NOOTERRA
SoonPaper 1 — Causally Structured Latent World Models
Nooterra is a research lab building a causally-structured decision substrate for institutions. We start from replay-honest data, climb a safety-aware substitution stack on top of it, and publish each rung as a falsifiable paper before moving on.
The substrate is early but running. Alexandria is a local Parquet + DuckDB ledger with event time, system time, monotonic writes, and replay-by-cutoff. Delphi is the first bridge layer: replay-correct panels, dependency-free discovery baselines, fitted SCM intervention and counterfactual queries, and a one-step intervention-ranking probe on synth_institution. Aegis is the authority surface; LTL synthesis and execution governance are future bridge-paper work.
AIXI is the navigation star, not the design target. The full destination is rung 17 of the computable-approximation tower — the rung where every named AIXI failure mode has a deployed substitution in the substrate. We're at rung 7. The first five bridge papers take us to 11.
Nooterra Labs
We are building the substrate that lets a learned world model train on institutional data without learning to cheat the future, with named safety substitutions for each of AIXI's structural failure modes built in at the math layer rather than bolted on.
Replay-honest data
Bitemporal first, world model second. Every observation is timestamped twice — when it happened in the world and when we learned about it. Training only ever sees what was knowable as of a chosen system time. Without that boundary, every world model trained on revised institutional records is built on look-ahead leakage.
Causal world modeling
Counterfactuals, not correlations. The world model has to answer what would have happened if, not what tends to happen near. Bridge paper #1 adds a Structural Causal Model layer on top of Delphi's latent transitions and benchmarks five dependency-free discovery baselines on a ground-truth synthetic institution with 12 variables.
Safety as substitution
Each AIXI failure mode gets a named replacement. Reward hacking → CIRL value learning. Cartesian dualism → embedded-agency self-modeling. Logical omniscience → logical induction. Opacity → mechanistic interpretability and Eliciting Latent Knowledge. Convergent instrumental goals → LTL-gated policy synthesis. Structure-blind universal prior → causal and hierarchical structured priors. The substitutions ship as bridge papers, not as a year-ten afterthought.
Current state
Early, but running. The v0.1 substrate report has a source-only arXiv bundle prepared and the endorsement packet ready; it is not yet on arXiv. The v0.2 causal-structure manuscript is an internal manuscript seed with a 10-seed aggregate, revision-stress diagnostic, and sample-size ablation committed alongside it. Repository-specific public links stay disabled until the lab changes repository access state.
Join us
We're a small lab — currently a solo researcher plus AI tooling, operating in public against a multi-year research target. The charteris the long version of why this lab exists. If the research direction resonates and you'd like to collaborate, contribute, or join, write directly.
Hiring + collaborationaiden@nooterra.ai