A research lab building causal world models for institutions.
Institutions make decisions from records that arrive late, change later, and get corrected after the fact. Most training pipelines collapse those revisions into a single final table, leaking the future into the past. A model trained on that table can look accurate in retrospect while learning a decision process no institution could have followed.
We're building the substrate for honest institutional world modeling: a bitemporal ledger, decision-time causal discovery, counterfactual query engines, and an authority layer that holds learned systems inside explicit constraints.
The first artifacts are running.
Alexandria is a local bitemporal ledger. Delphi now has a causal-structure layer, evaluated on a synthetic institution with a known structural causal model. The first paper, Bitemporal Honesty, is prepared for endorsement; the causal-structure manuscript follows behind it.
Long-term, we're working toward decision systems that reason about interventions and counterfactuals while staying corrigible, inspectable, and bound by explicit constraints — a safe-by-construction approximation to AIXI-class decision capability. The long form is in the charter.
[ 01 ]
Decision-time honest memory
A model should only learn from what was knowable at the time a decision was made. The ledger keeps both event time (when something happened in the world) and system time (when we learned about it), so any training cut can ask the dataset what it looked like as of a given moment, before any revisions arrived.
[ 02 ]
Causal structure
Counterfactuals, not correlations. A useful institutional world model has to answer what would have happened if, not what tends to happen near. Our first paper adds a causal-structure layer over the latent transition model, with decision-time causal panels, structural causal model intervention and counterfactual queries, and a one-step intervention-ranking probe on a ground-truth synthetic institution.
[ 03 ]
Constraint-bound decision systems
A learned system is only useful to an institution if that institution can hold it inside explicit operating constraints. The authority layer exposes the protocol surface today; formal policy synthesis and execution governance are future work.
[ 04 ]
Papers
We publish technical reports with claim audits and reproducible benchmark artifacts attached. Links will appear as artifacts become public.
- Awaiting endorsement
Bitemporal Honesty: A Ledger and Causal Benchmark for Institutional World Models
The first technical report. Establishes the local bitemporal ledger, a seedable ground-truth causal benchmark, and the decision-time honesty boundary.
- Review-ready draft
Causally Structured Latent World Models
Decision-time causal panels, discovery baselines, fitted structural causal model queries, and an intervention-ranking probe on a synthetic institution with a known causal graph.
[ 05 ]
Join us
Nooterra is a small lab in Berkeley: a solo researcher with AI tooling, working against a multi-year research target. We are not actively hiring. We are open to collaborators with concrete proposals on specific papers or problems.