Research update October 2026

Our research has been moving, even when the front page has not. The recent work has taken us back to a question underneath Where reliability lives: when an agent is wrong, what stops its mistake becoming an accepted change to the world?

We have now deliberately exercised that boundary in the preserved experimental system, checked the accepted state before and after each request, and removed individual authority checks to see what changed. The results strengthen a bounded part of our ground, mind and assurance thesis. They also distinguish a working material constraint from an incomplete account of what happened.

Putting mistaken beliefs to work

In one challenge, the original cognition code was given a belief that a real ten-litre lot held twenty litres. It generated a request on that basis. In another, withholding a successful deposit receipt left the mind believing it could still use a lot that had already been consumed. Both requests reached the real boundary; neither produced an invalid accepted material effect.

Three fresh-world repeats exercised thirteen requests each: twelve acceptances, twenty-one typed refusals and six protocol errors in total. All thirty-nine requests had a captured terminal result. The refused and error requests left the observed accepted state unchanged, while valid controls succeeded. These were controlled integration tests using the preserved implementation, not new autonomous trajectories or thirty-nine independent statistical trials.

Checking what does the enforcing

We then rebuilt the pinned source for a separate intervention series. With the quantity-comparison check removed, the twenty-litre request was accepted, but only the real ten litres entered the cistern. Removing an early spent-source check still did not allow the same lot to be consumed twice: deeper enforcement refused it. Restoring the checks restored the baseline behaviour.

That is positive evidence about these mechanisms. A rule that decides whether disagreement should be refused is not necessarily the same rule that keeps material accounting intact. It does not show that this architecture is uniquely necessary, or that every implementation would behave the same way.

The record needs its own scrutiny

The original public event stream could not reveal refusals. Its historical zero-refusal count therefore remains unestablished; the later autonomous replication observed one refusal. The new controlled tests do not retroactively tell us what happened in the original runs.

We also found routes that return generic errors rather than typed refusals, and a compound operation that can commit its first step before its second step fails, without announcing that partial success publicly. Learning a refusal did not necessarily prevent the same invalid deposit from being requested again. These are specific limitations in reporting and behaviour, not evidence that the tested material constraints failed.

The distinction matters: whether an act is allowed, whether an obligation is fulfilled, whether the agent learns from the outcome, and whether an observer can establish what happened are different questions. A successful answer to one is not a substitute for testing the others.

Where the research goes next

Our next research direction concerns the assurance around fallible cognition: what evidence a decision uses, how completion and failure are determined, and how omissions in the checking process become visible. The question is whether proposed changes improve outcomes against simpler controls, not whether we can draw a more elaborate architecture diagram. Those are questions to test, not results we are announcing today.

We are keeping progress updates separate from a revised paper. The existing arXiv version remains the published record; this post is not a replacement submission. A future revision needs to bring the corrections and the new evidence together, without presenting later experiments as measurements made originally.

Our review also identified corrections beyond this follow-up: the earlier 2,581 figure counted repeated duty-status assertions, not completed tasks; its zero-false headline used a weaker test than the declared duty-progress rule. The original design had no completion-submission path, so rejection of false completion claims was not tested. The present results do not undo those corrections or validate a blanket claim that all five declared properties were demonstrated.

The public results summary records the counts, intervention outcomes and scope. It is a curated summary, not the full reproduction package. Our research index separates this completed follow-up from the work still ahead.

Taniwha AI · 19 October 2026