What is Decision Record?
ObservabilityThe audit artifact that survives an agent's run: for one output, on one date, given one input, a structured record of why the system produced this result. It captures what a sparse log cannot: the model version, the prompt in effect, the retrieved context, and any human oversight action taken.
Why It Matters
When a regulator, a customer, or your own incident review asks why the system produced a specific output, a timestamp and a log line are not an answer. EU AI Act Article 12 makes this concrete: high-risk systems must record events so that behavior can be reconstructed, not just confirmed to have occurred, with retention of at least six months.
The decision record exists because reconstruction is the actual requirement. “The agent did something at 14:32” is a fact. “The agent produced this output because it ran this model version, under this prompt, over these retrieved documents, after this human approved it” is an answer.
What a Record Captures
Decision ID. A stable identifier that ties the record to the run and the output.
Timestamp. When the decision was produced.
Model version. The exact version at inference time, not the API alias, because providers roll weights behind the same name.
Prompt hash. A hash of the prompt template in effect at that moment, so later prompt edits do not rewrite history.
Retrieved context IDs. The exact set of documents passed to the model, pinned even as the index keeps changing.
Input and output. The data in and the result produced.
Oversight action. Any human intervention on the record: approved, rejected, modified, or expired.
Where It Breaks
Three gaps account for most failures, and none of them look like gaps until someone asks the question. Prompts get edited without a version tied to decisions. The retrieval index keeps changing while nothing captures which documents a given run actually saw. And model providers roll versions behind aliases, so a decision in March and a decision in July may have run on different weights while sharing a model name.
Each gap is a schema decision that has to be made before the system ships. None of them can be retrofitted onto data that was never captured.
How Flytebit Handles It
Decision records are part of the schema we instrument in every observability engagement, alongside the trajectory that carries them. The compliance framing and the three retrofit-proofing gaps are documented in Observability in Agentic AI.