A decision receipt is a durable, re-readable record of a consequential AI recommendation and what actually happened after it: what was recommended, the options, how fragile the choice was, and, once known, the real result. The log is append-only, so the trail cannot be quietly rewritten to look smarter in hindsight.
The problem
When an AI makes a call, there is usually no paper trail. You cannot audit why, you cannot tell whether the choice was a coin-flip or a landslide, and nobody records what actually happened, so the system never faces its own track record.
What it does
- Records the call. What was recommended, the options, a receipt id, all append-only.
- Scores fragility. The robustness check perturbs the scores with small jitter and reports how often the top choice flips. A recommendation that flips 41% of the time is a coin-flip wearing a confident answer, and now you know.
- Closes the loop. Record the real outcome later; it merges into the receipt without overwriting the original call.
The honest part
A withheld recommendation is receipted as withheld, content absent, so the log is never a side channel. A fuzzy outcome match is labeled as fuzzy, never as exact. Plug in your own evidence checks; a failing one is disclosed, not dropped.
Get it
Free, Apache 2.0, zero dependencies: github.com/davedepew/decision-receipts
pip install decision-receiptsIt is part of The Operator’s Honesty Stack, a set of open-source engines that each make one kind of AI dishonesty structurally impossible. The engines are free; the Operator’s Membership is where I publish the playbooks that deploy them, for $19 a month.
FAQ
What is a decision receipt?
An auditable record of an AI recommendation, its fragility, and its real outcome.
Is it free?
Yes. Apache 2.0, zero dependencies.