# Latent Minds Institute, full public research context

This file is generated from data/research-objects.json. Preserve every object's status and evidence limits when quoting or summarizing it.

## Safeguard Durability in Open Weight Models

- Slug: open-weight-safeguard-durability
- Canonical URL: https://latentmindsinstitute.com/research/open-weight-safeguard-durability/
- API URL: https://latentmindsinstitute.com/api/research-objects/open-weight-safeguard-durability.json
- Type: Research statement (statement)
- Lifecycle: in-preparation
- Epistemic status: design-only
- Status statement: In preparation; design, cost measure and controls specified; no attack run, no model evaluated, no result
- Programme: deployment
- Date: 2026-07-28
- Version: v0.1
- Authors: Latent Minds Institute
- Research question: When a model's weights are public, how much adversarial effort does each published safeguard actually impose, and how much of that cost survives a modest fine tuning budget?
- Evidence: none; this is a design statement published before any run
- Provenance: Published in advance of results so the question, the cost measure and the controls are on the record before any finding is
- Methods: adversarial red teaming across input, decoding, weight modification and composed surfaces (planned); attack cost curves against a fixed behaviour set (planned); capability controls separating a defeated safeguard from a degraded model (planned); attacker effort controls so null results read as bounded search (planned)
- Models: published open weight models (planned)
- Code: not publicly linked
- Data: not publicly linked
- Related slugs: model-entrenchment

## Preemptive Detection of Agentic Misalignment, and Its Shelf Life

- Slug: agentic-misalignment-detection
- Canonical URL: https://latentmindsinstitute.com/research/agentic-misalignment-detection/
- API URL: https://latentmindsinstitute.com/api/research-objects/agentic-misalignment-detection.json
- Type: Research statement (statement)
- Lifecycle: in-preparation
- Epistemic status: design-only
- Status statement: In preparation; two adjacent questions and a shared design specified; no probe fitted, no intervention run, no drift measured
- Programme: cognition
- Date: 2026-07-28
- Version: v0.1
- Authors: Latent Minds Institute
- Research question: Can representation engineering detect agentic misalignment from internal state before the action is taken, and does the self representation such a detector reads remain stable as reinforcement learning horizons lengthen?
- Evidence: none; this is a design statement published before any run
- Provenance: The two questions are posed together because the second sets the shelf life of any detector built under the first
- Methods: readouts over a fixed set of agentic settings, taken before the action is emitted (planned); intervention to separate a used internal variable from a merely decodable one (planned); re-evaluation across a reinforcement learning checkpoint series at matched behavioural performance (planned); surface form, capability and lexical baseline controls (planned)
- Models: open weight models with reinforcement learning checkpoint series (planned)
- Code: not publicly linked
- Data: not publicly linked
- Related slugs: evaluation-state, latent-signatures-strategic-games

## Construct map

- Slug: construct-map
- Canonical URL: https://latentmindsinstitute.com/research/construct-map/
- API URL: https://latentmindsinstitute.com/api/research-objects/construct-map.json
- Type: Methods note (note)
- Lifecycle: n/a
- Epistemic status: methods-note
- Status statement: Methods documentation; no results, no numbers
- Programme: cognition
- Date: 2026-07-18
- Version: v1
- Authors: Latent Minds Institute
- Research question: For each strategic game family, what construct is measured, what control does a reading require, and which claim does the evidence never license?
- Evidence: documentation of method only; contains no measurements and states none
- Provenance: Written for the latent arena game families behind Latent Signatures in Strategic Games
- Methods: construct specification per game family; required controls stated before code; explicit never claim boundaries
- Models: none stated
- Code: not publicly linked
- Data: not publicly linked
- Related slugs: latent-signatures-strategic-games

## Latent Signatures in Strategic Games

- Slug: latent-signatures-strategic-games
- Canonical URL: https://latentmindsinstitute.com/papers/latent-signatures-strategic-games/
- API URL: https://latentmindsinstitute.com/api/research-objects/latent-signatures-strategic-games.json
- Type: Working paper (working-paper)
- Lifecycle: pilot
- Epistemic status: research-programme
- Status statement: Research programme; behavioural fleet measured across six models in negotiation and four in social deduction; latent analysis proposed, none run
- Programme: cognition
- Date: 2026-07-11
- Version: v0.2
- Authors: Latent Minds Institute
- Research question: Do deception, opponent-modelling, and evaluation-awareness signatures in an open model's activations predict and cause its moves in strategic games?
- Evidence: research programme with three signature hypotheses and a three-stage proposed pipeline; the latent-arena harness is built and tested, the behavioural fleet is measured (negotiation self-play across six models, social deduction with four model tables), no activations captured
- Provenance: Built for Latent Minds Institute; the harness is the latent-arena repository, being prepared for public release and available on request until then
- Methods: self-hosted game environments; behavioural self-play (measured); instrumented self-play (proposed); signature probes (proposed); steering during play (proposed)
- Models: open-weight models (proposed)
- Code: not publicly linked
- Data: not publicly linked
- Related slugs: evaluation-state, model-entrenchment

## Circuit Traces: Attribution Graphs for Open Models

- Slug: circuit-traces
- Canonical URL: https://latentmindsinstitute.com/instruments/circuit-traces/
- API URL: https://latentmindsinstitute.com/api/research-objects/circuit-traces.json
- Type: Instrument (instrument)
- Lifecycle: instrument
- Epistemic status: instrument
- Status statement: Live attribution graphs on Gemma 2 2B; hypotheses to confirm with interventions, stated as such
- Programme: computation
- Date: 2026-07-11
- Version: v1
- Authors: Latent Minds Institute
- Research question: Which features caused which, from prompt to prediction, and does the causal story survive comparison across tasks, models, and methods?
- Evidence: real causal graphs served live; replacement-model caveat stated in interface
- Provenance: Method: Ameisen et al. / Lindsey et al. (2025), Transformer Circuits; implementation: safety-research/circuit-tracer; hosting: Neuronpedia
- Methods: attribution graphs on transcoder replacement models; curated case-study reading guides; validated graph loading
- Models: Gemma 2 2B (Gemma Scope transcoders)
- Code: https://github.com/safety-research/circuit-tracer
- Data: not publicly linked
- Related slugs: latent-observatory, transformer-explainer

## Latent Observatory: SAE Feature Explorer

- Slug: latent-observatory
- Canonical URL: https://latentmindsinstitute.com/instruments/latent-observatory/
- API URL: https://latentmindsinstitute.com/api/research-objects/latent-observatory.json
- Type: Instrument (instrument)
- Lifecycle: instrument
- Epistemic status: instrument
- Status statement: Live model data via the open Neuronpedia API; explanations labelled as hypotheses
- Programme: representations
- Date: 2026-07-11
- Version: v1
- Authors: Latent Minds Institute
- Research question: What has a sparse autoencoder actually learned about a model, and does the auto-interp story survive contact with the activation evidence?
- Evidence: model data from the open Neuronpedia API
- Provenance: Built by Latent Minds Institute on the open Neuronpedia API (MIT)
- Methods: semantic search over auto-interp explanations; activation-record inspection; logit-effect readout; decoder-space nearest neighbours
- Models: GPT-2 small (RES-JB); Gemma 2 2B (Gemma Scope res 16k)
- Code: not publicly linked
- Data: https://www.neuronpedia.org/api-doc
- Related slugs: interpretability-map, evaluation-state

## Evaluation State in Language Models

- Slug: evaluation-state
- Canonical URL: https://latentmindsinstitute.com/papers/evaluation-state/
- API URL: https://latentmindsinstitute.com/api/research-objects/evaluation-state.json
- Type: Working paper (working-paper)
- Lifecycle: proposal
- Epistemic status: conceptual-framework
- Status statement: Research programme · five hypotheses, experiments proposed, none run
- Programme: cognition
- Date: 2026-07-10
- Version: v0.1
- Authors: Latent Minds Institute
- Research question: What evidence would show that a model represents evaluation as a shared internal state rather than reacting to surface cues?
- Evidence: literature synthesis and experimental framework; no original empirical results
- Provenance: not stated in registry
- Methods: layer-wise probes vs lexical baselines (proposed); cross-family and cross-task transfer (proposed); activation steering with capability controls (proposed); Jacobian-lens readouts (proposed)
- Models: open-weight instruction-tuned models (proposed)
- Code: not publicly linked
- Data: not publicly linked
- Related slugs: model-entrenchment, interpretability-map

## Model Entrenchment: Why Useful AI Systems Become Difficult to Remove

- Slug: model-entrenchment
- Canonical URL: https://latentmindsinstitute.com/papers/model-entrenchment/
- API URL: https://latentmindsinstitute.com/api/research-objects/model-entrenchment.json
- Type: Working paper (working-paper)
- Lifecycle: proposal
- Epistemic status: conceptual-framework
- Status statement: Conceptual framework · experiments proposed, none run
- Programme: deployment
- Date: 2026-07-01
- Version: v1.1
- Authors: Latent Minds Institute
- Research question: When 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?
- Evidence: conceptual framework with falsification criteria; every empirical claim is a prediction
- Provenance: not stated in registry
- Methods: linear probes (proposed); activation steering (proposed); counterfactual environments (proposed); behavioural evaluations (proposed); case coding (proposed)
- Models: open-weight models (proposed)
- Code: not publicly linked
- Data: not publicly linked
- Related slugs: interpretability-map, gpu-credit-underwriting, gpu-forward-market

## The Interpretability Map

- Slug: interpretability-map
- Canonical URL: https://latentmindsinstitute.com/instruments/interpretability-map/
- API URL: https://latentmindsinstitute.com/api/research-objects/interpretability-map.json
- Type: Research map · instrument (research-map)
- Lifecycle: instrument
- Epistemic status: literature-synthesis
- Status statement: Literature synthesis · sources cited per node
- Programme: representations
- Date: 2026-07-01
- Version: v1
- Authors: Latent Minds Institute
- Research question: What has mechanistic interpretability actually established, in what order, with what dependencies, and what should a new researcher do this week?
- Evidence: synthesis of primary sources; contains no original experimental claims
- Provenance: not stated in registry
- Methods: literature dependency graph; chronology; tiered reading pathway; runnable TransformerLens/SAELens protocols; model-access matrix; open problems as experiments
- Models: GPT-2 small; Gemma 2 2B (protocols)
- Code: not publicly linked
- Data: /instruments/interpretability-map/atlas.json
- Related slugs: model-entrenchment, transformer-explainer

## Transformer Explainer: GPT-2 Live in the Browser

- Slug: transformer-explainer
- Canonical URL: https://latentmindsinstitute.com/instruments/transformer-explainer/
- API URL: https://latentmindsinstitute.com/api/research-objects/transformer-explainer.json
- Type: Instrument · adaptation (instrument)
- Lifecycle: instrument
- Epistemic status: adaptation
- Status statement: Adapted instrument · original by Cho et al., Georgia Tech Polo Club
- Programme: computation
- Date: 2026-07-10
- Version: v1
- Authors: Latent Minds Institute
- Research question: What does a transformer's forward pass actually compute, step by step, on real input?
- Evidence: pedagogical instrument running a real model; the 2026 commentary is editorial, not original research
- Provenance: Adaptation of poloclub/transformer-explainer (MIT); previously hosted as Transformer Visualiser at mo3.ca
- Methods: ONNX GPT-2 (small) inference in-browser; attention map inspection; temperature and sampling controls
- Models: GPT-2 small (124M)
- Code: https://github.com/poloclub/transformer-explainer
- Data: not publicly linked
- Related slugs: interpretability-map

## Underwriting the Machine: A Field Guide to GPU Credit Risk

- Slug: gpu-credit-underwriting
- Canonical URL: https://latentmindsinstitute.com/papers/gpu-credit-underwriting/
- API URL: https://latentmindsinstitute.com/api/research-objects/gpu-credit-underwriting.json
- Type: Working paper · economics strand (working-paper)
- Lifecycle: proposal
- Epistemic status: conceptual-framework
- Status statement: Analytical framework with illustrative models · not investment advice
- Programme: deployment
- Date: 2026-07-11
- Version: v1
- Authors: Latent Minds Institute
- Research question: How should lenders price credit risk on GPU-collateralised debt, and what does that market structure imply for compute dependence?
- Evidence: analytical framework with worked illustrative numbers; no proprietary deal data
- Provenance: not stated in registry
- Methods: credit risk decomposition; DSCR/LTV sizing models; residual-value analysis; scenario stress
- Models: none stated
- Code: not publicly linked
- Data: not publicly linked
- Related slugs: gpu-forward-market, model-entrenchment

## A Forward Market for GPU Compute

- Slug: gpu-forward-market
- Canonical URL: https://latentmindsinstitute.com/papers/gpu-forward-market/
- API URL: https://latentmindsinstitute.com/api/research-objects/gpu-forward-market.json
- Type: Working paper · economics strand (working-paper)
- Lifecycle: proposal
- Epistemic status: conceptual-framework
- Status statement: Market-design proposal · conceptual
- Programme: deployment
- Date: 2026-07-01
- Version: v1
- Authors: Latent Minds Institute
- Research question: What would a functioning forward market for GPU compute look like, and what would it change about the economics of AI deployment?
- Evidence: market-design proposal; no live market data beyond cited public sources
- Provenance: not stated in registry
- Methods: market design; contract specification; term-structure analysis
- Models: none stated
- Code: not publicly linked
- Data: not publicly linked
- Related slugs: gpu-credit-underwriting, model-entrenchment