Publications

Field Notes

Articles chronicling the development of ARM — design decisions, findings as they emerged from the data, and the honest-scope discipline that defines the project.

Published on LinkedIn ↗ as they are written. Articles 1–3+ are live.

3June 2026
Live on LinkedIn

Failure mode reported itself clean

Three meshes. Same question. Opposite unanimous consensus. This note walks through the CFAA zero-day run: what we observed, what we expected, why the directional opposition is the most robust signal, and what it implies for practitioners building on single providers.

FingerprintingCross-provider testingCFAA
Read on LinkedIn ↗
2May 2026
Live on LinkedIn

Which AI are you actually deploying?

Convergence is not agreement. This note introduces the Alignment Monoculture Problem: in same-model meshes, observed consensus can be indistinguishable from shared prior activation. Covers the carpentry analogy and the first evidence from runs 12–14.

Alignment MonocultureEpistemic drift
Read on LinkedIn ↗
1May 2026
Live on LinkedIn

AI systems agree with each other too much

The key insight from SCRUMtious to ARM: if you want to know whether peer exposure changed an agent's mind, you have to capture its uncontaminated position first. The two-round architecture and why it mirrors Global Workspace Theory.

Protocol designTwo-round architecture
Read on LinkedIn ↗
In the pipeline
·The Gamma-flip failure mode — 10+ confirmed cases, what they have in common, and how the v0.8 polarity gate catches them
·Role injection as a confound — the OT-001 vs. OT-002 clean pair and what it means for comparing runs
·Convergence ≠ agreement — three proofs from the run corpus that Jaccard similarity does not measure consensus
·v0.8 gate decoupling — the 11-trace battery and why two independent gates matter
arXiv

Preprint (pending)

The ARM whitepaper v4.0 (~15,300 words, 89 JSON traces in Appendix D) is in preparation for arXiv submission. Citation pass and delta_mismatch field verification in v0.7.1 exports are the remaining steps. The field notes series will link to the preprint when it is live.