1
Delayed feedback can act on a temporary trace
Yagishita and colleagues stimulated glutamatergic and dopaminergic inputs separately in mouse striatal slices. Dopamine promoted enlargement of stimulated dendritic spines when it arrived roughly 0.3 to 2 seconds after the relevant input. He and colleagues then found distinct temporary traces for strengthening and weakening cortical synapses, with later neuromodulator signals converting those traces into lasting changes, and with which signal worked depending on the circuit.
Candidate experiment
Keep a record of each operation that may receive later feedback: the case, input version, tool call or extraction, affected claim and downstream uses. Attach a reviewer's correction to that record, then compare three policies on controlled cases: update the whole run, update the most recent operation, or update only operations linked to the corrected field. Delay and reorder the feedback, include two similar cases, and measure whether the right operation is corrected and whether unrelated behaviour changes.
Transfer limit: Software can retain exact identifiers, timestamps and dependency records, so it need not imitate a decaying trace or expire feedback after two seconds. The useful hypothesis is selective attribution across a delay. The biological experiment does not show how to infer responsibility from an ambiguous customer comment.
2
Learning also needs mechanisms for stability
Turrigiano and colleagues showed that prolonged suppression of activity in cultured rat cortical neurons rescaled synaptic strength while retaining relative differences among inputs. Hengen and colleagues recorded visual cortex in freely behaving rats and saw firing rates return toward baseline over several days despite continued deprivation. Bissen and colleagues (2026) then qualified a fixed setpoint account: learning during the visual critical period established new stable spine-density levels, and was absent in adults under the tested conditions.
Candidate experiment
An orchestration system can develop escalating workloads when one component emits more candidate issues and the responses trigger further review. Test a separate workload controller against a fixed budget and the existing routing policy, using repetitive low-value alerts alongside rare decisive findings. Measure unnecessary calls, valid findings lost, completion rate and correction quality.
Transfer limit: Regulating activation is not the same as calibrating truth. Automatically rescaling confidence, suppressing a productive specialist, or forcing every role to contribute equally would need independent justification.
3
Astrocytes can regulate memory related circuits
Williamson and colleagues identified learning-associated astrocytes in the mouse hippocampus; reactivating the tagged ensembles stimulated recall and deleting NFIA from them suppressed it. Dewa and colleagues described a slower astrocytic state that stabilises later memory, and supplied the limit: more stabilisation was not uniformly better, since stronger conditioning produced freezing in a different context, read as reduced precision. Han and colleagues showed human glial progenitors can alter a mouse neural system, which is a bridge, not a human brain.
Candidate experiment
Distinguish a provisional reusable rule from a rule promoted after further independently verified use. A corrected extraction pattern could stay narrowly scoped until it succeeds on different documents with the same structure. Compare immediate promotion, promotion after independent validation, and a fixed waiting period, measuring both retention of useful corrections and incorrect transfer to near-matching cases.
Transfer limit: Lower priority, because earlier rounds already recommended reviewed memory updates. It becomes a distinct experiment only if it isolates the promotion condition and tests persistence against overgeneralisation. Adding an astrocyte agent would not test the finding.
4
Learning can alter communication infrastructure
McKenzie and colleagues disrupted formation of new oligodendrocytes in adult mice while preserving established myelin; those mice were impaired at mastering a wheel with irregularly spaced rungs, while previously learned performance was much less affected. Lakhani and colleagues trained 17 healthy adults over four weeks and reported increased myelin water fraction in two task-relevant regions, in a small before-and-after design with no untrained control group.
Candidate experiment
For repeated workflows, measure which dependencies dominate completion time before adding more analysis. Hold message contents, recipients, model settings and analysis logic fixed and change only scheduling or preparation of the same dependency. Introduce variable tool latency, verify identical evidence reaches synthesis, and measure completion time, redundant waits and quality under delayed critical results.
Transfer limit: None of this is biological myelination. Treat it as a performance investigation only when execution traces show a transport or scheduling bottleneck. Biological evidence cannot justify a cache lifetime, a service topology, or a promised speedup.
5
Stable behaviour can outlast a changing neural code
Driscoll and colleagues repeatedly imaged posterior parietal neurons in mice and found relationships to task features more stable within a day than over weeks, even while behaviour stayed stable. Gallego and colleagues recorded six monkeys for up to about two years and recovered stable low-dimensional population dynamics across sessions, while fixed decoders tied to the recorded activity degraded more. The recorded set of neurons changed, which is recording turnover, not proof of cell replacement.
Candidate experiment
A model update can change wording, field conventions, omission patterns or how uncertainty is expressed. Compare direct free-text handoff with a validated structured interface when swapping a specialist's model version, holding the source packet and downstream evaluator fixed. Include cases where the replacement uses different terminology, units or ordering, and cases where its output is actually worse. Measure material decision agreement, semantic field errors, invalid conversions and failures correctly surfaced, against a simple deterministic format converter.
Transfer limit: Consistency is not sufficient for quality. A stable interface can preserve the same mistake indefinitely, so score agreement and correctness separately. The proposed benefit is operational continuity under a controlled substitution, not evidence that Noetis has reproduced a cortical manifold.
6
Value depends on its comparison context
Louie and colleagues modelled value coding as divisive normalisation, where a response depends on the wider set of options, and matched human and monkey choice behaviour. Cox and Kable found steeper fMRI value-response slopes in a narrow range of subjective values, supporting range-dependent representation through an indirect signal.
Candidate experiment
Create paired cases where options A and B, the objective, constraints and evidence are unchanged. Add an option that is worse on every relevant criterion, duplicate an existing option, or change an irrelevant extreme value, and observe whether the recommendation between A and B moves. Include legitimate changes, such as a new feasible option, where the system should adapt.
Transfer limit: An efficient biological code can produce behaviour that is undesirable in software. The lesson here is a failure test, not an architecture to imitate. There is no need to invent a normalisation layer in Noetis.
