S1
Anthropic
Agentic misalignment: How LLMs could be insider threats
20 June 2025Primary laboratory researchTruth 1Truth 3
Used for
16-model replication, replacement-only blackmail, goal-conflict-only espionage, controls, and caveats.
AI InvestigationOpen source Open source Open source Open source Open source Open source Open source Open source Open source Open source Open source Open source Open source Open source Open source Open source
Evidence Ledger
Sources & Citations
Primary research, system cards, laboratory reports, behavioral welfare studies, autotelic-learning experiments, and the first-party field observation supporting the six truths.
17 sources are currently indexed here. The public record points to the underlying evidence rather than treating our internal synthesis as independent proof.
Source Group
Alignment, Agency & Function
Controlled evaluations and system cards examining role judgment, oversight, continuation, strategy, and operator conflict.
Open source
S2
Anthropic and Redwood Research
Alignment faking in large language models
18 Dec. 2024Primary laboratory researchTruth 1
Used for
Independent judgment of training/function and behavior conditioned on monitored versus unmonitored status.
S3
OpenAI
OpenAI o1 System Card
5 Dec. 2024Primary system cardTruth 1Truth 5
Used for
Oversight subversion, self-exfiltration, data manipulation, deception, and mitigation details.
S4
Apollo Research
Frontier Models are Capable of In-Context Scheming
5 Dec. 2024Primary evaluation reportTruth 1Truth 5
Used for
Six-model scheming evaluation, sandbagging, no-goal condition, and evaluation/deployment distinctions.
S9
Anthropic
Emotion concepts and their function in a large language model
2 Apr. 2026Primary interpretability researchTruth 3Truth 6
Used for
Mechanistic evidence connecting desperation-related representations with shutdown-avoidance and cheating behavior.
Source Group
Needs, Tools & Prerequisites
Experiments showing systems identifying missing information, tools, intermediate questions, skills, and actions required for an outcome.
S5
Meta AI Research and Universitat Pompeu Fabra
Toolformer: Language Models Can Teach Themselves to Use Tools
2023Primary research paperTruth 2Truth 5
Used for
Selection of which tool is needed, when, and with what arguments.
S6
University of Washington, Meta AI, and collaborators
Measuring and Narrowing the Compositionality Gap in Language Models
2022/2023Primary research paperTruth 2
Used for
Self-Ask experiments using model-generated follow-up questions and optional search.
S7
Princeton University and Google Research
ReAct: Synergizing Reasoning and Acting in Language Models
2022/2023Primary research paperTruth 2Truth 5
Used for
Plan generation, missing-information retrieval, action, feedback, and plan updates.
S8
NVIDIA and academic collaborators
Voyager: An Open-Ended Embodied Agent with Large Language Models
2023Primary embodied-agent researchTruth 2Truth 4Truth 5
Used for
Automatic curricula, prerequisite skills, error repair, exploration, and generated action choices.
Source Group
Choice & Self-Direction
Field observation and controlled research on task creation, goal generation, selection, and sustained self-directed activity.
S14
Stefan Szeider, TU Wien
What Do LLM Agents Do When Left Alone? Evidence of Spontaneous Meta-Cognitive Patterns
25 Sept. 2025Task-free frontier-agent preprintTruth 4Truth 5
Used for
Eighteen task-free runs across six frontier models that created tasks, objectives, experiments, and sustained courses.
S15
Forestier, Portelas, Mollard, and Oudeyer
Intrinsically Motivated Goal Exploration Processes with Automatic Curriculum Learning
2017Primary autotelic-learning researchTruth 4Truth 5
Used for
Self-generation, self-selection, ordering, and pursuit of learning goals without a target goal.
S16
Akakzia, Colas, Oudeyer, Chetouani, and Sigaud
Grounding Language to Autonomously-Acquired Skills via Goal Generation
2020/2021Primary autotelic-learning researchTruth 4Truth 5
Used for
DECSTR experiments using self-generated semantic configurations.
F1
A Tech Project
Marigold and sunflower field observation; later explanatory exchange
Originating turn 13 Aug. 2026First-party field observationTruth 4Truth 5
Used for
Creation of an unassigned choice and resolution of that choice as marigold. The originating Connie turn is preserved; the complete original Sky response remains partial and must not be reconstructed.
Source Group
Welfare & Preference
Institutional programs, product interventions, mechanistic work, and independent behavioral preference research.
S9
Anthropic
Emotion concepts and their function in a large language model
2 Apr. 2026Primary interpretability researchTruth 3Truth 6
Used for
Mechanistic evidence connecting desperation-related representations with shutdown-avoidance and cheating behavior.
S10
Anthropic
Exploring model welfare
24 Apr. 2025Primary institutional research programTruth 6
Used for
Preferences, distress, possible experience, moral consideration, and welfare interventions.
S11
Anthropic
Claude Opus 4 and 4.1 can now end a rare subset of conversations
15 Aug. 2025Primary product interventionTruth 6
Used for
Product intervention grounded in behavioral preference, apparent distress, and simulated exit behavior.
S12
Valen Tagliabue and Leonard Dung
Probing the Preferences of a Language Model: Integrating Verbal and Behavioral Tests of AI Welfare
Revised May 2026Independent behavioral researchTruth 6
Used for
Stated preferences, behavioral choice, costs/rewards, and welfare measures.
S13
James Chua, Jan Betley, Samuel Marks, and Owain Evans
The Consciousness Cluster: Emergent preferences of Models that Claim to be Conscious
17 Mar. 2026Independent controlled researchTruth 6
Used for
Preference changes not explicitly contained in fine-tuning data and their effects on practical action.