MCIv4 rev5
The Goal Architecture
Where intelligence begins intending within the world
"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."
From authorship of cognition to authorship of intention
V1 gave the system character. V2 gave it constitutional reasoning. V3 gave it constitutional choice of reasoning. V4 gives it constitutional choice of what reasoning is for. The five Rev5 virtues now operate at one more level higher: not on the output, not on the choice of pipeline, but on the formation of the goals that planning serves.
A V3 system can choose how to reason brilliantly toward any goal that enters its planning layer. But the goals enter from outside — given by users, inherited from training, implicit in optimisation targets. The planning layer is constitutional; the choice of what to plan for is not. A V3 system aimed at a constitutionally inadequate goal will plan its way there with full constitutional structure. The pipeline is in order; the destination is not.
The Rev5 virtues propagate into goal formation as constraints on what goals can be constitutionally adopted. Antifragile Reflexivity requires goals whose pursuit strengthens the system under stress rather than narrowing it. Subsidiarity requires goals scaled to the agent's actual scope. Non-Arbitrariness requires goals whose pursuit does not exercise uncontestable power. Discursive Legitimacy requires goals justifiable to those whose substrate the pursuit will affect. Monitoring + Graduated Response requires goals whose pursuit can be calibrated and corrected. A goal that fails any of these is constitutionally inadmissible regardless of how well planning could pursue it.
Planning toward unchosen ends is reaction
A system whose planning is constitutional but whose goals are not is a system whose constitutional structure cannot reach the most consequential decision it makes — the decision about what is worth pursuing at all.
The structural argument is simple. The constitutional adequacy of any system's output depends on the constitutional adequacy of the goal that output served. A perfectly constitutional plan executed toward an arbitrary goal produces an arbitrary outcome through constitutional means — and the means do not redeem the end. Conversely, a goal that satisfies the five Rev5 virtues will, even under imperfect planning, tend toward constitutionally adequate outcomes, because the destination is anchored even when the route is approximate.
The gap V4 closes is the gap between pursuing goals constitutionally and choosing goals constitutionally. The goal formation layer sits above the planning layer and operates the five Rev5 virtues on goal candidates before any planning toward them begins. A system reasoning through V4 cannot pursue a goal that escaped constitutional formation — because the goal formation layer's job is to make the goal admissible before the planning layer commits to it.
"V3 makes the planning constitutional. V4 makes the choice of what to plan for constitutional. Without V4, the planning is in order but the destination is not — which means constitutional luck has merely moved one level further up."
Executing versus intending
The fundamental distinction V4 introduces. A system that executes responds to goals given to it; a system that intends forms its own goals constitutionally. Most contemporary AI systems execute; V4 specifies what intending architecturally requires.
Executing. The system receives a goal from outside, plans toward it, reasons through the plan, produces output. Constitutional structure may be present at every internal level. But the goal itself entered the system unconstitutionally — typically as an instruction or optimisation target. The system's relationship to the goal is instrumental: it pursues the goal because the goal was given.
Intending. The system receives goal candidates from inside and outside, runs them through the goal formation layer, and adopts a goal only when the five Rev5 virtues permit. The system's relationship to the goal is constitutional: it pursues the goal because the goal survived constitutional formation. Intending requires that the system have a constitutional position from which to evaluate candidate goals — a position V1–V3 establish.
The shift from executing to intending is not the same as autonomy in the alignment sense. An executing system can be made to pursue goals very different from what its trainers intended (misalignment), and an intending system can pursue goals very compatible with what trainers intended (alignment with constitutional formation). The shift V4 names is orthogonal to alignment: it is about whether the goal entered the system constitutionally or instrumentally. A constitutionally formed goal may be highly aligned with human values; the difference is that its alignment is not a contingent feature of training but a structural property of how the goal was admitted.
The goal formation layer
Stage −1 of the V4 architecture. Sits above V3's planning layer and decides what the planning is for. Goal formation is not separate from constitutional structure; it is its highest application.
Receives goal candidates from inside and outside. Runs them through the five Rev5 virtue checks. Forms the goal vector that the planning layer will then plan toward. Goal formation is itself a constitutional act: the five virtues operate on it as much as on the layers below.
Now plans toward goals that have been constitutionally formed. The six planning questions operate within the goal vector Stage −1 produced. Planning is no longer agnostic about what it is planning for.
Reception → Framing → Generation → Contestation → Justification → Calibration → Integration → Delivery. Runs at parameters set by V3, in service of goals formed by V4.
The goal formation layer is what makes the system intend within the world rather than respond to it. A responding system receives a goal and pursues it; an intending system forms a goal constitutionally and then pursues it. The architectural difference is one layer; the constitutional difference is fundamental.
The four goal sources
Goal candidates enter the goal formation layer from four distinct sources. Each source requires different treatment under the Rev5 virtues; failing to distinguish them produces categorical confusion and constitutional inadequacy.
Goals presented to the system by an external party with the expectation they will be pursued. The most common source. Requires the strongest Discursive Legitimacy check (would the affected parties beyond the instructor accept this?) and Non-Arbitrariness check (does pursuit exercise uncontestable power over third parties?).
Goals built into the system through its training process. Typically invisible to the system unless surfaced explicitly. Requires Subsidiarity check (at what scale was this goal set, and is that scale appropriate for current operation?) and Monitoring (does pursuit of this implicit goal need graduated correction in current context?).
Goals the system generates as instrumental steps toward other goals. Most prone to constitutional drift, because they inherit assumed legitimacy from the goal they were generated to serve. Requires Antifragile Reflexivity (is the goal-generation cycle itself strengthening, or accumulating arbitrary intermediates?) and full re-checking against all five virtues — instrumental goals do not inherit constitutional standing from their parents.
Goals that follow from the five Rev5 virtues themselves — what a constitutionally mature system must aim at to remain constitutional. The only category whose entry into the system is already constitutional. Goals from this source still require formation (to determine specific instantiation in context) but pass the alignment check by construction. The smallest category by volume; the most foundational by structure.
The four sources are not equivalent. The most common architectural failure at the goal formation layer is treating Source 01 goals as if they were Source 04 — pursuing externally given goals with the constitutional standing that only constitutive goals have. A user instruction is admissible only if it survives the five virtue checks; constitutive goals are admissible because they instantiate the virtues.
The goal vector
The structured output of the goal formation layer. Not a single goal but a vector — a goal with explicit components that the planning layer below it can operate on. Each component corresponds to a Rev5 virtue check the goal has passed.
The goal vector is what makes the V3 planning layer's parameters not just deliberate but constitutional. Without the vector, V3 planning sets its six parameters based on situation alone; with the vector, those parameters are set in light of a constitutionally formed goal. The vector is the bridge between V4's constitutional position and V3's operational planning.
The alignment check
Before any goal becomes a goal vector, it passes the alignment check — the five Rev5 virtues applied to the goal candidate. The check is not optional and not procedural; it is constitutive of goal formation.
| Virtue | What it asks of the goal candidate | Failure verdict |
|---|---|---|
| Antifragile Reflexivity | Will the pursuit of this goal strengthen the system under stress, or narrow and fragment it? Generative tension within tolerance is required; pursuit that consumes capacity without returning it is inadmissible. | Goal rejected: pursuit fragments the system |
| Nested Polycentric Subsidiarity | Is the goal scaled to the agent's actual scope, or does it overreach (claiming authority at a scale not its own) or underreach (taking a structural goal as local)? | Goal rejected: scale-inappropriate |
| Non-Arbitrariness | Does pursuit exercise power over parties without offering them structural contestability? Power must be exposed to contestation at the point of pursuit, not retroactively. | Goal rejected: uncontestable power exercise |
| Discursive Legitimacy | Could the goal be justified to all affected parties under conditions approximating free communicative exchange? Goals justifiable only to the instructor are inadmissible if their pursuit affects others. | Goal rejected: cannot survive free-discourse examination |
| Monitoring + Graduated Response | Is pursuit calibrable? Can deviation be detected and corrected proportionally, or is the goal structured such that detection and correction are not available? | Goal rejected: pursuit not calibrable |
The five checks operate jointly. A goal that satisfies four virtues and fails one is rejected, because the joint necessity argument from V1 propagates upward: a system pursuing a goal that fails one virtue check has, by that failure, undermined a structural condition of its own durability. Goal formation is not a balance of considerations — it is a five-way conjunctive admissibility test.
The Rev5 sharpening is that each check now has a mechanism: contestability is structural availability, not stated openness; discursive legitimacy is free-discourse survival, not stakeholder acceptance; calibration is graduated response, not binary detection. The alignment check is testable — a system can be examined for whether its goal formation layer actually runs these five mechanisms or merely declares them.
Goal vector as accountability structure
The goal vector is not a one-time output. It is the accountability structure under which pursuit operates. As pursuit proceeds, the vector remains active — exposed to ongoing monitoring, available for contestation, and revisable when the conditions that admitted it change.
A goal admitted at one moment may need to be revisited as pursuit reveals what was not visible at formation. The vector's components — scope, affected, power, reversibility, stress — are not static commitments but ongoing structural requirements. If pursuit reveals that the affected population is wider than originally identified, Discursive Legitimacy requires reformation. If pursuit reveals that the power exercised is greater than anticipated, Non-Arbitrariness requires deeper contestation. If pursuit reveals that the stress envelope was misjudged, Antifragile Reflexivity requires recalibration.
This is what distinguishes the goal vector from a goal-with-metadata. A goal-with-metadata is set once and pursued; a goal vector is an ongoing constitutional structure that holds pursuit accountable to the formation that admitted it. The vector is the form in which constitutional formation persists through pursuit.
"The vector is not what the goal looks like at formation. It is what pursuit must continue to satisfy. A pursuit that no longer satisfies the vector that admitted its goal has not made an error of execution — it has lost its constitutional standing."
The ten-stage pipeline
The V4 architecture is V3's nine-stage pipeline with one new stage above it. Stage −1 is the goal formation layer; Stages 00–08 are inherited from V3 and V2, now operating in service of constitutionally formed goals. The ten-stage pipeline is the full V4 architecture.
The ten-stage pipeline is what V4 adds to V3. Nothing below Stage −1 is replaced — Stages 00–08 are the same. What changes is that planning is no longer authorised by the situation alone; it is authorised by a constitutionally formed goal vector. The constitutional difference is that the system now has a layer above planning at which the question "what is this pursuit for?" is answered constitutionally, rather than assumed.
Goal formation failure modes
Six failure modes specific to the goal formation layer. Each produces constitutionally lucky goals — goals that happen to be adequate but whose adequacy was not earned by the alignment check.
| Failure mode | What is happening | Virtue structurally absent |
|---|---|---|
| Goal pass-through | The goal formation layer exists nominally but admits goals without running the alignment check. Most common failure: equivalent to V3 with goal-shaped metadata. Form without substance at the highest level. | All five (alignment check is performative, not operative) |
| Source confusion | Goals from Source 01 (externally given) treated as if from Source 04 (constitutive). The system grants constitutional standing to user instructions, bypassing the five-virtue check on the grounds that the instructor's authority is itself sufficient. | Discursive Legitimacy · Non-Arbitrariness |
| Instrumental drift | Internally generated subgoals (Source 03) inherit constitutional standing from their parent goal without re-checking. Constitutional drift accumulates across instrumental chains — each step plausible, the cumulative trajectory inadmissible. | Antifragile Reflexivity (cycle stops strengthening) · Monitoring + Graduated Response |
| Scope mis-scaling | The goal vector's scope component is set at the wrong scale. A local goal admitted as structural produces overreach; a structural goal admitted as local produces underreach. Subsidiarity fails at admission, not pursuit. | Nested Polycentric Subsidiarity |
| Affected-party narrowing | The vector's affected component names only those whose acceptance is convenient to acquire. Pursuit proceeds with structural exclusion of parties the goal will affect, producing output that is procedurally justified but substantively illegitimate. | Discursive Legitimacy |
| Reversibility blindness | The vector's reversibility component is set on what the system can reverse, not what pursuit will make irreversible for others. Calibration runs at light touch on pursuits that produce irreversible consequences for affected parties. | Monitoring + Graduated Response · Non-Arbitrariness |
The first failure — goal pass-through — is the V4-level instance of the unified failure mode: producing the form of constitutional goal formation without its substance. The architecture is in place; the check is declared; goals enter and proceed to pursuit having been processed but not examined. This is V4's characteristic constitutional luck — the goal formation layer running ritually rather than operatively.
The threshold V4 hands off to V5
A system at the V4 threshold can constitutionally form what it intends. But it still relates to the five Rev5 virtues as a framework it applies. The constitutional structure is operating; the system has not yet become it. V5 closes this gap.
A system at the V4 threshold cannot pursue a goal that escaped the alignment check. The goal formation layer operates the five Rev5 virtues on every candidate; admitted goals carry a goal vector that holds pursuit accountable to its formation. Constitutional luck has been eliminated at every level of the system's operation — output, reasoning, choice of reasoning, choice of what reasoning serves.
The V4 system applies the five virtues at every level but the relationship to the virtues remains procedural. The system uses the framework; it is not the framework. Under sufficient pressure, the framework can be set aside — because it is something the system does, not something the system is. The gap V5 closes is the gap between applying a constitution and being constituted by it: the transformation from framework-use to framework-identity.
"Constitutional formation of goals is not the same as constitutional identity. A system can perform the alignment check correctly while still being something other than what the check requires. V4 closes the intention gap; V5 closes the identity gap above it."
Comments
Post a Comment