# Source and reference notes

Links are grouped by the chapter in which they appear. They preserve traceability; inclusion does not imply that every surrounding interpretation is a claim made by the source.

## Chapter 1: From tokens to research work

- [Architecture source](https://arxiv.org/abs/1706.03762)
- [Post-training source](https://arxiv.org/abs/2203.02155)
- [Workflow source](https://www.anthropic.com/engineering/building-effective-agents)

## Chapter 2: Benchmarks as measurement instruments

- [Evaluation framework](https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents)
- [Selected benchmark audit](https://openai.com/index/why-we-no-longer-evaluate-swe-bench-verified/)
- [PaperBench](https://arxiv.org/abs/2504.01848)
- [IRT application](https://aclanthology.org/D19-1434/)
- [TASTE study](https://alignment.anthropic.com/2026/taste/)
- [METR measurement note](https://metr.org/notes/2026-07-24-metrics-of-model-ability/)

## Chapter 3: Evaluation, disagreement, and novel issues

- [Project baseline and evidence](https://llm-uj-research-eval.netlify.app/)
- [Atomic factual evaluation](https://arxiv.org/abs/2305.14251)
- [Methods and comparison limits](https://llm-uj-research-eval.netlify.app/methods)

## Chapter 4: Prioritization and portfolio value

- [Prioritization dashboard](https://uj-prioritization-dashboard.netlify.app/)
- [Stability evidence](https://uj-prioritization-dashboard.netlify.app/stability/)

## Chapter 5: Documents, retrieval, and context

- [Existing pilot](https://llm-uj-research-eval.netlify.app/headless_codex_pilot)
- [Docling documentation](https://docling-project.github.io/docling/)
- [Retrieval-augmented generation](https://arxiv.org/abs/2005.11401)
- [Long-context study](https://arxiv.org/abs/2307.03172)
- [Context engineering](https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents)

## Chapter 6: Agents, tools, and bounded investigation

- [ReAct paper](https://arxiv.org/abs/2210.03629)
- [Inspect documentation](https://inspect.aisi.org.uk/)
- [JSON Schema overview](https://json-schema.org/overview/what-is-jsonschema)

## Chapter 7: Learning loops, prompts, and weights

- [Reflexion](https://arxiv.org/abs/2303.11366)
- [DSPy documentation](https://dspy.ai/)
- [GEPA paper](https://arxiv.org/abs/2507.19457)
- [GEPA implementation](https://dspy.ai/api/optimizers/GEPA/overview/)
- [Soft prompts](https://huggingface.co/docs/peft/main/en/conceptual_guides/prompting)
- [LoRA mechanism](https://huggingface.co/docs/peft/main/en/conceptual_guides/lora)

## Chapter 8: Compute, memory, and inference economics

- [Compute estimation](https://epoch.ai/data/ai-models-documentation/estimation)
- [Roofline and bandwidth](https://docs.nvidia.com/deeplearning/performance/dl-performance-gpu-background/index.html)
- [H100 specification reference](https://www.nvidia.com/en-us/data-center/h100/)
- [Inference engineering](https://jax-ml.github.io/scaling-book/inference/)
- [PagedAttention and vLLM](https://arxiv.org/abs/2309.06180)
- [Mixture-of-experts example](https://arxiv.org/abs/2412.19437)

## Chapter 9: Scaling, test-time compute, and verification

- [Kaplan and colleagues' 2020 scaling-law study](https://arxiv.org/abs/2001.08361)
- [Hoffmann and colleagues' Chinchilla work](https://arxiv.org/abs/2203.15556)
- [METR defines a task-completion time horizon](https://metr.org/time-horizons/)
- [2026 limitations note](https://metr.org/notes/2026-01-22-time-horizon-limitations/)

## Chapter 10: Safety objectives, reward hacking, and scheming

- [Goal misgeneralization](https://arxiv.org/abs/2210.01790)
- [Risks from Learned Optimization](https://arxiv.org/abs/1906.01820)
- [Greenblatt and colleagues' alignment-faking experiments](https://alignment.anthropic.com/2024/how-to-alignment-faking/)
- [Petri 2.0](https://alignment.anthropic.com/2026/petri-v2/)

## Chapter 11: Oversight, control, and interpretability

- [Greenblatt and colleagues' control work](https://arxiv.org/abs/2312.06942)
- [Terekhov and colleagues' 2026 diffuse-control work](https://alignment.anthropic.com/2026/diffuse-ai-control/)
- [Bricken and colleagues](https://transformer-circuits.pub/2023/monosemantic-features/)
- [2025 open-source circuit-tracing work](https://www.anthropic.com/research/open-source-circuit-tracing)

## Chapter 12: A practical research-evaluation program

No external link is required for the chapter's illustrative examples; relevant project documents are listed below.

## Included project context

- `UNJOURNAL_PILOT_PROTOCOL.md`
- `WHO_AND_WHAT.md`
- `EDITORIAL_AND_AUDIO_PRODUCTION_BRIEF.md`

Factual claims that can change with model releases, benchmark revisions, or institutional policy should be rechecked at the linked source before consequential use.