Model failures need more than success and error.

Video caption

Model failures need more than success and error.

Defaults can erase high-value work.

Completed, refused, unavailable, uncertain.

The case stays owned until a real decision exists.

#EricFieldNotes

Full written post / accessibility read

An agent may get a complete answer, an explicit refusal, no service response or an answer that cannot be trusted. Compress those into success and error, and downstream code will eventually interpret one as the other.

Picture a test report with a known owner. The model adapter returns a refusal, but the ticket writer maps missing classification to informational. The issue appears resolved precisely because the adapter lacked a refused state.

Completed must still pass business checks. Refused preserves the case for human or approved alternate handling. Unavailable permits bounded retry, then alert. Uncertain captures the failed evidence test. Never invent a successful empty value.

Inject each result in a disposable test and inspect the ticket store after the whole workflow. Do this because a beautiful error message is irrelevant if the database still marks the case complete.

#EricFieldNotes

Four-beat scene transcript

1. Model failures need more than success and error.

An agent may get a complete answer, an explicit refusal, no service response or an answer that cannot be trusted. Compress those into success and error, and downstream code will eventually interpret one as the other.

Visual: A refusal is different from timeout or bad evidence.

2. A blank response is not a blank case.

Picture a test report with a known owner. The model adapter returns a refusal, but the ticket writer maps missing classification to informational. The issue appears resolved precisely because the adapter lacked a refused state.

Visual: Defaults can erase high-value work.

3. Make the adapter return a tagged result.

Completed must still pass business checks. Refused preserves the case for human or approved alternate handling. Unavailable permits bounded retry, then alert. Uncertain captures the failed evidence test. Never invent a successful empty value.

Visual: Completed, refused, unavailable, uncertain.

4. Test the state transition, not the label.

Inject each result in a disposable test and inspect the ticket store after the whole workflow. Do this because a beautiful error message is irrelevant if the database still marks the case complete.

Visual: The case stays owned until a real decision exists.

Research and claim limits

• OpenAI Usage Policies (S133)
• Anthropic Usage Policy update (S134)

The examples identified as illustrative or simulated are design probes, not reported incidents. Vendor specifications do not establish workload performance.