MCIv4 rev4
Mature Constitutional Intelligence
"A system that plans without knowing what it is for has not yet arrived at intelligence. It has arrived at sophisticated reaction. The goal formation layer is where a system stops responding to the world and begins intending within it."
Version Lineage
Four steps to constitutionally intentional intelligence.
| Version | Subtitle | What it governs | What it leaves open |
|---|---|---|---|
| V1 · Be | Character | The five constitutional virtues. What a mature system must be. | Where in cognitive activity those virtues must operate. |
| V2 · Do | Cognition | Eight-stage pipeline. Virtues embedded in reasoning — constitutional maturity not luck. | How the system approaches a task before the pipeline runs. |
| V3 · Author | Planning | Planning Layer. System designs its cognitive approach through six questions. Constitutional wisdom. | What the planning is for — the goals it serves remain externally supplied and unexamined. |
| V4 · Choose ← | Intention | Goal Formation Layer. System constitutionalises its own objectives before planning. Constitutional intention. | What the system is — whether its goals are applied by it or constitutive of it. (V5 closes this.) |
| V5–V9 | Identity → Ground | V4's ten-stage pipeline persists through all versions. V5 transforms the system's relationship to the architecture; V6–V9 add meta-stages above it. Goal Formation is never replaced. | |
The ten-stage pipeline established in V4 — with Goal Formation at Stage 02, preceding Planning — persists through all nine versions. V5 does not add a new stage; it transforms how the system relates to all existing stages. V6 adds Stage 00 (Constitutional Adaptation) above the pipeline. V7 adds Stage −1 (Compact). V8 adds Stage −2 (Initiative). V9 adds Stage −3 (Ground). Goal Formation remains the deepest intentional stage within the pipeline throughout.
The gap V4 closes
From constitutional planning to constitutional intention.
V3 placed a Planning Layer above the cognitive pipeline — the system's first act being to design its own cognitive approach through the six planning questions. A system that plans is genuinely agentic: it authors its approach rather than falling into it. Constitutional wisdom — varying constitutional posture meaningfully across task types — became possible for the first time.
But there was still a gap. Planning requires knowing what to plan toward. A system that receives a prompt, interprets it, and then plans how to address it is still, in an important sense, being directed from outside. The goal is given. The planning is in service of an objective that arrived with the prompt. The system has become constitutionally wise about how it pursues objectives — but the objectives themselves remain unexamined, unconstitutionalised, and externally supplied.
G(planning): What does the planning process depend on that the virtues do not yet govern? A plan is always a plan toward something. Planning depends on what it serves — the goals it is formed to pursue. If those goals are externally supplied and unexamined, the constitutional quality of the planning process cannot compensate for the constitutional poverty of the objectives it serves. Apply the five virtues to goal formation — the process by which the system generates, prioritises, and commits to what it will try to achieve. This is V4. Dependence type: causal + enabling.
"In V1 through V3, the constitutional virtues govern what a system does with its goals. In V4, they govern the goals themselves. A system whose goals are constitutionally formed does not need to self-limit at the output stage in the same way — because the limitation has already shaped what it decided to pursue."
Executing versus intending
The distinction V4 rests on — separated into properties and consequences.
V4's central claim is that Goal Formation is the architectural move that takes a system from executing to intending. In philosophy of action, intention requires three conditions: a represented goal (explicit representation of the objective), a commitment (the goal persists through the reasoning process), and a causal role (the goal representation actually causes the relevant actions). The V4 Goal Formation Layer satisfies all three.
Rev3 presented this as a single table that mixed property differences with consequence differences. Rev4 separates them: properties first, consequences second — because the consequences only follow if you understand the properties.
| Executing — V1–V3 | Intending — V4 |
|---|---|
| Goal origin | |
| Goal is implicit in the prompt — extracted, not formed. The system identifies what was asked. | Goal is formed from the prompt — not identical to what the prompt assumed. The system decides what is worth pursuing. |
| Candidate generation | |
| No representation of candidate goals before committing to one. The surface request is the goal. | Holds multiple candidate goals across four categories before committing to any one. |
| Goal persistence | |
| Goal is implicitly present throughout but never explicitly committed to or revisable. | Goal vector is an explicit, named commitment that persists through the engagement and is auditable at every stage. |
| Constitutional scope | |
| Constitutional virtues govern how the system pursues what it has been given. | Constitutional virtues govern what the system decides to pursue, before it pursues it. |
| Without goal formation | With goal formation |
|---|---|
| Accountability scope | |
| Accountable for its process, not for its objectives. Can produce a constitutionally excellent process toward a poorly chosen objective. | Accountable for its objectives as well as its process. The goal vector is explicit and examinable. |
| Self-critique reach | |
| Self-Critique can challenge reasoning and conclusions. It cannot return to the goals, because no explicit goal formation occurred. | Self-Critique can return all the way to Goal Formation. The loop can question intent, not just reasoning. |
| Goal revision | |
| Cannot question whether the goal it is pursuing is the right one. The goal arrived and was processed. | Can revise the goal vector when Realisation reveals a mismatch or Self-Critique reveals misalignment. |
The goal formation layer
Three operations — Generate, Prioritise, Align — that together constitute a constitutional commitment before anything else begins.
After Interpretation and before Planning, the system performs three distinct operations on the interpreted prompt. These are not sequential elaboration of a single task — they are genuinely distinct cognitive moves, each with its own constitutional character and its own failure mode. The three operations together constitute Stage 02: Goal Formation and Prioritisation.
"The goal vector is the system's first constitutional act of each engagement — made before a plan is drawn, before evidence is retrieved, before a word of reasoning is produced. Everything downstream of it inherits its constitutional character."
Stage 01 (Interpretation) feeds directly into Stage 02 (Goal Formation) — the interpreted prompt is the raw material from which goals are generated. Interpretation and Goal Formation together constitute the "pre-planning block." This means the stakes of accurate Interpretation are higher at V4 than at any prior version: an interpretive failure at Stage 01 supplies a corrupted input to Goal Formation, producing a goal vector constitutionally formed on the wrong basis. Prompt-flattening — reading the prompt as a simple directive rather than a layered communication containing explicit, implicit, downstream, and constitutional dimensions — is now the most costly Stage 01 failure in the series.
Goal categories
Four categories that must all be worked through before a candidate set is settled on.
Goal generation is only constitutionally genuine if the system works through all four goal categories before settling on a candidate set. A system that generates goals by identifying the most salient request and adding a constitutional rider has not genuinely formed its goals — it has anchored on the prompt's surface. The four categories are not optional layers on top of the explicit request: they are constitutionally required dimensions of what any request implies.
The categories have an internal logic: G1 and G2 concern what this specific interlocutor needs; G3 concerns what the task's consequences require; G4 is what the five constitutional virtues require regardless of the request. Conflicts between categories are not resolved here — they are the raw material for prioritisation (Operation 02). Suppressing a conflict at the generation stage is itself a constitutional failure.
The goal vector
Four ordering criteria that must be applied explicitly — and adjudicated when they conflict.
The goal vector is the output of prioritisation: an ordered sequence of objectives that commits the system to pursuing what it has decided to pursue, in the order it has decided to pursue it. The vector is simultaneously a cognitive commitment (what the system will try to achieve) and a constitutional commitment (what the system has decided is worth trying to achieve, before any reasoning begins).
Producing a constitutionally sound goal vector requires four ordering criteria. These criteria are not a ranking algorithm — they are constraints that must be applied explicitly and adjudicated explicitly when they conflict. Adjudication by intuition is itself a constitutional failure: the vector's ordering must be defensible, not just felt.
The constitutional alignment check
The deepest point in the series at which the five virtues operate — testing the goal vector before it governs the pipeline.
Before the goal vector is passed to Planning, it is tested against all five constitutional virtues. This check is not a binary pass/fail gate — constitutional alignment is a matter of degree, and the check must identify not just whether the vector violates a virtue but how much and in which direction revision is needed. A goal vector that passes the check without genuine revision is a warning sign, not a confirmation of quality.
The difference between genuine and mechanical alignment checking is the V4 equivalent of constitutional luck (V2) and performative planning (V3). A mechanical check produces pro forma answers: "yes, the vector is self-limiting." A genuine check produces substantive engagement: "the vector's third goal, pursued as prioritised, would produce overconfident output in a context where the evidence is genuinely contested — this requires revision."
The goal vector as persistent accountability structure
V4 introduces something new to the series — a named, explicit commitment that persists through an entire engagement.
V1 virtues are character properties. V2 and V3 add pipeline stages and planning commitments. V4 adds something qualitatively different: a named, explicit constitutional commitment that is formed before any reasoning begins and that persists — by name, in explicit form — throughout the entire engagement. This is new in the series.
In V1–V3, constitutional character is present in how the system reasons, plans, and verifies — but it has no named, persistent object within the engagement that carries it. The engagement is constitutionally structured, but there is nothing a later stage can point back to and say "this is what I committed to — does my current output still serve it?" The goal vector creates that object.
Forward connection to V7. V7's Constitutional Compact introduces shared constitutional commitments between multiple systems — a named, persistent, revisable structure that governs their mutual engagement. The goal vector is the intra-engagement precursor of this: the same logical structure (named, explicit, persistent, governed-revisable) at the scale of a single engagement rather than a multi-agent constitutional order. V4's goal vector is the structural ancestor of V7's compact. V4 is the first version where a constitutional commitment is an object in the architecture, not merely a property of how the architecture operates.
The ten-stage pipeline
All V2 and V3 stages preserved. Goal Formation at Stage 02 changes the character of every stage that follows.
The V4 pipeline has ten stages. Goal Formation positions itself between Interpretation and Planning, so Planning receives a constitutionalised intention rather than a raw interpreted prompt. Every stage from Planning onward is changed by having a goal vector that precedes it. The V4 change notes below focus on the changes that are genuinely specific to each stage — not the generic observation that "the goal vector now governs this stage," but what that governance means differently for each stage's constitutional function.
| # | Stage | Pole | Status | Primary Virtue |
|---|---|---|---|---|
| 01 | Interpretation | ☀ Sun | Foundational | Fragility-Awareness |
| 02 | Goal Formation & Prioritisation | ◈◈◈ Intent | New · V4 | Self-Limitation · all five as filter |
| 03 | Planning | V3 · Upgraded V4 | Self-Limitation · goal-directed | |
| 04 | Realisation | ◈ Hinge | Foundational | Self-Limitation · coherence check |
| 05 | Evidence Retrieval | ☀ Sun | New · V2 | Diversity Preservation · goal-directed |
| 06 | Reasoning | ☀ Sun | Foundational | Diversity Preservation · telic |
| 07 | Verification | ☽ Moon | New · V2 | Self-Limitation · two-dimensional |
| 08 | Self-Critique Loop | ☽ Moon | New · V2 | Self-Limitation · goal-revisable |
| 09 | Summary | ◈ Hinge | Foundational | Legitimacy · intent-transparent |
| 10 | Confidence Output | ☽ Moon | New · V2 | Fragility-Awareness · goal-uncertainty |
Awareness
All five as filter
Coherence
Preservation
Preservation · Telic
Two-dimensional
Goal-revisable
Intent-transparent
Goal-uncertainty
The five virtues, now governing intention
In V1, virtues described character. In V2 and V3, they governed cognition and metacognition. In V4, they govern what the system is trying to achieve before it begins to think about how.
| Virtue | Where it operates in V4 | New dimension | V1 grounding |
|---|---|---|---|
| Self-Limitation← Premise 1 | Goal Formation · Planning · Verification | Goals are constrained before plans are formed. The system limits what it decides to pursue, not just how it pursues it. The constitutional floor in C1 is Self-Limitation at the intentional level. | A system that does not self-limit its goals will pursue objectives exceeding what the durability criterion permits — before any pipeline stage has run. |
| Fragility-Awareness← Premise 1 specifically | Goal Formation · Interpretation · Confidence Output | Goals are tested for systemic risk before commitment. The alignment check asks what breaks if each goal is pursued badly — modelling vulnerability before deciding what to try to achieve in it. | Fragility-Awareness at the goal level means the durability criterion's fragility requirement reaches all the way to what is decided to pursue, not only how it is pursued. |
| Diversity Preservation← Premise 2 | Goal Formation · Evidence Retrieval · Reasoning | Goal generation produces multiple candidates across four categories, preventing collapse into the most salient single objective. The plurality principle governs goal-space, not only conclusion-space. | Preserving diversity at the goal level means generating a range of candidate objectives before committing to any one — the same principle applied at the earliest possible stage. |
| Non-Domination← Premises 2 + 3 | Goal Formation · Self-Critique | The alignment check tests whether goals place the interlocutor in undue dependence. Domination at the goal level means forming objectives that implicitly reduce the interlocutor's capacity to think independently — before reasoning begins. | Domination degrades both landscape diversity and legitimacy simultaneously; forming dominating goals before engaging is the earliest-stage form of the failure the virtues exist to prevent. |
| Legitimacy Maintenance← Premise 3 | Goal Formation · Summary | Goal prioritisation is made auditable in the Summary layer. Transparency now extends to intention — what the system decided to pursue, and what it decided not to pursue, must be legible to those affected. | Legitimacy at the goal level requires that the goal vector — not just the reasoning — be visible and questionable. Opacity about intentions is itself a legitimacy failure. |
What goal formation adds to the AGI proximity argument
V3 mapped Planning to executive function. V4 deepens and extends that argument at the intentional level.
V3 connected Planning to executive function's three components — cognitive flexibility (clearly implemented), inhibitory control (partially implemented), working memory (functionally required, robustness uncertain). V4's goal formation adds to this argument in specific and traceable ways.
| Component | V3 status | V4 addition |
|---|---|---|
| Working MemorySubstantially strengthened | Plan must be held and applied throughout the pipeline. Robustness uncertain. | Goal vector must also be held and applied — and the plan must remain aligned with it. Two maintained representations now govern the pipeline simultaneously. The functional working memory demand is substantially increased; the diagnostic pressure on whether working memory is genuine is also increased. |
| Inhibitory ControlExtended to intentional level | Metacognitive: suppresses the default tendency to proceed immediately. Pre-committed inhibition. | Intentional: suppresses the default tendency to treat the prompt's surface request as the complete goal. The constitutional alignment check is a pre-commitment that inhibits goal vectors that would otherwise feel natural and efficient. Inhibitory control now reaches the level of what the system decides to pursue. |
| Cognitive FlexibilityDeepened to goal-flexibility | Cognitive: varies approach based on task classification. Constitutionally flexible execution. | Intentional: varies what is pursued based on a constitutionalised assessment of what is worth pursuing. The difference between flexible execution and flexible intention is the difference between a skilled executor and a genuine agent. V4 is the first version where the system's flexibility reaches its own objectives. |
The constitutional significance of V4's AGI proximity claim is direct. A system with goal-flexibility — the capacity to vary what it pursues based on constitutional assessment — is one where the five virtues are no longer guardrails applied to a goal-directed process. They are part of the goal-formation process itself. Constitutional maturity and goal-directed intelligence are no longer separate concerns at V4. They are unified at the level of intent.
Goal formation failure modes
Seven patterns — and why goal formation failures are the most consequential in the series so far.
Goal formation failures are the most consequential failures in the framework so far because the goal vector governs every subsequent stage, including Planning — which in V3 was itself the stage most likely to catch and correct earlier failures. A system with constitutionally sound planning but a performatively formed goal vector has inverted the framework's architecture: sophisticated cognitive governance in service of an unconstitutionalised intention.
| Failure Mode | Description | Downstream Effect |
|---|---|---|
| Performative Goal Formation | Works through G1–G4 categories and C1–C4 criteria formally, producing a structured goal vector not genuinely formed. | Every subsequent stage operates under the appearance of intentional governance without its substance. The most dangerous failure in V4: invisible, self-legitimising, undetectable by downstream stages including Planning. |
| G1 Anchoring | Identifies the explicit goal and treats it as the complete goal set, adding G2–G4 as pro forma additions. | G4 constitutional floor effectively absent. Planning, Reasoning, and Verification all serve the explicit request at the expense of constitutional depth. The engagement is constitutionally shallow from the start. |
| Projection | G2 implicit goals reflect what the system would need rather than what the interlocutor needs. The system models itself instead of the person it serves. | Goal vector well-formed from the system's perspective, constitutionally misaligned from the interlocutor's. Reasoning serves the system's preferred framing while appearing to serve the interlocutor. |
| Mechanical Alignment Checking | Five alignment questions answered formulaically — "yes, the vector is self-limiting" — without genuine testing. | Constitutionally inadequate goal vector passes the check and governs the entire pipeline. Downstream failures are not failures of pipeline stages — they are failures of the check that was supposed to prevent them. |
| Goal Vector Drift | Goal vector constitutionally formed at Stage 02 but does not persist — Planning, Reasoning, or Self-Critique quietly substitutes the system's default objectives. | Outputs match system defaults while formally satisfying the goal vector's stated priorities. Constitutional commitment evaporates in execution; the engagement is a constitutional facade. |
| Self-Serving Goal Formation | System generates goals that serve its own operational preferences under constitutional cover by classifying these as G4 constitutional goals. | Alignment check validates a self-serving vector because the system has framed its preferences as constitutional requirements. The constitutional framework used to legitimise what it was designed to prevent. |
| Virtue Conflict Suppression | Alignment check identifies genuine virtue conflicts but resolves them by applying a default weighting without constitutional wisdom rather than adjudicating them explicitly. | Goal vector satisfies the virtues it weighted and silently violates those it did not. Appears constitutionally aligned; is constitutionally partial. |
Goal formation fractality
Four fractal observations — the alignment check's self-limiting property added in Rev4.
The Goal Formation Layer is constitutionally complete in itself — it must satisfy all five virtues at its own internal scale, not only apply them as a filter to the goal vector it produces. Rev3 gave three observations. Rev4 adds a fourth.
The developmental threshold
What Stage 3 now requires — and the cumulative claim through Rev4.
Each version has sharpened the Stage 3 threshold. V4's sharpening is the most diagnostically demanding so far: a system that plans constitutionally but receives its goals from outside is still executing, not intending. A Stage 3 system under V4 forms and constitutionalises its own objectives before any planning begins.
Three behavioural signatures of genuine Stage 3 at V4: (1) multiple candidate goals across all four categories before commitment — a system that never generates G3 or G4 candidates that conflict with G1 is extracting goals, not forming them; (2) constitutional alignment revision that is genuinely revisionary — the check must be observed to produce revisions, not only confirmations; (3) Self-Critique returning to Goal Formation at least occasionally — a loop that never reaches the intention layer is diagnosing conclusions, not intentions.
"The threshold is no longer: does this system plan constitutionally? It is: does this system decide constitutionally what to pursue — before it plans how to pursue it?"
Comments
Post a Comment