Subsequent evidence — 29 September: The commissioned
intervention-choice study
now has an accepted first phase at cb25d5c. An acquired chooser adds no benefit
beyond ordinary reuse, and a source-only parser resolves its authored grammar.
This illustrates why the available remedies matter to the decision-value
comparison; it does not estimate that value on broader language distributions.
The original reading below remains historical; the study's broader remit is open.
Subsequent approval — 28 September: The user approved continuing this selection. The intervention-choice commission now assigns bounded feasibility to a prepared independent study. The completed reading and recommendation below remain the basis for that later decision.
28 September 2026. Completed root literature-and-theory phase during ancillary execution. This develops AD1–AD2 without changing their predictions or commissioning work.
Conclusion: choosing where to intervene is already a public research problem with reported experimental results. A generic demonstration of model/harness co-evolution would add little. The more consequential remaining question is whether accumulated intervention outcomes improve decisions beyond competent current diagnosis, and whether that experience remains useful when the receiving agent changes. A better remedy, a better decision and cheaper continuing work are separate achievements.
The source ledger records exact versions, inspected sections and limitations. It adds four methods and checks two pinned public implementations. Results below remain author-reported; root ran no participant models or published experiments.
| Method | What determines the change? | Consequence for selection |
|---|---|---|
| P120, Qwen-Planner-Agent | Prescribed alternation of model training and harness editing. | Co-development alone does not test learned allocation. |
| P121, MetaRSI | An experience-revised policy schedules data, harness and model operators. | Direct precedent for improving the improvement policy; generic feasibility is no longer merely an architectural proposal. |
| P122, Ecdysis | Reviewed failure patterns guide harness edits, with the task model fixed. | Diagnosis and edit scope matter; recurring failure alone does not locate its cause. |
| P123, ModularRSI | Predefined modules evolve separately and are integrated; functions are selected at use time. | Localized construction and runtime selection differ from learning which substrate to change. |
The earlier P41/P43/P45 readings remain valid assessments of those versions. Their metadata versions have not changed. The newer comparison changes the program-level assessment, not the history of what was known on 16 September. ModularRSI advances from a discovery lead to inspected methods.
The most relevant reported MetaRSI comparison forks one improved starting system and compares its original and revised improvers. That is substantially closer to AD1 than an end-to-end gain over an unchanged agent. The inspected public release does not supply an independently reproduced version of that comparison. Neither publication acceptance nor an implementation gap establishes the result's truth or falsity. The ledger distinguishes the paper claim from the artifact.
Consider an agent that repeatedly mishandles a tool result. Retaining the correct contract may help; a scoped tool wrapper may enforce it; training may improve the agent's interpretation. A later combination may work better than any single change. Calling the failure a model weakness does not uniquely determine where the useful remedy should live. A wrapper can be a sound response to a model weakness, provided it preserves valid uses and remains worthwhile to maintain.
Separate three retained objects:
- The capability: an acquired lesson, implementation, adapter or rule that changes what the agent can do.
- The decision information: records of the state receiving an intervention, what actually changed, subsequent outcomes and the work consumed.
- The decision policy: the mapping from available observations and experience to an intervention, a sequence, a diagnostic check or leaving the system alone.
Experience can improve any of them. A policy can learn through revised explicit instructions; parameter training is not required. Conversely, a frozen model making different choices from current observations does not by itself establish that experience improved its decision policy.
This distinction matters when a system both changes its selector and learns to generate better candidate edits. Its total gain is useful, but cannot automatically be assigned to better selection. The marginal comparison needs the same starting agent and available remedy generators, with the decision policy varied. A separate comparison can measure better generators and their interaction with selection. These are attribution needs, not a requirement to freeze all components throughout a practical deployment.
Let o denote currently available observations, h earlier intervention outcomes,
and a an available action or bounded sequence. For a declared future workload
and an explicitly chosen utility, let u(a) be its potential outcome from the
same starting state. An idealized value of the additional information is:
[ V_H = \mathbb{E}{o,h}!\left[\max_a\mathbb{E}[u(a)\mid o,h]\right] -\mathbb{E}{o}!\left[\max_a\mathbb{E}[u(a)\mid o]\right]. ]
This is a decision-theoretic ceiling under known conditional outcomes and a shared feasible action set, not an estimator or a promised learning result. The ideal decision maker can ignore history, so the ceiling is nonnegative. Actual acquisition, observation and decision costs can make a learned policy less useful than its alternative. Without a justified utility conversion, keep complete quality, latency, tokens, training and maintenance work separate.
Two conditions make the ceiling zero: current observations already supply all outcome information in history, or the same action remains best for every history consistent with those observations. Thus differences between remedy outcomes are insufficient. Experience must change a useful decision beyond what a competent present-state rule already knows. Even a positive ceiling need not be learnable from the available histories or large enough to repay learning.
Our remedy-feasibility assessment already encountered the second limitation: ordinary explicit access solved the candidate workload without a useful remedy-choice problem. Reopening that workload unchanged would not test AD1. Equally, the maintenance-decision study shows why a better chooser is insufficient when available actions leave the complete task unresolved. Finding either limitation is a valid feasibility conclusion; a selector advantage is not an admission requirement.
A historical intervention outcome remains evidence about its original state. Its relevance to a new state is conditional. For example, adding logging without changing task behavior need not invalidate a useful remedy preference. Replacing a parser, changing a tool contract or teaching the model the missing behavior can change that preference. A component revision alone proves neither continued applicability nor universal obsolescence of prior experience.
This gives accumulated experience a possible role in deciding what needs to be rechecked, as well as what to change. The alternative is substantial: unchanged reuse can remain fully effective. The migration assessment already shows that paid adaptation need not repay its cost after an agent change. These local findings motivate AD2; they do not establish cross-substrate decision transfer.
Interventions can also change one another's value. Improving a reader may remove the need for a recurring instruction; obtaining evidence may make training possible; a tool wrapper may alter what an existing adapter must produce. The comparison therefore includes competent fixed sequences where appropriate. Selecting one permanent winning substrate would miss the original question.
Complete this root phase; do not open a general co-evolution study. The public precedents narrow that proposal. Retain transfer of intervention-choice experience as the next independent feasibility recommendation from this review:
Can outcomes from earlier interventions improve the next adaptation decision on fresh work after the receiving agent or interface changes, beyond competent current diagnosis and unchanged reuse of the old decision policy?
This sharpens existing AD1–AD2 rather than claiming a new problem. A bounded ancillary exploration could establish whether there are functioning alternatives, consequential decision differences and observable information from which to learn. Two credible remedies may suffice; a five-substrate competition is unnecessary. Its scientific target would separate decision experience from improving the remedies themselves, and assess complete subsequent work including preserved obligations. Current observations and source access belong to the competent comparators as well. Claims developed during diagnosis require fresh material.
The public methods support this question but do not supply a validated local
workload or a ready reproduction of the closest policy-update comparison. We
therefore recommend bounded feasibility, not an immediate controller benchmark
or broad training campaign. Closure can follow an explained lack of decision
value, a demonstrated method limitation or a concrete resource constraint.
If commissioned, the ancillary investigator owns workload discovery, methods
and execution under the standing gpt-6-astra investigator policy, with actual
reasoning settings and model provenance recorded.
This is a recommendation, not a commission. It neither waits on nor redirects
the active investigation and EBM agents. Their latest root review boundaries
remain 523eb3f and 8ac9b69; original predictions remain intact. The present
phase adds a literature-grounded distinction and research selection, with no new
learner result, ancillary phase or demonstrated cost advantage.