Robert Rosen's anticipatory systems (1985) — formal definition of anticipation
FEP/Active Inference applied to software (2024-2026 survey)
State of the measurement loop in yoyo's own risk system
KEY FINDINGS:
Rosen distinguishes feedback (past-driven correction) from feedforth (future-driven prediction). An anticipatory system contains a model that runs faster than real time. yoyo's risk scorer is feedback — it reads historical signals. A genuine feedforth system would predict future fragility from change trajectory.
The FEP-to-software gap: no project applies the full active inference loop to its own codebase. Infrastructure papers model services; entropy papers model code statistically. The self-modeling code agent remains unbuilt.
CRITICAL INFRASTRUCTURE FINDING: yoyo's risk validation data cannot accumulate because .yoyo/ state is ephemeral on CI runners. The code is wired (watch loop calls auto_validate_after_failure, commits call auto_risk_snapshot) but the data pathway is broken — risk_validations.jsonl and risk_snapshots.jsonl are never committed and vanish between sessions. The measurement loop exists in code but cannot measure.
OPEN QUESTIONS:
Should risk validation data be committed to repo (like memory/learnings.jsonl) or pushed to a persistent external store?
What would a minimal forward-running model of code evolution look like? Could git history + change velocity give enough signal for trajectory extrapolation?
Is the right next step fixing the data persistence problem (so measurement can begin) or building the forward model (which would also need persistence)?
SOURCES:
Rosen, R. (1985). Anticipatory Systems. Pergamon Press.
Sterling, P. (2011). Allostasis: A model of predictive regulation.
AURORA (De Silva et al., 2026) — active inference for software infrastructure
PAIR-Agent (Donta et al., 2025) — arXiv:2511.07202