Organizations make consequential decisions every day, increasingly through a mix of people and AI. Months later someone asks: how was this decided, by whom, against what information, with what review? Usually the answer has to be reconstructed from email and memory — slowly, unreliably, sometimes not at all.
The Decision Provenance Standard closes that gap. It is a record format: a defined way to write down how a decision was made, so it can be found, understood, and trusted later.
A small, fixed set of building blocks makes an organization's decisions affirmable, auditable, and resumable — whether a human or an AI did the underlying work — and that is how responsibility for decisions stays human as AI takes on more of it.
There are only a few blocks, and the set is closed: you cannot invent new ones at runtime. That is what makes records comparable and gradeable across any organization. The rest of this is those blocks.
A Charter governs a recurring kind of decision (for example, launch-readiness calls, or pricing exceptions) — not a single decision. It states up front, in writing: the single human accountable, the mode (human-led or AI-led), the schedule of records it will keep, and the triggers that force the decision to be reopened.
It moves through five states in one direction — open → mode declared → required fields set → fields completed → closed — and never backward. Requiring each field before the next state is what makes a Charter consistent by construction, not by after-the-fact review. It is the foundation everything else sits on.
Every decision is dispatched in one of two modes — this answers the question auditors care about most: who actually authored this?
When an AI authors content that reaches a person, the EU AI Act (Article 50) requires transparency. The Standard does not satisfy that obligation — it structures the inputs a human needs to satisfy it, via a small disclosure block with five required fields: the responsible human (declaring authority), the AI system identity, the jurisdiction, the content type, and the timestamp.
A useful rule: when records are later anonymized, four of the five fields survive unchanged — only the named person may be replaced by a placeholder. You can hide who without erasing that AI authored this, where, of what type, when.
A decision record moves draft → reviewed → affirmed. The load-bearing rule: a record only becomes affirmed when a named human performs an explicit act — a signature — and at that moment it is sealed (a tamper-evident hash is recorded).
There is no passive promotion. A record cannot drift into "approved" because time passed or a box auto-checked. This is what separates the Standard from a logging system, and it is the answer to "won't AI just decide everything?" — no, because a human signature gates every consequential record. (A four-layer safeguard, where no two layers share an actor, guards against a Mode-1 record quietly becoming AI-authored without anyone noticing.)
Three levels, each building on the last, and self-declared — there is no certifying body:
The most important slide. The records are audit-ready decision provenance — structured input that counsel and auditors convert into evidence, certifications, or attestations.
It is a small, self-contained set: the core Standard, four Companions (regulatory cross-references, a worked-charter library, implementation guidance, and the diagrams), an Appendix for governance and references, and an open-source reference implementation. It is published openly under Creative Commons Attribution; the name is a protected trademark; a Founding Steward governs the text but does not certify anyone's conformance.
Adoption is voluntary and incremental — the path is Charter-first: author one Charter for a real recurring decision, name the accountable human, declare the mode, produce records, self-declare Level 1, then work upward. One Charter, run honestly, is a conformant start.
A few fixed blocks — the Charter, the two-Mode authorship grammar, the human-affirmation-and-seal, and self-declared conformance levels — turn "how was this decided?" into a record you can find in thirty seconds and trust. The records are input to the humans who judge them; they never replace that judgment. And a human signature gates every consequential decision — which is how responsibility stays human as AI does more of the work.
All comments collected during your review. Edit, remove, or add general notes.
No comments yet. Navigate to any slide and click on highlighted sections to add feedback.
Jump to slide:
Enter ↵ to go • Esc to close
A short, plain-language walkthrough: the problem it solves, the Charter, the two Modes, Article 50 disclosure, the human-signature gate, conformance levels, and what it is not.