Part V, The Decision Machinery

Derived from Axioms 5 (transparency), 6 (evidence over assertion), 10 (self-correction). This is the engine model5 shows as Evidence → Decision → Review·Pause·Correct, specified in full, including the means/ends frontier, inter-domain conflict, and the governed role of AI.

V.1 The decision loop

Every significant governing decision follows one published loop. The loop is the accountability: each stage produces a public artefact.

MANDATE          What the people have asked for (Part III) within the Charter (Part I)
   ↓
EVIDENCE &       Data, models, scenario testing, risk assessment, with stated
MODELLING        uncertainty and adversarial (red-team) review
   ↓
OPTIONS &        Real alternatives, each with costs, benefits, winners, losers,
TRADE-OFFS       and the assumptions behind them
   ↓
VALUE CHECK      Is this really a technical choice, or a hidden value choice?
   ↓             (If value → escalate to democratic decision, §V.4)
DECISION         A named, accountable choice with published reasoning,
                 assumptions, and trade-offs (§V.3)
   ↓
IMPLEMENTATION   With predicted outcomes recorded in advance
   ↓
MEASURE          Actual outcomes vs prediction, on the outcomes ledger (Part VI)
   ↓
REVIEW·PAUSE·    Error-correction: confirm, adjust, pause, or reverse (§V.7)
CORRECT          → feeds back to MANDATE / EVIDENCE

V.2 Evidence and modelling standards

Because decisions claim to be evidence-based (Axiom 6), the evidence itself is held to a standard:

• Open data and open models. The data, model, and code behind a major decision are public and, where possible, reproducible. "Trust our model" is not acceptable; "re-run our model" is.
• Stated uncertainty. Point estimates without confidence ranges are forbidden for material decisions. Decisions must say how confident they are and what would change the answer.
• Adversarial / red-team review. Major decisions undergo structured challenge by an independent team paid to find the flaw, institutionalised dissent against groupthink (Part IV).
• Falsifiability (Axiom 6), tiered honestly. Where a decision rests on an empirical prediction, it states what observation would prove it wrong and commits to measuring it. Many governing decisions are not falsifiable in that sense (long-horizon, counterfactual-dependent, or value-laden); for these the standard is explicit success-criteria plus a defined review trigger, not a forced pseudo-test — and independent review checks that stated tests are non-trivial, so falsifiability drives real evaluation rather than a ritual metric set to be trivially passed (links §V.6, Part VI, §XIX.2).

V.3 Stated assumptions and trade-offs, radical transparency of reasoning

This is model5's "stated assumptions and trade-offs", made into a hard rule. Every significant decision publishes, in plain English:

1. The recommendation, in one sentence.
2. The assumptions it rests on.
3. The trade-offs, what is being sacrificed for what, and who benefits and who bears the cost (named groups, not abstractions).
4. The alternatives considered and why they were rejected.
5. The falsification test, what result would mean this was the wrong call.
6. The named decision-owner, the person accountable (no anonymous power, Axiom 4; "the algorithm decided" is never a defence, §V.6).

A decision published without these is procedurally void.

V.4 The means/ends frontier: detecting hidden value choices

The integrity of §0.5 depends on correctly sorting technical questions (experts decide) from value questions (the people decide). The danger is value choices disguised as technical ones — the discount rate applied to future generations, the acceptable level of risk, the distribution of who-wins-who-loses. These look technical and are actually values. This is the single most important anti-technocracy mechanism in the model, so it is specified to the point of being operable, not left to a single word of judgement.

V.4.1 The value-flagging test, and what "material" means

A decision is flagged as a value choice, and routed out of expert hands, if it materially does any of:

• changes the distribution of benefits/costs between identifiable groups;
• trades present against future citizens (intergenerational, Axiom 13);
• sets a risk tolerance the public would reasonably contest;
• touches a rights or moral question.

Because almost every decision touches distribution or risk to some degree, "material" is defined so the test neither escalates everything (paralysis) nor nothing (capture). A decision is material if any one of these holds:

• Scale — the distributional shift, cost, or risk exceeds a published threshold (an absolute figure and a share of the relevant budget or affected population), set per domain and versioned in the repo;
• Identifiability & concentration — the winners or losers are identifiable groups and the effect on them is concentrated rather than diffuse;
• Irreversibility — the decision is hard to undo (§V.7), which lowers the threshold;
• Reasonable contestation — reasonable citizens with the same facts could rationally disagree about the choice (the decisive test for the risk and moral limbs).

Worked examples are published with the test: setting the level of a pollution limit is a value choice; choosing the cheapest compliant engineering to hit a legislated limit is technical; a clinical-priority rule is a value choice; the logistics of delivering it are technical. The line is a spectrum, and the design says so: the test manages the ambiguity with a published, challengeable threshold rather than pretending a crisp boundary exists.

V.4.2 Where a flagged value question is decided (one deterministic rule)

A flagged question has exactly one destination, fixed by its stakes, so it cannot be steered to a forum the proponent can win:

• Tier V-a, ordinary value trade-offs are decided by the Citizens' Sortition Chamber (§IX.2). Its ruling binds the experts' execution, but is itself suspensory before the elected Assembly and the people: the Assembly may substitute its own value judgement only by a transparent supermajority, and either may be put to the people by referendum on petition (§XIX). The Chamber decides in the first instance; it does not hold the last word over the electorate.
• Tier V-b, constitutional-magnitude or rights-touching value questions (anything reaching a Class A right, the franchise, the Union, or the core, §I.9.2) are decided by the people directly, by binding referendum, after Sortition-Chamber and Assembly deliberation.

This resolves the ambiguity by which §V.4, §IX.2, §XVIII D3 and §XIX previously appeared to name different final deciders: the Sortition Chamber is the standing decider of ordinary value trade-offs (suspensory to the people), and genuine constitutional or rights questions go to the electorate.

V.4.3 Who decides classification (self-executing, not court-gated)

Classification can itself be gamed — a body may insist a value-laden choice is "merely technical" to keep it. So:

• The default is pro-democratic: where classification is genuinely contested, the presumption is escalation — the question goes to the people or the Sortition Chamber, not to experts.
• Escalation is self-executing on a citizen petition above threshold, or on a request by the Sortition Chamber, the Integrity bodies, or its materiality-affirmation panel (§XIX.3): it does not require a court's permission. The Rights & Constitutional Court rules only on rights-legality (whether an escalated decision would breach the Charter), never on whether escalation was warranted — so the boundary of democratic authority is not drawn by the least democratic institution.

V.4.4 Anti-abuse, in both directions

The classifier is a weapon that can be misused to smuggle values in (capture) and to force competent execution out (obstruction). Both are defended:

• Against smuggling: the pro-democratic default and self-executing escalation above.
• Against obstruction: a challenge to a classification, and a §V.7 Pause triggered by petition, are subject to a vexatious-challenge filter; a raised threshold and cooling-off for repeat challenges to the same decision; a rule that a resolved escalation cannot be re-triggered without materially new grounds; and time and cost bounds on escalated deliberation, so a motivated actor (a well-funded interest, a populist campaign, a foreign influence operation) cannot mire a sound technical decision — a vaccine schedule, a grid upgrade, a pollution limit — in endless re-litigation. The rate of escalation-abuse is tracked as a backsliding indicator (§VI.7).

V.5 Inter-domain conflict resolution

Domains will clash, Economy vs Environment, Security vs Liberty, Health vs Economy. The model resolves this explicitly rather than leaving it to whoever shouts loudest:

1. Against the mandate. Conflicts are resolved by reference to the people's published priorities (the mandate, Part III) and the wellbeing objective (§0.2), not by inter-departmental power.
2. Cross-domain arbitration. A standing arbitration process weighs the trade-off transparently, applying §V.3 (publish who-wins-who-loses).
3. Value conflicts escalate. If the clash is fundamentally a value trade-off (it usually is), it triggers §V.4 escalation to democratic decision, the experts do not get to pick the winner between competing values.
4. Rights are not balanced away. Where a conflict touches Class A rights, the Charter (Part I) governs absolutely; there is no trade-off to arbitrate.

V.6 The role of AI and computation, augment, never rule

"Given all technological advances" is a core premise, but technology is a tool of governance, never a governor:

• AI is decision-support, not decision-maker. It strengthens evidence, modelling, scenario-testing, anomaly detection, and translation of complexity into plain English. It never holds authority over a citizen.
• A human is always accountable. Every decision has a named human owner (§V.3.6). "The algorithm decided" is never a defence and never a shield from accountability.
• Governed AI (Part VIII). Any AI used in governance must be registered, auditable, explainable, bias-tested, and reproducible. Secret or unexplainable algorithms are forbidden in government (Axiom 5, 7).
• Automation-bias guard. Decision processes are designed so humans genuinely evaluate AI outputs rather than rubber-stamp them; dissent and override must be easy and recorded.
• No autonomous coercive power. No automated system may, on its own authority, deprive a citizen of liberty, money, or rights. A human decides; the system advises.
• Cleared before deployment. No AI enters a governing function until a published algorithmic impact assessment, purpose, data, tested error and bias, affected rights, and fallback, has passed independent review, with the highest-impact uses held to the strictest tier (technical detail, §VIII.5).
• Never the decider in rights-critical functions. AI may inform but never decide where a citizen's liberty, the use of force, or eligibility for an entitlement or legal status is at stake; there a named human decides on the evidence (§XII.2, §XII.8).

V.7 Review · Pause · Correct, the error-correction engine

The mechanism that makes the system self-correcting (Axiom 10) rather than ossified:

• Triggers: measured outcomes diverge from prediction; new evidence; detected harm; a failed falsification test (§V.2); a citizen petition above threshold; an Integrity Assembly or Court referral.
• The Pause power is the two-tier instrument of §VI.1, held by the Court: an immediate emergency interim freeze (triggerable by a single judge or a senior Integrity Enforcement officer, auto-referred to the Court within hours) for fast irreversible harm, and a time-boxed, per-decision-capped policy Pause for contested-but-not-urgent decisions, appealable on proportionality. The Enforcement Office requests; the Court affirms. It is never an unbounded standing veto by a single body.
• Mandatory post-implementation review: every major decision is reviewed against its predicted outcomes on a published schedule. No "decide and forget."
• The error register: mistakes are recorded, openly, and feed learning, because a system that hides its errors cannot correct them (and hiding them is itself a failure mode, §0.6).
• Reversibility by design: decisions are structured, where possible, to be reversible or piloted first, so error is recoverable.

V.8 Speed versus deliberation

The system must be able to move fast without being reckless:

• Standard track: full loop with deliberation, the default.
• Fast track: for time-critical decisions, a compressed loop with a named owner, recorded reasoning, and heightened, mandatory post-hoc review, speed is bought with stronger after-the-fact accountability, never with less transparency.
• Crisis: genuine emergencies are governed separately and bounded hard by Part VII and the Charter's crisis doctrine (§I.6).

V.9 Failure modes and safeguards

Part V ends. Next: Part VI, Integrity, Safeguards & Self-Review: the system's immune system, and the answer to "who guards the guardians".