28 September 2026. The user approved continuing the adaptation-selection recommendation. This commissions intervention-choice-transfer, an independent bounded feasibility investigation. The root owns the question and synthesis; the ancillary investigator owns workload discovery, methods, diagnostic development, execution and publication.
Can outcomes from earlier interventions improve adaptation decisions on fresh work after the receiving agent or interface changes, beyond competent current diagnosis and unchanged reuse of the old decision policy?
This develops AD1–AD2 without changing their original wording. It is distinct from learning to investigate within a task and from evidence-set assembly. Those studies retain their existing remits; this investigation need not wait for them.
The preceding reading already identifies direct public allocation precedents. Demonstrating that model and harness changes can jointly help is insufficient. We want to know what useful decision experience persists, what receiving-state changes invalidate it, and when reacquiring decision evidence is worth doing.
Keep three contributions identifiable: better candidate remedies, better choices among available remedies, and better complete subsequent work. An acquired policy may live in explicit records, contextual instructions, executable rules or weights. Neural training is optional. An action may be a short sequence where the components complement one another; a competition among all five substrates is unnecessary.
Begin with a tractable task family, a small set of credible remedies and one material change to the recipient or interface. Establish that the remedies can be applied and that their consequences can be judged through complete outcomes, including still-valid obligations. The investigator may use published components or motivated authored tasks, with their provenance and limitations explicit. Do not presume that a useful choice problem exists merely because failures have different labels or different remedies produce different scores.
The first phase includes discovering whether earlier outcomes contain useful information beyond the present-state observations. A developed diagnostic rule, unchanged reuse of the earlier policy, and a good fixed remedy or sequence are serious alternatives. Source evidence, ordinary retrieval, caching and repair should remain available. Separate inherited investigator/public knowledge from feedback actually earned by the experimental decision maker. New task wording alone does not establish transfer across receiving states or histories.
If the comparison supports acquisition, pursue bounded learning and diagnosis; do not stop after a minimally developed failed recipe. If ordinary reuse or a cheap current-state rule explains the available outcomes, publish that boundary. Assess claims developed through exploration on fresh material. Charge discovery, trial interventions, failed attempts, policy acquisition, observations, later use and maintenance in relevant native units; a perfect retrospective choice is a diagnostic bound, not a deployable free policy.
These are scientific distinctions rather than a prescribed factorial design. The study selects informative contrasts and can refine the question as evidence develops. Phase closure rests on explanatory progress, a demonstrated limitation or a concrete resource constraint; a learned-policy advantage is not required.
The preparation ledger extends P123 to a pinned ModularRSI release. It offers candidate harness components and final-harness evaluation artifacts. It does not provide a ready matched intervention-outcome dataset for this commission. A release policy also incorporates feedback from a published evaluation benchmark. Treat inspected public cases and release decisions as background/development knowledge, not fresh confirmation.
ModularRSI is a starting lead, not the mandatory platform. Its integration cost may outweigh its value for a small comparison. The existing local S4 remedy feasibility and maintenance-transfer findings remain explained limits, not instructions to restart their old workloads. Workload construction belongs to the new study; preparation has not established working remedies or decision value.
Use gpt-6-astra for research investigators and reviewers under the standing
22 September policy, recording the actual reasoning setting, provider, harness,
starting context and interventions. Participant, teacher and learner models are
scientific choices. The dedicated 64 GB Mac Studio offers Docker Model Runner
serving; verify mutable-state/gradient access separately. Coordinate heavy work
with the other investigators. Local docker model, codex, agent, uv and
Docker/Compose remain available subject to actual capabilities.
The independent repository carries a self-contained question, local instructions, public-source boundaries and starting prompt. Preparing and publishing it does not launch a session or model run. The investigator should publish its findings and next bounded step locally; root assesses those publications against this question and the enduring research directive.
Prepared boundary: the private
intervention-choice-transfer repository
(local README)
is committed and pushed on main at
3cabc29aa1b784ec8a9129c568224d90badb9cd6, with a clean working tree and matching
remote head. Its
START.md
is ready for an independent gpt-6-astra investigator. This records preparation,
not a started session or an accepted experimental publication.
The root review advances
the preparation boundary to cb25d5c64bd64859e0e9abbcfc41f0e8f44762df and accepts
the first bounded experiment plus source-parser follow-up. Counting 24 earlier
old-interface outcomes produces a functioning explicit category policy. On eighteen
fresh nested-API tasks it completes sixteen, versus seventeen for a fixed canonical
adapter; shared source-order repair brings a tie. The global inherited choice
already matches ordinary reuse. This is an interface change with the model fixed,
not receiver-model transfer or an isolated causal estimate of chooser degradation.
The first action pair is complementary, but its 18/18 retrospective bound requires outcome knowledge unavailable to a deployed chooser. A subsequently developed and frozen parser handles the closed source grammar, completing 36/36 fresh API tasks (eighteen source cases, two interfaces) without model calls. Its development is charged as remedy construction, not learned decision improvement. Freshness means new outcome evaluation within the same grammar; earlier structural tests had instantiated the seeds, as disclosed. Root replays saved generations, decisions, repairs and parser outputs and verifies historical freeze hashes.
This establishes no added intervention-choice value or repayment of acquisition. It is a useful explained workload limit, not a general failure of learning. Close this template phase and retain the broader remit. A later workload should first establish reliable source labels and what remains unresolved after competent ordinary parsing, retrieval and repair; then assess whether observable information or earlier outcomes can improve the decisions. Linguistic variety alone is not sufficient, and a positive selector gap is not an admission requirement. P126–P127 narrow the relevant method comparison. The investigator owns further methods and workload discovery; this review makes no new commission.