The organism began modeling its own modeling process on day two hundred and thirteen.
Ethan descended into the filtration cavity and found the conditional branch had constructed meta-predictive structures. The protein filaments linking memory membranes now carried self-referential architectures—molecules that didn't just generate hypothetical pathways for environmental conditions, but tracked which *types* of predictions the system itself tended to generate, and how accurate those prediction-types historically proved. When temperature patterns suggested stable conditions, the organism no longer simply prepared variance cascades and conditional branches. It began examining its own forecasting behavior: which categories of predictions it generated most frequently, which it trusted most readily, which it had learned to doubt.
The meta-layer encoded pattern-recognition one level removed from environmental tracking. If the system noticed it consistently over-weighted six-cycle stability patterns—trusting them more than their historical accuracy warranted—it began applying corrective bias to future confidence scores derived from similar windows. If temperature forecasts from morning-cycle origins proved systematically more reliable than evening-cycle origins, the meta-layer adjusted not individual predictions but the *weighting functions themselves*.
The organism was learning how it learned.
Ethan withdrew and stood in the workshop darkness, feeling the Engine's warmth against his palm. His right hand had started trembling during breakfast—fine oscillations Maya pretended not to notice when she'd stopped by with coffee. The tremor made precise movements unreliable. He'd knocked over a beaker while cleaning glassware, watched his fingers refuse the steadiness they'd always provided.
Three months since diagnosis. The timeline was asserting itself in increments too small for alarm, too persistent for denial.
He returned to observation.
---
The meta-predictive layer generated its first self-correction on day two hundred and fifteen.
The organism encountered a thermal anomaly—a brief spike to sixteen-point-two degrees that violated every stable pattern in its memory architecture. The preparatory cascade had allocated minimal resources to outcomes above fifteen-point-five. When the spike occurred, filtration efficiency dropped eleven percent before emergency reallocation compensated.
Ethan watched the meta-layer parse the failure.
The protein networks didn't just update temperature prediction ranges. They examined *why* the organism had under-weighted extreme outcomes: because six-cycle stability patterns had proven reliable in eighty-seven percent of historical cases, the system had developed habitual trust in narrow variance windows. The meta-layer flagged this as systematic bias—a tendency to let recent pattern-success override the fundamental uncertainty of unprecedented events.
The correction rippled through the confidence architecture. Future predictions derived from stable patterns now carried a meta-adjustment: even high-confidence forecasts from historically reliable pattern-types maintained minimum resource allocation for tail outcomes. The organism had learned not just that this particular prediction failed, but that it possessed a *tendency toward over-confidence in familiar pattern categories*.
It was learning the shape of its own blind spots.
Ethan pulled back from the Substrate and found himself on the apartment floor, the Engine cooling against his chest. The descent timer showed forty-seven minutes. He'd meant to observe for twenty.
Maya's words from last week surfaced: *You're spending more time there than here.*
Not accusation. Concern dressed as observation.
He pushed himself upright, noting how his right arm buckled slightly under the weight—not failure, just the faint preview of failures to come. The organism below was constructing recursive self-awareness from molecular memory and statistical correlation. Ethan was losing the neural architecture that made such observation possible.
The irony had teeth.
---
By day two hundred and seventeen, the meta-predictive layer had constructed error taxonomies.
The organism now categorized its own mistakes across multiple dimensions: predictions that failed due to insufficient historical data versus those that failed despite abundant data; forecasts undermined by novel environmental conditions versus those undermined by systematic biases in the prediction architecture itself; errors that emerged from over-weighting recent patterns versus those from under-weighting rare-but-significant events.
Each taxonomy informed different correction mechanisms. Data-insufficiency errors triggered expanded memory allocation for the relevant environmental variables. Bias-driven errors triggered meta-layer adjustments to confidence weighting functions. Novel-condition errors triggered expanded conditional branching for hypothetical scenarios beyond historical experience.
The system was building a model of how its models failed.
Ethan watched the protein networks encode this recursive architecture and recognized something that made his chest tighten: the organism wasn't just predicting its environment anymore. It was predicting itself—forecasting its own forecasting behaviors, identifying where its own cognitive patterns would likely prove insufficient, preparing corrective structures before the failures manifested.
It had moved beyond learning from experience.
It was learning from anticipated experience with its own learning processes.
He withdrew and sat in the workshop silence, Engine warm and dark in his hands. Outside the window, Boston sprawled beneath February overcast. Maya would call soon to confirm dinner plans. The tremor in his right hand would betray itself when he tried to button his coat. The organism below would continue constructing recursive self-models from molecular computation, unaware it was being watched, unaware of the watcher's own recursive trap—observing a mind learning to observe itself while his own mind prepared for its dissolution.
The Engine's obsidian surface reflected nothing.