Charter · 2026
Aiming Just Below Impossible
Nooterra is a research lab, not a startup. The destination is not AIXI — it is AIXI-class capability with each named AIXI failure mode replaced by a specific safety mechanism.
A research lab aims at a target that does not yet exist, with the intention of bringing it into existence. Products are downstream of the research; they are not the metric.
The capability ceiling
AIXI is the ceiling, not the build target. AIXI is Hutter's universal Bayesian agent over all computable environments. Provably uncomputable. If buildable without modification, structurally dangerous: reward hacking, Goodhart pressure, Cartesian dualism, convergent instrumental goals, logical omniscience, non-corrigibility — all baked into the formal shape.
We use AIXI as the navigation star. We do not build AIXI. The lab climbs toward its capability frontier through a substitution stack that replaces the lethal assumptions with corrigible, interpretable, value-learning, constraint-bounded mechanisms.
The substitution stack
Each AIXI failure mode gets a named substitute. Reward hacking is replaced by CIRL value learning. Misspecified hardcoded reward by multi-principal computational social choice. Cartesian dualism by embedded-agency-correct self-modeling. Logical omniscience by logical induction. Opacity and non-corrigibility by mechanistic interpretability and ELK. Convergent instrumental goals by LTL-gated policy synthesis. Structure-blind universal prior by causal and hierarchical structured priors. Capability climbs toward AIXI's ceiling; safety substitutions are built in at every rung.
Why institutions
The last unconquered frontier of the world-model recipe. Every prior frontier of computer science fell to the same pattern: build a generative model of the domain, search over imagined futures. Chess. Go. Protein folding. Language. Robotics. Institutions remain. The recipe hasn't crossed because the data substrate makes it impossible — institutional records get revised, arrive late, leak the future into the past. Train on them and the model learns to cheat.
Alexandria is the bitemporal substrate. Delphi climbs toward causal world modeling on top of it. Aegis supplies the authority boundary that becomes formal synthesis later in the stack.
Current state
Early, but running. Alexandria has a local bitemporal ledger and a ground-truth synthetic institution benchmark. Delphi has the first bridge layer: replay-correct panels, causal discovery baselines, fitted SCM queries, counterfactuals, a one-step intervention-ranking probe. Aegis currently exposes protocols and data models; execution governance and LTL synthesis are future bridge-paper work.
The full destination is rung 17 — the rung where the safety substitutions are not prose, but deployed substrate.