Intelligence as growth
The architecture now contains many capabilities: ingestion, retrieval, consolidation, reinforcement, policy, multi-agent reasoning, arbitration, identity, goals, world modeling, reflection, attention, temporal cognition, cognitive economics, forgetting, affect, narrative, meta-cognition, constitutional stability, grounding, social cognition, causal intelligence, curiosity, dreaming, and abstraction. Having all of these capabilities implemented does not mean the system should use all of them at full strength from the first moment of operation.
A newly deployed system lacks the calibration history, memory depth, and policy stability that high-stakes operations require. The developmental cognition engine models the growth of capability maturity over time and gates higher-capability operations on demonstrated readiness.
Capability stages and phase thresholds
The development engine tracks the system through five phases, each unlocking additional subsystems:
- Seed (mean capability below 0.25): basic operations only. The system can ingest, retrieve, and store experience, but nothing more.
- Novice (0.25 to 0.48): adds ingestion, retrieval, consolidation, and policy. The memory substrate is operational but the agent loop and social systems are not yet trusted.
- Apprentice (0.48 to 0.68): adds the cognitive agent loop, multi-agent decomposition, and grounding. The system can reason and act but not yet model identity or affect.
- Integrative (0.68 to 0.85): adds agent society, world model, goals, attention, affect, and meta-cognition. The full self-regulation stack is now operational.
- Open-ended (above 0.85): all subsystems including open-ended evolution mode.
Experiment 27 demonstrated a mean capability of 0.30 producing a novice phase transition (from seed), unlocking ingestion, retrieval, consolidation, and policy. A mean capability of 0.75 produced an integrative transition (from novice), adding agents, world model, goals, attention, affect, and meta-cognition.
Phase unlocks are cumulative: each higher phase inherits all lower-phase subsystems. An integrative-phase system has everything a novice-phase system has, plus the additional integrative-phase subsystems.
Curriculum emergence
The development engine can identify task sequences that build competence. Rather than assigning all tasks uniformly, it prioritizes tasks where difficulty approximately matches current capability: easy enough to succeed reliably, hard enough to push the capability score upward.
This curriculum is not externally assigned. It emerges from the system's own performance history. The engine looks for tasks where the capability score in the relevant domain is between 0.3 and 0.7 times the task difficulty (the "readiness zone"). Tasks too easy (capability much higher than difficulty) produce no learning. Tasks too hard (capability much lower than difficulty) produce only failures and negative reinforcement.
The readiness-based curriculum selection is a form of self-directed learning. The system does not need a human instructor to sequence its development; it selects tasks based on where the expected learning is highest. This is particularly important during the apprentice-to-integrative transition, where the system needs enough successful multi-agent experiences to demonstrate readiness for the full society and self-regulation stack.
Phase transitions and qualitative change
Some developmental changes are gradual. A capability score that rises steadily produces gradual improvements in performance. Other changes appear as phase transitions: qualitative shifts in what the system can do when multiple prerequisites align.
The integrative phase transition is a good example. The system does not gradually begin using identity and affect; it either has them in its active subsystem set or it does not. The transition from apprentice to integrative is discrete, not continuous, because the subsystem inclusion decision is binary.
Phase transitions create moments of increased risk. A system that was operating stably at the apprentice level has just gained access to affect modulation, meta-cognition, and the full multi-agent society. These additions expand what the system can do but also expand the ways it can fail. The constitutional stability layer is particularly important immediately after a phase transition: it monitors for unsafe drift in the newly activated subsystems before they have accumulated sufficient calibration history.
Progressive unlocking and safety-aware growth
The progressive unlocking structure is a safety mechanism. Higher-capability operations (open-ended evolution, full self-modification proposals, autonomous long-horizon planning) require demonstrated competence in lower-capability operations before they become available.
This prevents the system from accessing powerful self-modification capabilities before it has shown reliable calibration on basic retrieval and reinforcement. A system that attempts to modify its own reasoning strategy before its world model is calibrated is likely to make modifications based on flawed predictions.
The unlocking criteria are capability-based rather than time-based. A system that learns quickly and accumulates strong calibration evidence can transition through phases faster than one with weak performance history. A system that operates in a difficult environment and accumulates many calibration failures may remain in an earlier phase longer.
This approach differs from fixed-schedule deployment (activate all features on day 30). Fixed schedules assume that time is a reliable proxy for readiness. Capability-based unlocking uses actual performance evidence as the proxy, which is both more accurate and more resistant to deployment-schedule pressure.