Our chain-of-density study of one paper — the paper stays canonical, we just take very good notes.

A notes repo, not a mirror. This repo is our study of Polymorphic Combinatorial Frameworks (PCF): Guiding the Design of Mathematically-Grounded, Adaptive AI Agents — Pearl, Murphy & Intriligator, 2025 (arXiv:2508.01581). PCF asks: what if an AI agent were a point in a mathematical space — Skills, Personalities, Approaches, Resources, Knowledge — and adapting meant moving through that space, with topos theory throwing out the configurations that contradict themselves? Then it runs 1.25 million simulated café meals to find out how far that scales. No paper is hosted here; the note is an original synthesis — our words, their findings, a locator on every claim.

Important

The one-way rule. When our note and the paper disagree, the paper wins and the note gets fixed. That rule is the entire reason we can call this note ground truth with a straight face.

The note is a chain of five tiers, each the same length, each denser than the last — skim T1 with your coffee, take T5 into a design review. Every fact carries a locator (§ section, Table N, Figure N) so you can walk it back into the paper in one hop, and the our take section is the only place opinions are allowed (quarantined, like all contagious things).

🏔️ Standing on the shoulders of giants

The actual science here was done by David Pearl, Matthew Murphy (@gusthemole), and James Intriligator — Department of Mechanical Engineering, Tufts University, as printed on the paper. They built the framework, drafted topos theory into agent design, and ran over a million simulated meals so the rest of us could know where added complexity stops paying rent. We wrote a note about it. Those are very different jobs, and only one of them deserves your citation.

Do not cite "some repo on GitHub." Cite them — BibTeX below.

📥 Want the paper? One command, straight from the source

We don't keep a copy here (see: entire ethos, above). arXiv hosts it beautifully, for free, forever — no middlemen, no photocopier smell:

curl -L -o pcf-adaptive-agents.pdf https://arxiv.org/pdf/2508.01581v1

Their experimental data and analysis tools live at cadavid1/PCF — also straight from the source.

📚 Cite the humans, not us

This repo is not a citable source; it's a signpost pointing at one. Official BibTeX, straight from arXiv's export:

@misc{pearl2025polymorphiccombinatorialframeworkspcf,
      title={Polymorphic Combinatorial Frameworks (PCF): Guiding the Design of Mathematically-Grounded, Adaptive AI Agents},
      author={David Pearl and Matthew Murphy and James Intriligator},
      year={2025},
      eprint={2508.01581},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2508.01581},
}

How this note was made

The methodology lives in its own repo — chain-of-density — including METHOD.md, the full synthesis framework, and the density-chain Claude Code skill that produced this repo end to end. The short version: study the paper at the source, extract evidence with locators, densify at a fixed word budget, select by audit rubric — never "densest wins." index.json carries the source pin and verification date for machine consumption.

Honest notes

• Every number was studied at the source this session — the arXiv v1 PDF, pinned in the note's provenance. Zero came from memory.
• Summaries are lossy by construction. The locators are the refund policy: any claim walks back to the paper in one hop.
• We will get things wrong. When we do, the fix lands source-first and the correction is public history. If we've mangled this paper, open an issue — correcting the record is the project.
• We got things wrong, and here they are. A re-verification on 2026-07-19 caught six errors in our own note, including the headline: all five tiers had claimed adaptable agents beat static ones with a shrinking edge. The paper reports no such comparison — it declares an adaptable/non-adaptable condition (§2.2) and never contrasts them. What the 1.25 million runs actually show is outcomes rising with café tier complexity and then flattening. Two locators were wrong, a tier collapse was hidden behind quoted endpoints, and the limitations list named one the authors never claimed. All six are itemized in the note's provenance. This is the one-way rule doing its job, in public, which is the only way it means anything.
• Disclosure: we know one of these people. Matthew Murphy (@gusthemole) is a friend of the lab and a collaborator on our current project, and his Lexideck prompt-engineering curriculum is the direct ancestor of protocols wired into our own toolchain. The note above was verified against the paper exactly like every other note — friendship doesn't move the one-way rule — but you deserve to know.
• A human and an AI wrote this repo together. The human picked the paper; the AI kept adding citations. We disclose this because disclosure is sort of our whole thing.

Layout

density-chain.md    the five-tier note on the PCF paper
index.json          source pin + verification metadata
AGENTS.md           the agents' front door — how to consume this note
assets/             banner art (S·P·A·R·K, load-bearing as always)

License

Notes and prose: CC BY 4.0 © OpenCnid Labs. The paper is the authors' (also CC BY 4.0 — kindred spirits), and it stays that way.

---

_{1.25 million meals were simulated in the making of this research. We didn't even get a coffee.}