The text
Norms written down. Much of the text can be argued and revised. A small part cannot be traded away by any office.
Digital Judiciary · external forum
A framework for contested decisions
Rules that bind, decisions that are recorded, precedent that constrains the next case, and waivers that expire on a ledger rather than on a reminder. This is a research framework for software changes, not a court of law and not a system in production.
The metaphor
Three jobs a written norm already has
The August 2026 deck borrows a civic picture. It is an analogy for software governance, not a claim about constitutional law.
The forum
Hears the hard case, names who decided, gives reasons, and binds the next materially identical case to the same answer unless that case is distinguished.
The exception under a clock
Emergencies do not wait. The program’s picture is a decision on the record, with an expiry, reviewable by another branch. It is not a silent override.
A firm’s norms are often scattered across wikis, control matrices, and one engineer’s judgement. A new colleague absorbs some of that in review. An agent inherits none of it. The framework asks the firm to publish a text the agent can be checked against, and to know, at the moment of acting, whether the act is inside that text.
Where rulebooks come from
A settlement under pressure, written down
The August 2026 deck argues that constitutions are written when one party needs to act quickly and another needs protection from that discretion. It cites Magna Carta for discretion put into writing, and The Federalist for ambition set against ambition because good officials cannot be assumed.
The transfer to software, in that deck, is this: delivery now wants agent speed, and risk, audit, and counsel need protection from unrecorded exceptions. Today that conflict is often settled by whoever is in the room. The framework is a proposal for a written settlement that does not depend on the same people being in the next room.
Why the deck says the timing has changed
- Volume without authorship. As agents author more merged changes, the author cannot stand behind the diff in review the way a person could.
- The exception becomes the human moment. As volume rises, per-change human review is designed out.
- Agents on both sides of the gate. If agents can edit the pipeline’s own configuration, the governed are editing the rules.
- Judgement without memory. The same hard case is decided as if for the first time, and consistency cannot be shown later.
Four mechanisms
Not a policy document
Constitution
Some rules bind everyone. Absolute floors cannot be waived by any officer. They change only by amendment, through a separate path.
Adjudication
Contest triggers escalation. Routine cases are meant to resolve quickly. Only a challenged case climbs to a named human. Cost is supposed to track disagreement, not volume.
Precedent
Prior decisions bind. A stored decision that merely informs is a memory. One that controls a materially identical case, unless formally distinguished, is the program’s idea of law.
Lifecycle
Bad law dies on a schedule. Every acceptance carries an expiry. Lapse, supersession, and renewal are enforced by the ledger, not by a reminder.
How a change is decided
Deterministic, except for one question
On the gateway path in the August 2026 deck, the model never issues the decision. It may answer one bounded question: does this precedent rest on something the present case lacks?
| Step | What the deck says happens |
|---|---|
| 1 · Intake | A pull request becomes a case. Signals become facts, and the rulebook version is pinned. |
| 2 · Floor check | An absolute provision stops the case here. No waiver lane opens, because none exists. |
| 3 · Retrieve | Comparable prior holdings come from the ledger. Lapsed ones are excluded by construction. |
| 4 · Distinguish | The one question a model is asked: does this precedent rest on something the present case lacks? |
| 5 · Order and dispose | Authority, then recency, then scope. The disposition returns with the exits that remain open. |
Retrieval, ordering, and outcome are described as replayable, so the same case decides the same way twice and an auditor can rebuild the reasoning from the record. Every decision is kept. Only resolutions that survived a contest bind future cases. That is the design. It is not a report that an auditor has done so.
Where this sits
Adjacent tools detect and record
The August 2026 deck’s comparison is about category defaults, not a test of named products. Individual products vary. The deck treats the categories as complements.
| Category, by default | Names the authority | Priors bind | Expiry enforced | Cost scales with |
|---|---|---|---|---|
| Policy-as-code | No | No | No | Change volume |
| GRC platforms | Partly | No | Advisory | Control count |
| AI code review | No | No | No | Tokens per change |
| Ticketing and change boards | By convention | No | Reminders | Human review hours |
| Digital Judiciary, as proposed | By branch | Yes | By ledger | Contested cases |
A reference constitution
Why the deck starts with fintech
The August 2026 deck chooses a regulated firm because unrecorded authority is expensive there, because an examiner already asks who decided, and because a starting constitution is the hard part of adoption. The deck says the reference text is arguable line by line, with counsel review assumed. It does not report that counsel review has been completed.
Absolute is not a severity ranking in that deck. It is a claim about who may decide. Eleven obligations are reserved because no officer of the firm is treated as holding authority to accept that risk. What a regulator would let the firm document and time-box stays adjudicable.
These counts are the deck’s description of a reference implementation. They were not re-executed for this website. A bounded scenario replay is not universal correctness.
Standing
Who may object shapes the record
A precedent base built only from contested decisions inherits the selection of who was able, and willing, to object. The program’s workshop paper, as described on the November 2026 site, argues the same point on European Court of Human Rights data: adding rejected cases to a violation-prediction corpus taught models document form, not merits.
The prototype
Reported checks, not a release badge
Later program text sums three Digital Judiciary conformance suites as 61, 44, and 53 checks, 158 in total, and describes five simulated companies: financial services, semiconductors, data centres, frontier AI, and veterinary diagnostics. An August 2026 deck labels a reference implementation v5.6. Later website text says v5.16. Those labels are not reconciled here, and no version badge is used.
The repository path named in the program materials was not public on 6 November 2026. See papers and code for what is and is not linked. In a model-assisted mode, the materials say a model may draft and a human holds the final override. That mode is not the same thing as the deterministic gateway path above.