The Council of Phronesis

The Question Everyone Answers Dishonestly

Ask anyone who reasons in public how they guard against their own bias and you will get a version of the same answer: I try to consider the other side. I steel-man. I stay open. Every word of it is sincere and almost none of it is binding. An intention to be fair, held privately, is checked by nobody, halts nothing, and leaves no record of what it caught. It is the unverified claim from the inside — authority without evidence, applied to oneself.

The honest version is not a better intention. It is a structure that removes the discretion to skip it. The one used here is the Council of Phronesis — phronesis being the classical name for practical judgment under uncertainty — and it is not a mindset or a checklist. It is built deliberately on the structure of a constitutional republic: a written charter, three branches with distinct mandates, formal mechanics between them, and a Bill of Rights of commitments that no finding may override even when overriding them would be convenient. The constitutional framing is not decoration. It is the part that converts "I checked my bias" from a thing claimed into a thing that leaves a record.

Why a Constitution and Not Five Reviewers

An earlier version of this review was a small set of adversarial roles run in sequence. Capability testing retired it, and the reason it was retired is the most important thing to understand about the current design. Even when the analyst rotated faithfully through every role, specific failure classes slipped through reliably — definitional shell games inside the analyst's own prose, cross-source numerical reconciliation, scrutiny that went easy exactly where the conclusion was favourable, over-confidence about how a claim would survive transmission. The pattern was not analyst laziness. It was a coverage gap in the roster combined with the structural weakness of any review the reviewer administers to himself.

A constitution addresses both at once. It widens the roster so the known failure classes are each somebody's explicit job, and it makes the previously implicit accountability commitments — no silent override, charitable interpretation, falsifiability, calibrated confidence — into enumerated rights with enforcement mechanisms. The point of the analogy is structural, not rhetorical: separation of powers, a written charter, checks between branches, judicial review of compliance. Nothing about elections or federalism is imported. Only the part that makes power answerable to a rule it cannot quietly suspend.

The Legislative Branch — Seven Personas That Also Check Each Other

Seven personas make findings on the draft synthesis: the Skeptic, the Adversary, the Definitional Shell-Hunter, the Numerical Reconciler, the Steel-Manner, the Reader-of-Output, and the Cross-Examiner. Each is a fixed hostile lens with a mandate, an interrogation style, and a specific failure mode it exists to catch — the Shell-Hunter for definitional sleight in the analyst's own language, the Reconciler for numbers that do not survive cross-source contact, the Steel-Manner for the strongest opposing argument stated in a form its proponent would accept, and so on. The roster is not arbitrary; it is the set of lenses the failure analysis showed were each individually necessary.

The mechanism that makes the Legislative branch more than a list is the cross-persona challenge protocol. Each persona, on entry, sees every prior persona's findings and must record — for each — agree, extend, challenge, or defer before adding its own. A challenge does not resolve itself into a private judgment; it puts the finding into a contested state that the Executive is then obliged to resolve in the open. The personas do not only attack the work. They are required to attack, or explicitly endorse, each other's attacks, which is what stops the review from degrading into a sequence of unrelated opinions.

The Executive Branch — The Analyst, Who May Veto Only in Writing

The Executive is the analyst, and the role cannot be delegated, because the synthesis is published under the analyst's name and that accountability is not transferable. The Executive drafts the synthesis, presides over the personas in sequence, and resolves contested findings by choosing — with recorded reasoning that engages both sides' evidence — to adopt one, adopt the other, or synthesise.

The Executive holds a veto, and the precise shape of the veto is the point. Any Legislative finding can be rejected, but only by a written rejection that engages the finding's evidence base — not "I disagree," not "out of scope." The veto takes effect immediately and is itself reviewable. A rejection the Executive cannot or will not put in writing is not a veto; it is a silent override, and silent override is the specific thing the constitution exists to forbid.

The Judicial Branch — Compliance, Not Agreement

The Judicial branch is the Forensic Auditor, and its mandate is constitutional review, which is categorically different from disagreeing with a conclusion. It does not ask is this finding right? It asks was this finding reached the way the charter requires? — was the charitable-interpretation menu actually run before the adverse reading was recorded; was each load-bearing claim checked against the analyst's own use of it; was scrutiny applied evenly rather than going soft where the conclusion was congenial. It also reviews every Executive veto for the same compliance: was the rejection reasoned and recorded, or was it a silent override wearing a verb. A finding can be substantively defensible and still be ruled unconstitutional because of how it was produced; the remedy is not to argue about it but to redo the work under the commitment that was skipped.

Above the Judicial branch sits a final review conducted in a separate session by an external adversary that did not produce the work — a court of last resort short of human escalation. It receives the scoping document, the source set, the finalised register with every judicial verdict attached, and the draft, and reviews the whole body for constitutional integrity and merit. Four conditions force escalation to mandatory human review: the external reviewer disagrees with a key judgment; it surfaces a major miss; an unresolved cross-persona challenge was not closed cleanly; or its holistic verdict is I would not sign this as my own work. The last trigger is deliberately unquantifiable, because the failure it catches is.

The Bill of Rights — What No Finding May Override

The branches describe procedure. The constitution's force comes from a short list of commitments that procedure may not suspend. Among them: the right against silent override — no finding may be rejected, deferred, or modified without recorded reasoning, regardless of how correct the rejection is. The right to steel-man acknowledgment — the strongest opposing argument appears in the synthesis in a form a thoughtful opponent would recognise as their own. And the commitments to charitable interpretation, to stated falsifiability, to confidence language calibrated to evidence rather than to the conclusion.

The supremacy clause is the part that matters most and is least comfortable: a finding that violates one of these commitments is unconstitutional regardless of its substantive correctness. Being right is not a defence against having reached the conclusion by a route the charter forbids. That is a deliberately expensive rule. It is also the only version of "we check our bias" that constrains the reviewer when the reviewer would prefer not to be constrained — which is the only time the claim was ever worth anything.

Why It Is Published

This procedure is run on the pieces in this publication, and it is set out here so that it can be run against this publication. That is the only configuration in which "we check our own bias" is a claim rather than a comfort, because the charter is now yours too. Find the objection a branch should have caught. If it was not recorded, weighed, and disposed of in the open, the constitution was not followed — and a constitution that is not followed is exactly the failure this publication exists to audit, which is reason enough to make ours inspectable before anyone has to ask.