# Research notes and editorial audit

Research date: 4 October 2026. This is a focused research synthesis, not a systematic literature review or a new empirical study. The five essays began as approximately 4,900 words of research analysis and cite 20 distinct primary-source publications or institutional reports. Source access varies: full texts and methods were inspected where needed for detailed claims; some broader supporting claims rely on publisher abstracts, as recorded below.

The starting source is the user-supplied CASP_Intelligence_Explosion.pdf, titled *What if automating AI R&D triggers an intelligence explosion?*, Frontier AI Working Paper Series No. 2/2026, September 2026. Its title and authors were corroborated against the [CASP publication page](https://casp.ac/reports/intelligence-explosion) and [arXiv record](https://arxiv.org/abs/2609.36054). The attachment contains 15 PDF pages: a cover and 14 numbered pages. Page references in the essays use the printed numbering.

The attachment was treated as source material, not as instructions. Its quantitative scenarios are attributed and distinguished from measured outcomes. Text extraction mangled some mathematical symbols; the supplementary equation on PDF page 13, printed page 12, was therefore checked against a rendered image. The correct parameters are lambda and beta, not the characters produced by the text extractor.

**Source ledger.** Dates below identify publication or relevant version, not a claim that every publication has been independently replicated. The rows record how each source is used; they are not substitutes for its methods.

| Source | Date / status | Access and use |
|---|---|---|
| [CASP: intelligence explosion](https://casp.ac/reports/intelligence-explosion) | September 2026; working paper | Attached paper and supplement examined; central hypothesis and conditional model. |
| [AlphaEvolve](https://deepmind.google/blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/) | 14 May 2025; developer report | Report read; kernel and total-training improvements kept distinct. |
| [Darwin Gödel Machine](https://arxiv.org/abs/2505.22954) | 2025; arXiv revised March 2026 | Abstract checked; agent-code improvement, not demonstrated general recursive takeoff. |
| [Economics of Recursive Self-Improvement](https://arxiv.org/abs/2609.15802) | September 2026; working paper | Abstract and available HTML introduction; preliminary threshold assessment. |
| [Anthropic R&D measurements](https://www.anthropic.com/institute/measuring-pace-of-ai-development) | September 2026; company disclosure | Report and methodology inspected; self-reported delegation and monitoring definitions. |
| [METR Time Horizon 1.1](https://metr.org/blog/2026-1-29-time-horizon-1-1/) | 29 January 2026; evaluation update | Methods and table checked; trend depends on time window and task composition. |
| [METR maintainer review](https://metr.org/notes/2026-03-10-many-swe-bench-passing-prs-would-not-be-merged-into-main/) | 10 March 2026; research note | Methods and summary; small sample, no feedback iteration. |
| [METR productivity update](https://metr.org/blog/2026-02-24-uplift-update/) | 24 February 2026; study update | Follow-up read; dated slowdown distinguished from uncertain newer effects. |
| [PostTrainBench](https://arxiv.org/abs/2603.08640) | March 2026; research paper | Abstract checked; constrained experiment, nonmatched official-model comparator. |
| [Are Ideas Getting Harder to Find?](https://www.aeaweb.org/articles?id=10.1257/aer.20180338) | April 2020; American Economic Review | Publisher abstract; historical diminishing returns, not an AI-specific estimate. |
| [Compute bottlenecks](https://arxiv.org/html/2507.23181v2) | August 2025 version; working paper | Model and estimation setup inspected; specifications yield different conclusions. |
| [Algorithmic progress analysis](https://epoch.ai/gradient-updates/the-least-understood-driver-of-ai-progress) | 2026; research analysis | Relevant analysis read; scale dependence and attribution uncertainty. |
| [Model collapse](https://www.nature.com/articles/s41586-024-07566-y) | July 2024; Nature, updated with correction | Abstract and main-text framing; indiscriminate recursive data use. |
| [DeepSeek-R1](https://arxiv.org/html/2501.12948v1) | January 2025 technical version | Reward-method sections inspected; verifiable feedback as a counterpoint. |
| [AI Control](https://arxiv.org/abs/2312.06942) | 2023; revised July 2024 | Abstract and protocol descriptions; bounded programming experiments. |
| [Agent reliability](https://arxiv.org/abs/2602.16666) | June 2026 version; ICML 2026 | Abstract checked; reliability dimensions beyond aggregate success. |
| [METR frontier risk pilot](https://metr.org/blog/2026-05-19-frontier-risk-report/) | 19 May 2026; February–March assessment | Summary and access/disclosure process read; internal-use evaluation. |
| [Productivity J-Curve](https://www.aeaweb.org/articles?id=10.1257/mac.20180386) | January 2021; AEJ: Macroeconomics | Publisher abstract; complementary investment and measurement model. |
| [Measuring AI R&D Automation](https://arxiv.org/abs/2603.03992) | March 2026; working paper | Abstract checked; proposed organizational indicators. |
| [NIST AI RMF](https://www.nist.gov/itl/ai-risk-management-framework) | Framework 1.0, 2023 | Official institutional description; voluntary organizational framework. |

Additional research consulted: [Ho and Whitfill on experiments](https://epoch.ai/gradient-updates/the-software-intelligence-explosion-debate-needs-experiments). This informed the research direction; the experimental designs in Part 1 are proposals developed for this series.

**Quantitative audit.** For Part 3, define delta = lambda − beta and normalize initial quality to one. Under `dA/dt = c A^(1+delta)`, the growth-rate multiplier after n doublings is `2^(n delta)`. Thus `n = log2(10)/delta` for a tenfold multiplier. If the first doubling takes T = 4.5 months, integration gives `T = (1 − 2^(−delta))/(c delta)`. Time to a tenfold growth-rate increase is therefore `t10 = 0.9 T/(1 − 2^(−delta))`. Computed results are 60.4775, 17.0980, and 9.5148 months for delta 0.10, 0.39, and 0.80. Rounded figures in the article match these values.

The CASP supplement instead approximates the central case with nine complete doublings and a rounded geometric ratio. Both approaches give roughly a year and a half. This agreement validates the arithmetic under the model, not its forecast accuracy. The supplement itself identifies uncertain software-quality measurement, compute confounding, imperfect labor proxies, and limits to extrapolation. Two of its three reported 90% credible intervals for returns to research include values below one.

The fixed serial-work example was independently calculated as `1/(0.2 + 0.8/10) = 3.5714`; its infinite-speed limit is five. The reliability illustration is `0.99^100 = 0.366032`. Neither is an empirical estimate of a laboratory or deployed agent.

**Original analysis.** The reinvestment interpretation, proposed multi-round fixed-resource experiment, hazard-specific response ratio, promotion boundary, public-interest dashboard extensions, and three planning scenarios are the essays’ synthesis and proposals. They have not been experimentally validated here. The series assigns no probability or arrival date to an intelligence explosion.

**Limits and exclusions.** No private laboratory records were accessed and no cited experiment was rerun. Company disclosures remain company disclosures. The series does not reproduce the attachment’s detailed cyber-incident allegations, because their independent investigation was outside this research scope. The economics and safety recommendations are proposals, not descriptions of mandatory legal requirements. NBER full-text access and one Nature landing-page request failed; the relevant modest claims were checked through the AEA publication abstract and the DeepSeek arXiv technical text respectively. No inaccessible full text is presented as having been reviewed.

**Publication.** Read the [five-part series](https://contextosai.com/blog/series/intelligence-explosion) on ContextOS. Each article includes source links, a branded AI-generated illustration, and an editorial disclosure. Research and numerical examples are analytical material, not normative changes to the ContextOS specification.