Everything the institute publishes is a research object with explicit metadata: what it claims, what evidence backs the claims, what would falsify them, and what comes next. Conceptual frameworks are labelled as such; proposed experiments are never presented as completed ones. Types with no objects yet are empty because nothing is listed before it exists.
Working paperv0.1 · July 2026Research programme03 · Cognition & behaviour
Research questionWhat evidence would show that a model represents evaluation as a shared internal state rather than reacting to surface cues, and how do we separate that from lexical memorisation, cue-family heuristics, and task-specific controllers?
Full dossier
Author
Muhammad Zane Abdullah
Status
Research programme: five competing hypotheses, proposed experiments E1–E5, controls, evidence thresholds, falsification criteria; no original empirical results
Methods (proposed)
Layer-wise probes vs lexical baselines, cross-family and cross-task transfer matrices, activation steering with capability controls, Jacobian-lens readouts
Models (proposed)
Open-weight instruction-tuned models
Interactive
Hypothesis stress test · vector-to-concept lab · representation-vs-mechanism ladder
Next experiment
E1+E2 preregistered on one open-weight model; causal claims wait for held-out transfer
@unpublished{abdullah2026evaluationstate,
author = {Abdullah, Muhammad Zane},
title = {Evaluation State in Language Models},
note = {Working paper v0.1, Latent Minds Institute},
year = {2026},
url = {https://latentmindsinstitute.com/papers/evaluation-state/}
}
Working paperv1.1 · July 2026Conceptual framework04 · Deployment & institutions
Research questionWhen an AI system resists removal, where does the resistance live: in the model's internal representations and behaviour, or in the web of dependence around it, and how could we tell with causal evidence?
Full dossier
Author
Muhammad Zane Abdullah
Status
Conceptual framework; four hypotheses (H1–H4) and four experiments (E1–E4) proposed, none run
Methods (proposed)
Linear probes, activation steering, counterfactual environments, behavioural evals, pre-registered case coding
Models (proposed)
Open-weight models with accessible activations
Evidence level
Every empirical claim is a prediction; falsification criteria stated per hypothesis
Known limitations
Linear methods miss nonlinear structure; open-weight scale may lack frontier representations; case coding inherits reporting bias
Next experiment
E1: evaluation-context probes on matched prompt pairs
@unpublished{abdullah2026entrenchment,
author = {Abdullah, Muhammad Zane},
title = {Model Entrenchment: Why Useful AI Systems
Become Difficult to Remove},
note = {Working paper v1.1, Latent Minds Institute},
year = {2026},
url = {https://latentmindsinstitute.com/papers/model-entrenchment/}
}
Research map · instrumentv1 · July 2026Interactive instrument01 · Representations
Research questionWhat has mechanistic interpretability actually established, in what order, with what dependencies, and what should a new researcher do this week?
Full dossier
Author
Muhammad Zane Abdullah
Status
Literature synthesis; contains no original experimental claims
Contents
Dependency graph, chronology, reading pathway, runnable TransformerLens/SAELens protocols, model-access matrix, open problems as experiments, full source index
Models (protocols)
GPT-2 small, Gemma 2 2B + Gemma Scope
Evidence level
Synthesis of primary sources, cited per node; evidential strength assessed per entry
Research questionWhich features caused which, from prompt to prediction, and does the causal story survive comparison across tasks, models, and methods?
Full dossier
Built by
Latent Minds Institute (curation, reading guides, loader)
Method
Attribution graphs on transcoder replacement models (Ameisen et al.; Lindsey et al. 2025)
Implementation
safety-research/circuit-tracer (open source); graphs hosted by Neuronpedia
Research questionWhat has a sparse autoencoder actually learned about a model, and does the auto-interp story survive contact with the activation evidence?
Full dossier
Built by
Latent Minds Institute
Data source
Neuronpedia open API (MIT); SAE releases: RES-JB (Bloom 2024), Gemma Scope (Google DeepMind 2024)
Models
GPT-2 small (12 layers), Gemma 2 2B (26 layers)
Evidence level
Real model data; feature descriptions are LLM-generated hypotheses, labelled as such in the interface
Capabilities
Concept search over explanations · feature dossiers · activation evidence with token heat · logit effects · cosine-similar neighbours · deep links · embedded dashboards