An outsider built a working epistemic instrument.

It started with SCRUMtious— a multi-agent Agile management app. The problem was a practical one: downstream agents received conclusions but had no visibility into the reasoning that produced them. They couldn't audit the logic, only accept or reject the output.

The fix was to pass full reasoning traces between agents — not summaries, not conclusions, but the complete chain of thought including assumptions, discarded paths, and the challenge surface. A downstream agent receiving a trace could see why the upstream agent concluded what it did, and where that reasoning was vulnerable.

From there, the isolation-then-deliberation insight followed: if you want to know whether peer exposure changedan agent's mind, you have to capture its uncontaminated position first. Measure before sharing. That's Round 1.

The design mirrors Global Workspace Theory: specialized parallel processors that broadcast to a shared workspace, with integration happening after isolation, not during. Agents that are allowed to share before they reason aren't deliberating — they're conforming.

ARM didn't set out to find model-level epistemic fingerprinting. It was looking for drift patterns. The CFAA finding emerged from running the same question through three provider meshes and noticing that the results weren't just different — they were directionally opposite, with each mesh reporting high-confidence unanimity.

That's the uncomfortable part. It wasn't three agents disagreeing. It was three consensus systems disagreeing, each with the signature of reliability. A practitioner using any single provider would have gotten a confident, unanimous answer — and had no signal that an isomorphic system on a different provider would give the opposite answer with equal confidence.

ARM made that visible. That's the whole thesis.

Erik's working principle is credible over impressive. The honest-scope discipline — stating every confound, flagging every n=1, calling prototypes prototypes — is not hedging. It's the instrument doing what it should: not overclaiming.

ARM is a smoke detector, not a sprinkler system. Through v0.7.1 it logged failure modes; v0.8 begins to catch them at the gate. The roadmap to a sprinkler is visible in the data. The work to get there is honest.