is missing, record the known sequence or state the uncertainty.
  4. Check execution claims against the transcript and inspect the result.