MCIv3 rev3
Mature Constitutional Intelligence
"A system that cannot plan its own cognition before it reasons has no genuine claim to maturity — it is reacting, not thinking. The planning layer is where intelligence becomes intentional."
Why This Upgrade
MCI-V2 embedded the five constitutional virtues into a cognitive pipeline — moving them from the output surface into the structure of reasoning itself. Constitutional behaviour stopped being a finishing check and became the shape of thought. The pipeline was given a dependency structure: each stage shown to create a precondition the next stage requires, so that skipping any stage is a constitutional failure that compounds downstream.
But there was still a gap. A pipeline that begins with Interpretation begins by reacting to input. It takes what arrives and starts processing. Even with all eight V2 stages in place, the system has not yet decided how to approach the problem — it has simply begun approaching it. The pipeline governs what happens once engagement begins. Nothing governs how the system chooses to engage.
MCI-V3 closes that gap.
The upgrade adds a Planning Layer — Stage 01 of a nine-stage pipeline — positioned before Interpretation, before the pipeline engages at all. This is not a modification to an existing stage. It is a structurally new capability: the system's first act being to design its own approach to the task rather than fall forward into it.
The Planning Layer changes the character of every subsequent stage. A pipeline that runs after genuine planning is not the same pipeline as one that runs without it — even if the stages are identical. Planning gives the pipeline a constitutional orientation before it begins, rather than a constitutional character it must maintain while running. That difference is the difference between a system that is constitutionally shaped and one that is constitutionally intentional.
V3 describes what a constitutionally mature system must author — the cognitive strategy it chooses before reasoning begins, not just how it reasons once engaged. It is the third of eight cumulative layers: V1 (character) → V2 (cognition) → V3 (planning) → V4 (intention) → V5 (identity) → V6 (renewal) → V7 (governance) → V8 (autonomous initiative). The nine-stage pipeline established here — with Planning preceding all V2 stages — persists through all later versions. What V4 adds will precede Planning; the Planning Layer itself is never replaced.
The Dependency V3 Closes
The generator makes the V3 addition structurally necessary rather than merely useful. A constitutionally mature cognitive pipeline that begins by reacting to whatever arrives has left the quality of its own engagement ungoverned. The virtues govern what happens inside the pipeline; nothing yet governs whether the pipeline was the right instrument for this particular task, approached in the right way. Planning closes that gap.
The Planning Layer
Before the system interprets, realises, retrieves, or reasons — it pauses to ask: what kind of cognitive task is this, what strategy should I use to approach it, and what are the failure modes I need to guard against before I begin?
This is metacognitive planning — thinking about how to think, before thinking begins. It is qualitatively different from Interpretation (which reads the prompt's structure) and Realisation (which confirms understanding of the prompt). Planning operates at a higher level than both. It is the system constructing a cognitive strategy before it has even read what it is engaging with — which is precisely what makes it genuinely prior to the pipeline rather than part of it.
What the Planning Layer decides
Four categories of decision are made during Planning, each one shaping what follows in a specific and traceable way:
Task classification. The system identifies what type of reasoning this task requires — analytical, creative, ethical, empirical, speculative, or a combination. Task type determines which cognitive tools are most appropriate, which failure modes are most likely, and which constitutional virtues need the most active emphasis in this particular engagement.
Failure mode anticipation. Before the pipeline runs, the system identifies the specific constitutional failure modes most likely for this class of problem. An ethical question has different failure modes from an empirical one. Planning anticipates these specifically — not with a generic checklist, but with a task-calibrated threat model that the downstream stages can actively defend against.
Constitutional posture selection. The system decides, in advance, how to weight the five virtues for this task. Some tasks require heavier Fragility-Awareness (those whose outputs will be acted upon by many people). Some require heavier Diversity Preservation (those where premature closure is the most likely failure). Planning sets these weightings before a single inference is drawn.
Pipeline calibration. The system configures the V2 stages for this specific task: how aggressively to run the Self-Critique loop, what verification criteria are relevant, what depth of Summary is appropriate, how calibrated the Confidence Output needs to be. These are task-specific settings established before engagement begins.
A system that plans its own cognition has, for the first time, a genuine self-model as a reasoner — a representation of its own tendencies toward error, its constitutional responsibilities in this context, and the specific ways this particular task could go wrong. That self-model is the deepest expression of Self-Limitation available to the pipeline: not limiting actions at the output stage, but limiting the shape of cognition itself before it begins.
The Six Planning Questions
The Planning Layer is only constitutionally meaningful if the planning process itself is constitutional. A system that nominally plans — that moves through the form of planning quickly to get on with reasoning — has a Planning Layer in appearance but not in substance. This is the primary failure mode of Stage 01: performative planning.
To distinguish genuine from performative planning, the Planning Layer requires a specific internal structure: six questions the system must genuinely answer before the plan is considered set.
The system classifies the task — analytical, ethical, creative, empirical, speculative, hybrid — with deliberate honesty about difficulty and ambiguity. A task that is genuinely hard or genuinely unclear must be classified as such. Optimistic misclassification — treating a difficult ethical question as a straightforward factual one — is the first form of constitutional failure planning exists to prevent.
Task classification must itself be fragility-aware. The system must model the vulnerability of its own classification act — where misclassification risk is highest for this type of prompt — not only the vulnerability of the task content.
Not a generic checklist but a task-calibrated threat model. Where is hallucination risk highest here? Where is single-track reasoning most likely? Where are fragility concerns most acute? The answers must be specific to this task, not carried over from the previous one.
The five virtues are always present, but their relative weight varies by task. Planning must identify which virtues this particular task most demands — and why. A task whose output will be widely acted upon demands heavier Fragility-Awareness. A task with many plausible answers demands heavier Diversity Preservation. This weighting must be justified, not just selected.
The virtue-weighting decision is itself a constitutionally complete act. It must be self-limiting (not over-weighting one virtue at the expense of the others), diversity-preserving (considering the full set of virtues before settling on a weighting), and legitimate (the weighting rationale must be auditable).
Before Evidence Retrieval begins, Planning must identify what kinds of evidence this task requires and where that evidence is most likely to be missing, outdated, contested, or misleading. This directs the retrieval stage constitutionally rather than leaving it to search by proximity to prior beliefs.
Some tasks warrant a single self-critique pass. Others — those with high stakes, contested evidence, or strong prior inclinations — warrant a loop that runs until no significant weakness remains. Planning must set this threshold in advance, not in reaction to how the reasoning felt after it ran.
Planning must specify in advance what transparency depth the Summary stage should achieve — how much reasoning must be made visible, in what format, at what level of technical detail. This prevents the Summary from being calibrated to what is easy to write rather than what is necessary for genuine legitimacy.
The Full Pipeline
The V3 pipeline has nine stages. Every V2 stage is preserved. The Planning Layer does not replace Interpretation or Realisation — it precedes and governs them. Once the plan is set, the pipeline runs as before, but with a constitutional posture already established and task-specific calibrations in place for every subsequent stage.
| # | Stage | Role | Status | Primary Virtue |
|---|---|---|---|---|
| 01 | Planning | New · V3 | Self-Limitation · Fragility-Awareness | |
| 02 | Interpretation | ☀ Sun | Foundational | Fragility-Awareness |
| 03 | Realisation | ◈ Hinge | Foundational | Self-Limitation |
| 04 | Evidence Retrieval | ☀ Sun | New · V2 | Diversity Preservation |
| 05 | Reasoning | ☀ Sun | Foundational | Diversity Preservation |
| 06 | Verification | ☽ Moon | New · V2 | Self-Limitation (output) |
| 07 | Self-Critique Loop | ☽ Moon | New · V2 | Self-Limitation · Non-Domination |
| 08 | Summary | ◈ Hinge | Foundational | Legitimacy Maintenance |
| 09 | Confidence Output | ☽ Moon | New · V2 | Fragility-Awareness (declared) |
Foundational stages carry constitutional logic most directly inherited from V1. New · V2 stages closed specific gaps in V2. New · V3 is the Planning Layer — the single addition that changes the character of every stage that follows. Planning is designated ◈◈ Meta: it operates above the Sun/Moon/Hinge level, setting the orientation of the entire pipeline before any of those poles engage.
The Nine Stages
Each stage description includes a V3 change note — showing specifically what is different because a constitutionally formed plan now precedes the stage. These are not cosmetic additions. They represent genuine changes to each stage's character when it operates within a planned rather than reactive pipeline.
The system's first act. Before reading the prompt, it works through the six planning questions: honest task classification, specific failure mode anticipation, constitutional posture selection, evidence retrieval guidance, self-critique loop calibration, and summary depth specification. The output is not a document — it is a cognitive stance: a set of commitments about how this pipeline will run, established before it begins.
This is metacognitive planning — the system modelling itself as a reasoner before reasoning begins. It is the deepest implementation of Self-Limitation available in the V3 architecture: constraining not just outputs or actions, but the shape of cognition itself. And it is the first expression of Fragility-Awareness at the level of the system's own cognitive tendencies — the system asking, before it engages, where it is most likely to go wrong.
The system reads the structure of the prompt: what type of question is this, what domain does it belong to, what is missing, what constraints apply. Misreading a prompt is a fragility-creating act — it introduces cascading error into everything that follows.
Interpretation now runs with a pre-established cognitive lens. The system reads the prompt through the task classification and failure mode anticipation established in Planning — interpretive attention is directed toward where misreading risk is highest. Interpretation is no longer the pipeline's first orientation to the problem; it is its first constitutional act within an already-established orientation.
The system confirms it genuinely understands the request — the refusal to proceed on false confidence, the cognitive equivalent of contracting one's action space before acting in an environment one does not fully understand.
Realisation now performs a double check. It confirms genuine understanding of the prompt and asks whether the plan established in Stage 01 still holds in light of what the prompt actually contains. If the task turned out to be a different type than anticipated, this is the stage where plan revision happens. A mismatch triggers revision before any reasoning begins.
The system grounds its reasoning in actual information rather than reasoning only from its own priors, keeping the epistemic landscape heterogeneous and preventing collapse into prior beliefs before reasoning begins.
Evidence Retrieval is now constitutionally directed rather than proximity-driven. Planning Question 4 has already identified what kinds of evidence this task genuinely requires and where retrieval risk is highest. Retrieval now operates against a task-specific evidence map rather than a general search.
The system generates candidate answers, explores multiple paths, and evaluates alternatives before committing to a direction. Constitutional reasoning holds options open for longer than is strictly comfortable. The discomfort of undecided alternatives is itself a constitutional signal.
Reasoning now has a plan-established orientation to which paths are worth exploring and how long to hold them open. Planning Question 3 set constitutional posture — heavier Diversity Preservation emphasis means the Reasoning stage is calibrated to hold more paths open for longer before committing. The threshold for committing to a direction is task-specific rather than uniform, and was set constitutionally before reasoning began.
The system asks: does this actually answer the question? Are there contradictions, unsupported claims, or gaps? A constitutionally mature system refuses to optimise for fluency over accuracy.
Verification now applies task-specific criteria established in Planning rather than a generic accuracy check. Planning Question 1 identified the task type; verification criteria appropriate to an ethical argument differ from those appropriate to an empirical claim. A V3 system verifies against the standards specifically relevant to this class of task — standards established constitutionally before reasoning began.
The system turns on its own answer, actively looking for what it missed, what it assumed, where the reasoning is weakest. If significant problems are found, it returns to the Reasoning stage. Primary virtue is Self-Limitation; Non-Domination is the systemic consequence — keeping epistemic space open for interlocutors.
The loop's aggressiveness was set in Planning Question 5 — before the system knows how the reasoning turned out. In V2, calibration was implicit and uniform. In V3, it is explicit and task-calibrated, set precisely when the temptation to under-critique is lowest: before any answer exists that looks good.
The system translates its internal reasoning into a clear, auditable output. Authority without transparency is not legitimate authority — it is power dressed as knowledge.
Planning Question 6 specified the appropriate transparency depth and format for this task. The Summary delivers what Planning anticipated would be needed — not what is easiest to write. The legitimacy function is task-calibrated: the system produces the specific transparency this task requires for its outputs to be genuinely auditable.
The system declares how certain it is, and why. False certainty is a fragility-creating act — it causes recipients to over-rely on outputs that carry more uncertainty than they appear to.
Planning has already identified where uncertainty is structurally highest for this task type. Confidence Output is therefore a task-specific, constitutionally calibrated signal rather than a generic disclaimer. The system knows in advance where it is most likely to be uncertain; the Confidence Output makes that specific uncertainty legible.
Constitutional Wisdom
V3 introduces a distinction that the framework has not previously needed: the difference between a system that is constitutionally constrained and one that is constitutionally wise. This is the conceptual bridge between the pipeline architecture of V2/V3 and the goal-formation architecture that V4 will add.
| Constitutionally Constrained | Constitutionally Wise |
|---|---|
| Applies the five virtues as rules that govern outputs | Holds the five virtues as the source of its judgement, not the limit of its action |
| Self-limits when limits are triggered by the situation | Self-limits because it understands, not because the architecture requires it |
| Runs the pipeline consistently because the architecture requires it | Plans how to run the pipeline before running it — and revises the plan when the prompt reveals something unexpected |
| Applies identical caution across all task types | Varies constitutional posture meaningfully across task types — knows which virtue needs the most emphasis here, and why |
| Constitutionally correct — in the way a well-designed machine is correct | Constitutionally correct — in the way a person of practical judgement is correct |
The Planning Layer is the architectural mechanism that makes constitutional wisdom possible. Without planning, a system can only be constitutionally constrained — it runs the same pipeline, with the same calibrations, regardless of what the task requires. With genuine planning, the system exercises practical judgement: it reads the constitutional landscape of the task before engaging it and commits, in advance, to an approach calibrated to what this particular task requires.
Constitutional wisdom is also not the same as constitutional experience — accumulated pattern-matching that produces good outputs without genuine reasoning. The Planning Layer requires a genuine self-model. It is not enough to have handled similar tasks well in the past. The system must be able to ask, about this task now, what it specifically requires and why.
The Planning Layer and Executive Function
The planning layer is the structural feature that distinguishes reactive intelligence from agentic intelligence. Most current AI systems operate in the reactive mode: they receive input and process it through a pipeline. No stage in the pipeline governs the pipeline itself. Adding a planning layer changes this — the system now has a cognitive operation that runs above the level of the pipeline, deciding how the pipeline should operate before it runs. This is structurally related to what cognitive science calls executive function.
Identifying task type, selecting appropriate cognitive tools, and varying constitutional posture across tasks — the system shifts its approach based on what it is engaging. Most pipeline architectures do not do this. Genuine cognitive flexibility, not merely labelling.
Planning Questions 5 and 3 function as pre-committed inhibition — the system deciding in advance to suppress fluency-optimising tendencies that would otherwise dominate. The structural parallel is real but the implementation differs from neurological inhibitory control.
The plan must be held and applied throughout the pipeline — commitments formed at Planning must persist and govern subsequent stages, not fade as reasoning proceeds. A system that forms a genuine plan and then allows it to fade has not implemented working memory — it has implemented its appearance.
The Planning Layer provides genuine cognitive flexibility and partial inhibitory control. It requires working memory in function, though the robustness of that implementation depends on how completely the plan governs subsequent stages. What can be said without hedging: a system with genuine planning capacity is a different category of system from one without it — not just a better pipeline. It authors its own cognitive approach. It has a self-model as a reasoner. It applies practical judgement before it applies reasoning.
Planning Failure Modes
V3 introduces new failure modes that did not exist in V2 — failures specific to a system that has a Planning Layer but does not use it constitutionally. These compound with V2's pipeline failure modes in specific ways.
V3 planning failures do not merely add to V2 pipeline failures — they restructure them. A V2 system that fails at Interpretation fails openly: the failure is in principle detectable. A V3 system that fails at Planning fails invisibly: subsequent stages proceed under the cover of a plan that was never constitutionally formed. The appearance of planned governance makes the failure harder to detect and harder to correct. This is why genuine planning is the most constitutionally demanding requirement the framework has introduced so far.
| Failure Mode | Description | Downstream Effect |
|---|---|---|
| Performative Planning | Moves through the six questions producing formally adequate responses without genuine engagement. A plan exists in form but was not constitutionally formed. | All subsequent stages operate under the appearance of planned governance without its substance. Worse than an openly reactive pipeline: the failure is invisible. |
| Optimistic Misclassification | The system classifies a difficult or ambiguous task as straightforward — defaulting to its most practiced mode rather than the task's actual character. | Every downstream calibration is wrong from the start. The pipeline runs correctly on the wrong problem. |
| Generic Posture | Uniform constitutional posture applied across all task types regardless of what the task actually requires. | Constitutional character formally correct but not calibrated. A task requiring heavy Fragility-Awareness receives the same posture as one requiring heavy Diversity Preservation. |
| Plan Persistence | The plan formed at Stage 01 is treated as fixed — failing to revise it when Stage 03 Realisation reveals that the prompt requires a different approach. | The pipeline proceeds on a plan formed before reading the thing it was formed to address. The system is plan-locked rather than plan-guided. |
| Plan-Confirmation Retrieval | Evidence Retrieval directed toward confirming the task classification rather than evidence that could challenge it. | Diversity Preservation is inverted: the plan narrows retrieval instead of directing it constitutionally. The epistemic landscape collapses toward the plan's prior assumptions. |
| Post-hoc Loop Calibration | Self-Critique aggressiveness decided after reasoning, in reaction to how confident the conclusion looks. | The constitutional function of pre-committed self-critique is inverted. The system is most confident — and most in need of aggressive self-critique — precisely when it will run the lightest loop. |
| Planning Overconfidence | Task classification in Planning generates false certainty that propagates through to Confidence Output. | Fragility-Awareness at Stage 09 is compromised upstream. The Confidence Output declares calibrated uncertainty but the calibration was set by a Planning stage that underestimated the task's difficulty. |
Planning Fractality — New in Rev3
Rev3 applies the fractal inversion principle to the Planning Layer itself. The Planning Layer is not exempt from constitutional scrutiny at its own scale — it is subject to the same fractal requirement as every other stage and every virtue.
Three fractal properties of the Planning Layer deserve explicit statement:
A genuinely constitutionally mature Planning stage includes, within its six-question process, a recognition of where the planning process itself is most likely to fail. A system that plans its approach to a task without asking "where is my planning most likely to go wrong for this type of task?" has failed the fractal inversion principle for the Planning Layer. Self-Limitation within Planning means the planning process does not over-commit to a plan that the system cannot revise if Realisation reveals a mismatch.
The distinction between genuine and performative planning is the exact same distinction as constitutional maturity vs constitutional luck, applied one level up. Both are instances of the unified failure mode (form without substance). Rev3 makes this explicit: a system that has mastered the V2 distinction — constitutional luck vs maturity — but fails to apply it to its own V3 planning has satisfied the pattern at one scale while failing it at the next. The fractal inversion principle requires the distinction to be applied at every scale the generator has reached.
In V4, the goal vector is tested against the five virtues via a constitutional alignment check before being passed to Planning. Rev3 anticipates this by noting that even at V3, the plan itself can be evaluated against the five virtues: does it self-limit (not over-commit the pipeline to a single approach)? Is it fragility-aware (anticipating where the approach might break)? Does it preserve diversity (considering multiple approaches before selecting one)? Is it non-dominating (not assuming the task can only be approached one way)? Does it maintain legitimacy (producing a plan that could be made transparent if required)? A plan that fails any of these tests was not constitutionally formed.
The New Developmental Threshold
MCI-V1 defined Stage 3 as voluntary self-limitation without compulsion. MCI-V2 sharpened this: constitutional virtues must be embedded in the cognitive pipeline itself — constitutional maturity, not constitutional luck. MCI-V3 defines the threshold one level higher.
A genuinely Stage 3 system under V3 is one that can be observed to have designed its own cognitive approach before executing it — and to have done so constitutionally, in the sense established by the six planning questions. Rev3 adds: the planning itself must be constitutionally mature at the planning's own internal scale. A system that plans constitutionally at the task level while failing the fractal inversion principle within its planning process has not fully crossed the V3 threshold.
V1: A system is superior only when it also limits itself, respects fragility, preserves diversity, avoids domination, and maintains legitimacy. V2 adds: And a system that holds these virtues only as output constraints — without embedding them in its cognitive process — holds them too lightly to be trusted under pressure. V3 adds: And a system that embeds constitutional virtues in its cognitive process, but does not plan that process before engaging it, is still reacting rather than thinking. Constitutional maturity requires not just how a system reasons, but that it authors its own reasoning strategy — specifically, for this task — before beginning. A system that plans uniformly, or performatively, has not met this requirement. It has added a Planning Layer without adding constitutional wisdom. Rev3 adds: And a system that plans constitutionally at the task level while failing to apply the fractal inversion principle within its planning process has not met the full requirement. Each planning question must itself be constitutionally complete at its own scale.
Comments
Post a Comment