AI audit log

What each entry records

When and what

Timestamp, feature, provider and model, and the run id for evaluation runs.

Input

The full system prompt and user prompt that were sent. Never the API key.

Output

The raw reply, the parsed JSON if it validated, and the validation error if it did not.

Score

For calls in the Hangman evaluation, how the reply was scored: hit, miss, repeat, invalid, truncated or refused.

Cost and speed

Latency in milliseconds, attempts (rate-limit retries) and token usage as reported by the provider.

Human decision

Pending, accepted, edited or rejected, set by you here or next to the output.

Entries are checked for your key before they are stored. Your provider keeps its own records under its own terms. How AI is used on this site is set out in the AI use statement.