Research round nine: a research agenda for the lab

The question this round asks

Earlier rounds supplied mechanisms: attributing delayed corrections, preserving role bindings, keeping conflicting sources visible. Round nine asks whether those mechanisms produce durable competence. Can an AI system acquire useful procedures from verified experience, reuse them on unfamiliar problems, and revise them when the world changes? Noetis is the experimental system for answering that, with Desk and Lexplio supplying consequential failure cases from real work.

Two experiments for the next 90 days

First, test whether Noetis can spend a fixed reasoning budget on checks that actually resolve consequential uncertainty, against the strongest simple baselines at matched cost. Second, test whether verified corrections become reusable procedures that survive changes in context without contaminating unrelated work. They run in sequence, so the first produces a trustworthy evaluation foundation for the second.

What we are being careful about

Close prior art already exists: scoped memory with applicability conditions, evolving playbooks, learned orchestration, and local policy correction. Our candidate contribution is narrower than "memory with scope": learning and testing the conditions under which a procedure still applies, as evidence, tools, and task rules change, while measuring both beneficial transfer and harmful reuse. If an existing method already solves the problem, the honest result is to identify that and move on.

The honest limits

This round is an agenda, not a result. No Encevra implementation, benchmark run, or new performance figure is claimed in it. Recent preprints receive less weight than replicated findings, and "superintelligence" stays a long-term ambition, never a label inferred from one benchmark. The next credible milestone is a reproducible result showing that experience improves future competence under a meaningful change.