AI audit log

What each entry records

When and what

Timestamp, feature, provider and model, and the run id that groups the calls of one comparison run.

Input

The full system prompt and the batch of messages sent, the JSON schema the reply was held to, and the generation settings (temperature, output-token limit, thinking effort). Never the API key.

Output

The raw reply, the parsed labels when they matched the schema, the validation error when they did not, and the provider's message and category when the model refused.

Cost and speed

The final attempt's response time, the total over all attempts (retries and back-off waits included), the stop reason and the token usage the provider reported.

Human decision

Pending, accepted or rejected, with an optional note saying why: set here, or for a whole run next to its results. (The log also supports “edited” for features whose output a person can change; the spam comparison has none.)

Mock runs

Runs of the keyword rule that stands in for an LLM are logged too, marked “mock · no AI called”, so you can see what an entry looks like without a key.

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