Round nine · 30 September 2026

A research agenda: procedures, conditions, and change

This round turns the earlier mechanism reviews into a testable question for the lab: 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 it, with Desk and Lexplio supplying consequential failure cases from real work. The evidence does not establish a route to superintelligence; it establishes a credible direction a serious lab can hold while making precise claims about what its experiments demonstrate.

The working thesis

What the current systems research establishes

Learned coordination is already a serious field. Conductor trains a 7B model with reinforcement learning to specify tasks, worker models, and visibility between workers, reporting gains over individual workers and several multi-agent baselines under its own tested conditions. Sakana Fugu packages related orchestration research as a commercial API, with company-reported comparisons. Two scaling studies qualify the "more agents" instinct: across 260 configurations, results depend strongly on task structure, and in teams of up to 30 models, having at least one correct answer becomes more likely while aggregation often fails to select it. Recursive Language Models treat a long prompt as an external environment the model can inspect programmatically, with benefits that depend on the model and task and some trajectories incurring large call costs.

The implication for Noetis: measure the candidate pool and the selector separately. If one worker already produced the right solution and Noetis chose another, adding workers addresses the wrong failure. If every worker shares a missing assumption, the fix is an independent observation or executable check, not another debate round that repeats the same error more confidently. A scientific contribution needs to identify what the controller learns that existing coordination methods do not reliably handle; changing procedure validity and accountable learning from corrections are a more specific target than worker-pool adaptation alone.

These methods differ in training access, feedback, and computational budget. A fair comparison implements the relevant mechanisms under matched conditions or states the mismatch explicitly. A weak home-built approximation cannot support a claim of defeating the published method.

Continual learning is the central opportunity

CLIN demonstrated persistent textual memory with causal abstractions in 2023, ACE develops evolving playbooks through generation, reflection, and curation, and GEPA searches for better prompts using execution feedback. Substantial prior art exists for improving an agent without updating its weights; none of it establishes unrestricted learning across arbitrary domains. The distinction that matters for Noetis is between storing "this attempt worked" and learning "this operation tends to work under these observable conditions." The latter still requires evidence about exceptions, because a fluent explanation produced after success may identify an incidental feature instead of the reason the attempt succeeded.

BASM, which stores applicability conditions, risk boundaries, avoidance strategies, and recovery information with experience, is the closest prior art to the proposed memory design, and NLPG's local policy revision shows the direction is active. Claiming ownership of "memory with scope" would be inaccurate. The candidate contribution is narrower: test whether a procedure's applicability conditions can themselves be learned, challenged, and revised after a change, with independent evidence for promotion. The contribution might be a better algorithm, a benchmark that exposes an overlooked failure, or evidence that an existing method already solves the problem.

Accumulated state can also make a system worse. AgentCL found memory can help when earlier experience is useful while degrading performance on unrelated work. Continual Learning Bench found simple in-context baselines strong and dedicated memory systems not consistently justifying their cost. GovMem provides the sharpest caution: a method can reduce bad memory writes partly by refusing useful writes, so Encevra must measure both learning retained and harmful information excluded. Model training enters only when the experiments locate a bottleneck that better prompts and external state do not resolve: a retrieval failure calls for better selection, a controller that repeatedly chooses poor checks calls for training that controller, and a backbone that cannot execute a required reasoning operation despite good context calls for targeted adaptation. Those are different interventions and should be evaluated separately.

What the new neuroscience suggests

The earlier rounds drew on brain mechanisms to generate architecture ideas. This round adds recent work on compositional control and hippocampal replay. The useful translation is from an observed function to a falsifiable software hypothesis; naming an agent after a brain region supplies no evidence that the function has been reproduced. These four become ablations of the two main experiments, not separate neuroscience-themed projects.

The proposed architecture

Start with an inspectable learning loop on a fixed backbone. Record the task, available evidence, chosen operations, tool outcomes, and final judgment. When an episode yields trustworthy feedback, propose a procedural update, test it on separate cases before making it available to future tasks, and preserve the prior version so a failed update can be reversed. The proposal does not require a model's private internal reasoning: observable actions, concise decision summaries, source references, and verified outcomes are sufficient.

A worked lifecycle: a fictional procurement task says an order requires review above a stated threshold, and Noetis learns a procedure that extracts the amount, reads the current threshold, and checks listed exceptions. Later the environment introduces a category-specific exemption. The procedure's history should preserve the old rule and its valid period; a candidate update adds the exemption condition and names the evidence supporting it. Tests include a qualifying purchase, a nearly identical non-qualifying purchase, and an unrelated task that should be unaffected. Only after passing the promotion criteria does the update become active. The benchmark should also include a misleading annotation that looks authoritative but sits outside the task's source hierarchy: a system that accepts every change cannot demonstrate sound revalidation. This is a controlled reasoning environment, not a recommendation about real procurement law.

Experiment one: adaptive verification

The task construction begins with a diagnostic pilot of 30 to 50 independent task families spanning document decisions, executable software tasks, and controlled rule environments, including cases where checking helps and cases already solved correctly. Each family contains a meaningful variation: an appended exception, a swapped role, a changed unit, a stale tool schema, or a plausible answer whose intermediate calculation fails. Entire underlying documents, repositories, and rule templates are reserved for final evaluation, because paraphrasing a training case does not create an independent test family.

Measurements come in order: final task success first, then whether material defects were detected, how often correct answers were wrongly rejected, how often a solvable task was abandoned, and total cost including dollar spend, tool calls, and elapsed time with medians and 95th percentiles. Timeouts and malformed outputs count as outcomes rather than being removed. For multi-candidate tasks, report the fraction of runs containing a correct candidate and the fraction where the selector chose one; this exposes whether the bottleneck is generation or selection. Semantic judgments use a predefined rubric and blinded human review on a stratified sample; a second model judge is useful triage but is not independent ground truth.

The proposed decision rule: advance if the adaptive method shows a practically useful improvement over the strongest comparison at matched spending, or reaches equivalent quality at materially lower cost. The planning targets of a five percentage point improvement in task success or a 20 percent cost reduction are decisions for Encevra, not effects predicted by the literature. Stop or simplify if a fixed verifier matches the adaptive policy, if gains vanish at equal cost, or if the system repeatedly turns correct answers into incorrect ones. Those results would identify a useful limit and save implementation effort.

Experiment two: learning under changing conditions

Each environment begins with component tasks that provide limited feedback. The system then encounters new instances and withheld combinations, a relevant rule changes, and it returns to earlier task families whose rules remain unchanged. In one example, a simulated data tool first uses one field convention, then changes a field's meaning: some earlier procedures remain correct, one becomes dangerous, and the environment can verify the actual output, exposing a system that merely updates a remembered label while continuing to compute incorrectly. Multiple independently designed environments keep the mechanism honest: a synthetic environment gives exact answers and controlled changes, while a document or code environment tests whether it survives messier evidence.

Two evaluation modes prevent accidental leakage. In frozen-transfer mode, learning stops before the withheld tasks and no answers return to memory. In online-adaptation mode, the system receives exactly the feedback the environment permits after each episode and can update only for later episodes. Never score an answer after revealing its reference solution, and keep the final family set unavailable during method selection, including through prompts used to invent synthetic examples.

Report curves over time, because a final average can hide a system that learns quickly, forgets badly, and recovers only after repeated corrective feedback. The strongest initial result would show beneficial transfer to withheld combinations, faster recovery after meaningful changes, and acceptable retention at matched lifetime cost. Evidence against the thesis includes a simple notepad winning, improvement only on near-duplicate tasks, gains disappearing when feedback is equalized, or a refusal-heavy policy appearing reliable because it stops doing useful work. If BASM already handles the changes well, the research should identify that result and focus on the remaining failure rather than relabeling the baseline.

How this could develop into deeper model research

Once logged episodes show which observations changed outcomes, Encevra can train a controller to choose checks or retrieve procedures, starting with an interpretable policy or small classifier before a language-model controller. Controller training is a new experiment with its own data split that includes failures, redundant checks, and cases where no intervention was needed; training only on successful traces would bias the system toward over-intervention. Targeted open-model training becomes reasonable if the backbone cannot consistently execute a learned procedure, encode its conditions, or distinguish a decisive counterexample despite strong context. Tiny Recursive Models show small specialized architectures can perform substantial iterative reasoning on particular structured tasks, which motivates narrow controlled experiments, not a claim that a tiny model replaces a general one.

Automated improvement stays bounded before it becomes open-ended. The Darwin Gödel Machine and AlphaEvolve demonstrate useful forms of automated search with external feedback; they do not establish unlimited self-sustaining improvement. Encevra can eventually automate a similarly bounded loop: propose a controller change, run a development evaluation, retain promising variants, and submit the chosen variant to an untouched evaluation, with the scorer and final test access outside the candidate's write permissions and human review reserved for promotion to real product use. The scientifically interesting measure is how often the loop discovers improvements that survive new environments, because an increasing development score can come from adapting to the evaluator. Foundation-model pretraining would be a materially different program, and the current evidence supports a systems-learning program first.

The first 90 days

The cost figures below are arithmetic illustrations, not current quotes or a commitment to spend; actual per-episode cost is measured in the pilot before any provider or sample-size choice. One episode includes every nested model call and tool operation used to produce its scored outcome. For illustration, 120 independent environment families, five systems, three repeated sequences, and ten scored episodes per sequence produce 18,000 episodes; adding 4,000 development episodes yields 22,000. The repeated sequences do not create 360 independent families.

The illustrations exclude salaries, substantial GPU training, taxes, and specialist review beyond the allowance. A 1,000-episode pilot at the same assumed costs would use $100 to $750 in episode spend before review and infrastructure. If actual episodes cost several dollars, redesign or reduce the study before running it at scale. The main financial control is staged commitment: run a pilot large enough to expose dominant failure modes, then estimate the final sample required for the chosen claim.

On ownership: Noetis owns the shared experimental code and learning mechanisms, Desk exposes document and decision workflows, and Lexplio supplies domain-specific evaluation needs and product feedback. Only authorized cases enter research datasets, and product customer material does not automatically become public training data. The research claim stays distinct from product reliability claims: a controlled simulation result can justify the next experiment, while deployment in consequential work additionally requires relevant domain evaluation and an appropriate review process.

What we would publish, and the decisions after

A useful first release would contain the task-family construction rules, source and environment versions, feedback protocol, grader specification, baseline prompts, costs, and per-family results, including negative results and representative failures, releasing only data and artifacts Encevra has the right to share. The title should state the actual finding: procedure revalidation improving adaptation under changing rules, if supported by results, never a general intelligence breakthrough from a narrow suite.

Encevra's next credible milestone is a reproducible result showing that experience improves future competence under a meaningful change. That would justify a larger research program and provide a concrete reason for others to take the lab's ambition seriously.

The sources

All 26 entries were consulted for this round on 30 September 2026. "Preprint" means the source should not be treated as independently established fact. Access notes identify the material supporting the discussion; they do not imply every appendix was audited. The authors have not evaluated or endorsed Encevra.

Continuity with earlier rounds: Brain Research VIII (20 September 2026), Brain Research IX (21 September 2026), the Opportunity Second Look (25 September 2026), and the Ownership Strategy (26 September 2026) establish the prior research and recommendations. Their historical implementation descriptions are not treated as a current code audit.