Specification · Request for comment
DDR-1 — The Defensible Decision Record
| Valid time | When the fact held in the world | Distinguishes a stale fact from a false one |
|---|---|---|
| Transaction time | When the system recorded the fact | Establishes what could have been known |
| Decision time | When the conclusion was issued | Fixes the coordinate for reconstruction |
Collapsing these to a single timestamp is the most common and most damaging defect in deployed systems. A record MUST carry all three.
5. Required content
A conformant record MUST contain the following.
5.1 Identity
record_id — stable and unique. decision_time — when the conclusion was issued, with timezone.
5.2 Claim
claim — the conclusion, stated so that a non-specialist can evaluate it. disposition — the action recommended or taken. authority — the scope the decider was permitted to decide within.
5.3 Temporal coordinates
transaction_horizon — the latest transaction time of any evidence the decider could see. valid_window — the period in the world the claim concerns.
5.4 Evidence
For each item: a content-derived evidence_id; source; source_version; recorded_at; valid_from and valid_to where known.
Evidence MUST be addressable and MUST NOT be mutable in place. Where two items derive from a common upstream source, that relationship MUST be recorded — correlated evidence treated as independent inflates confidence without adding information.
5.5 Inference
rationale — the argument from evidence to claim, in language a reviewer can dispute. method — the deciding model or rule, with version and any parameters affecting output. load_bearing — which evidence items the conclusion depends on, distinguished from those that merely corroborate.
5.6 Alternatives
considered — conclusions examined and not adopted. rejected_because — the reason for each.
A record that reports only what was concluded, and never what was ruled out, gives a reviewer no way to assess whether the decider looked in the right places.
5.7 Falsifiers
would_change_if — the conditions under which this conclusion would not hold, stated at decision time rather than reconstructed afterward.
5.8 Confidence
confidence is OPTIONAL. Where present, it MUST carry calibration_reference identifying the measurement establishing what the value means: the population, the period, and the observed outcome frequency.
An uncalibrated confidence value MUST NOT be recorded. It is not an estimate, and its presence misleads.
5.9 Actor
decider — human, automated, or composite, with the delegation under which it acted. reviewed_by where a human affirmed, modified, or overrode the conclusion.
5.10 Supersession
Records are append-only. A correction is a new record carrying supersedes; the superseded record MUST remain retrievable. Retraction of evidence is a new assertion, never a deletion.
6. Conformance levels
L1 — Recorded.
Sections 5.1, 5.2, 5.4, 5.10. The decision, its evidence, and its immutability. A reader can see what was decided and on what.
L2 — Reconstructable.
L1 plus 5.3 and 5.5. An independent party can rebuild the knowledge state at decision time and re-derive the conclusion.
L3 — Contestable.
L2 plus 5.6, 5.7, 5.8, 5.9. A reviewer can argue with it — identify what was missed, what would change it, and who is accountable.
L2 is the minimum for any decision subject to regulatory review. L3 is the minimum for any decision that materially affects a person.
7. The reconstruction test
The single criterion by which conformance is judged:
Given only the decision record and the evidence store, an independent party who was not present MUST be able to rebuild the knowledge state as ofdecision_timeand re-derive the recorded conclusion — without access to any evidence recorded aftertransaction_horizon.
If later-arriving evidence is required to explain the decision, the system is not conformant. If the reconstruction produces a different conclusion, either the record is incomplete or the method is non-deterministic; both MUST be disclosed.
8. Anti-patterns
- The bare score. A record consisting of a number and a disposition. Not reconstructable, not contestable, not conformant at any level.
- The single clock. One timestamp doing the work of three. Late-arriving evidence makes past decisions appear negligent.
- Deletion on retraction. Removing discredited evidence, which silently falsifies every record that relied on it.
- The moving model. A
methodfield naming a model without a version, where the deployed model has since changed. - Independence by assumption. Multiple signals from one upstream source counted as separate corroboration.
- Retrospective rationale. A justification composed after the outcome was known.
9. Open questions
Comment is invited on the following.
- Whether
rationalerequires a structured form, or whether natural language is sufficient given that reviewers are human. - Whether L3 should require a stated confidence, or whether requiring calibration when present is enough.
- How to represent decisions where the decider is a chain of automated agents, each conditioning on the last.
- Whether
would_change_ifcan be validated automatically, or is inherently a matter of judgment. - Whether non-determinism should bar L2 conformance or be disclosable within it.
DDR-1 is a working document and will change. It specifies requirements, not implementations, and no conformant implementation is described here. Comment to phil@philpierce.com.