Explorer B: originate questions from success with unresolved contributions
Result and chosen case
Successful work can originate a question about which contribution made the success possible, without first being classified as recovery. The promising unknown is often a difference between what the result establishes and what its explanation attributes to a participant, artifact, or operation. Recoverable records should let later inquiry distinguish those contributions, not merely reconstruct the final deliverable. This is a proposed steering mechanism, not demonstrated improvement in search.
I chose the 2026-08-30 Commonplace revision, rather than another pass over the tag discrepancy. It records successful revision with model search, retained project material, operator corrections, and subsequent repository development. The unexplained regularity is productive continuation despite repeated global corrections. Did project-specific knowledge shape useful proposals, did operator selection rescue otherwise generic proposals, or did both make different contributions?
The actual prior account already calls this a human-inclusive case and explicitly denies an ablated causal effect, autonomous selection, and demonstrated gains from more computation. Those limits matter: discovering that the operator contributed would not overturn it. The attainable advance is a sharper account of where retained knowledge mattered and where the existing evidence cannot tell. This follows the turn-1 decision without inheriting its example as the only question worth asking.
Documentary observations and their limits
The current evidence note reports that the model retrieved project artifacts, compared formulations, and generated revisions. It reports successive operator objections to a defense portfolio, bootstrapping as a defense, and wording that treated computation as a future addition. The resulting account instead located the problem in selection and credit assignment within an already computational process. These are documentary reports, not a newly observed execution trace.
I recovered the contemporaneous record directly from local Git: commit 2b406eead12db44fde628b92fe7d34249531693a, path kb/work/theory-mediated-self-improvement-series/computational-theory-guided-conversation-episode-2026-08-30.md. It states the objective as making the research program coherent and exposing claims and tests that could change its status. It records the operator/model division and the transition from defense to provisional strategy. The current evidence note is therefore not the only accessible rendering. But the two records are not independent confirmation of the reported causal sequence.
Git also preserves a narrower observation independently of that narrative: accepted artifact changes. Searching the workshop README in the parent and accepted commit shows that the earlier document already distinguished computational closure from other achievements. The accepted version explicitly says the process already uses computation and separates computational search from human-assisted global selection. Thus the defensible contrast is increased specificity about the current process, not conversion from a uniformly non-computational prior. The commit message lists successive reframings, but does not supply the missing conversation.
The later commits named by the evidence note, 643908b9 and fe7e893e, are available. Their change summaries concern pre-decisive search control and independently executed symbolic operations respectively. This verifies preserved downstream development, not that a particular earlier sentence caused it. The evidence note also records later defects requiring repair. Successful revision cannot silently mean semantic correctness, reduced cost, or complete compliance.
Qualitative before/after comparison
A useful comparison would hold the retained episode and acceptance aim fixed while asking each account to explain the same distinctions. This is a documentary comparison proposed here, not a behavioral experiment.
| Common case feature | Coarse account | More discriminating account |
|---|---|---|
| Model generates substantial revisions | Computation contributed | Search occurred computationally; selection authority remains separately identified |
| Operator changes the framing repeatedly | Human help improved output | Feedback altered scope and program fit, not merely final acceptance |
| Project-specific concepts appear in output | The KB caused success | Retrieval and specificity support involvement, but not necessity or effect size |
| Later artifacts extend the accepted framing | Learning succeeded | Retention and continuation occurred; truth and autonomous reuse remain unsettled |
The coarse column is an analytical foil, not a quotation of the historical prior. The actual evidence note already makes most right-column distinctions. This comparison therefore demonstrates an available way to assess an account, not a fresh discovery that the KB previously lacked them. What remains learnable is the unresolved allocation of contribution within those distinctions.
Progress could consist in fewer unsupported entailments while retaining all observed changes. An account improves if it explains why a successful composite need not be an autonomous learner, and what additional evidence would decide the latter. It need not encode the episode in fewer words. A longer account may expose differences previously hidden by “success.” No literal coding gain, exact residual length, or universal progress score has been established.
The confidence–soundness note supplies a parallel argumentative distinction: confidence alone does not certify soundness, while its empirical association with soundness requires separate evidence. That constraint does not predict that the next confident answer will be wrong. Likewise, “successful composite work does not establish autonomous selection” restricts an inference; it predicts neither failure without the operator nor a frequency of operator intervention. Asking for event accuracy from either constraint would misclassify its contribution.
Strongest rival and discriminating evidence
The strongest case-level rival is operator-guided generic synthesis: a capable model could have generated comparable revisions from task context and feedback, while project artifacts supplied vocabulary rather than distinctive search constraints. This rival accommodates successful edits, explicit citations, and downstream continuity. It is stronger than claiming the KB was never read. A mixed explanation also remains live: retained material constrained local proposals while the operator supplied global selection.
A discriminating record would connect a specific artifact passage to a proposal before the relevant operator correction, preserve the correction, and show how the proposal changed. If the proposal already respected a distinctive project constraint that the task statement and feedback did not supply, that would strengthen the retained-material explanation. If the constraint appeared only after the operator supplied it, credit should move toward feedback. Neither observation alone eliminates latent model knowledge or redundant causes.
A withheld or information-matched replacement comparison could test a causal difference more strongly, but none is authorized or performed. The current missing full prompt and tool-call transcript blocks retrospective attribution at that resolution. The proper result is partial: artifact change and reported contribution are recoverable; necessity, magnitude, and comparative task benefit remain unresolved.
The strongest steering rival is explanation-seeking inquiry with explicit alternatives, without compression as an additional objective. It can choose this question for exactly the same reason: successful continuation has an unresolved explanation. Compression-informed inquiry would need evidence of different prioritization, lower search cost, or better discoveries to establish added value. This case supplies none. Learning value remains real as a candidate even without immediate delivery savings.
What acquisition must learn, and what must remain recoverable
Acquisition must learn more than whether the output passed. To distinguish these explanations it needs the task and acceptance boundary; which evidence reached the worker before each consequential proposal; actual proposal and correction text; and the outcome those corrections were meant to improve. Costs and operator effort are separate observations, currently unknown here. These are question-relative requirements, not a proposed memory store.
Three evidence states require different responses:
- Never acquired: no observer recorded a proposed causal input or distinction at the time. Later summaries cannot recreate it. Acquire a comparable future encounter if separately authorized; mark the historical gap.
- Acquired then discarded: an interaction was available during work but its recoverable record was not kept. Recollection can supply testimony, not the original sequence. The absent transcript here does not establish whether it was never durably captured or later discarded.
- Retained but unconsumed: the earlier episode survives in Git, even if a later worker reads only the current note. Retrieval can improve the account without new world observation. Availability does not establish that the historical worker consumed the same material.
An approximate account is not necessarily a lossy archive. A synopsis plus exact versioned artifacts can permit reconstruction of those artifacts while offering only an approximate explanation. A prose “exception” saying that the operator corrected the theory cannot reconstruct the correction's words, ordering, or context. Calling it a residual is insufficient: specify the recoverable object. Git here can recover accepted document bytes and their parent state, not the complete interaction or alternative proposals never committed. Even preserving every supplied message would not recover unobserved causal states.
Recoverability therefore protects future questions: later inquiry may need to ask whether an operator introduced a distinction or only selected a model-originated one. If only the final agreed wording survives, that distinction is lost regardless of summary quality.
Next question and parked alternatives
Prefer the question “Which project-specific distinction first appeared before operator correction, and where did it come from?” over another generic claim that memory helped. It targets an unresolved contribution within productive work, permits a meaningful negative result, and can reveal what future acquisition must capture. The present record cannot finish it; that boundary itself changes which inquiry is promising next.
Park a general retention architecture because it would outrun the discriminating question and belongs to adjacent workshops. Park numerical compression measures because no recoverable prose coding relation has been specified. Park causal intervention because execution is unauthorized, not because the documentary account makes it unnecessary. Broader causal-credit and redundant-cause literature remains an ungrounded lead; no literature result is claimed.
Only this result is authored. No current-turn A assignment, result, or return was read. No source mutation, implementation, experiment, or library repair was performed. Broader connection discovery through cp-skill-connect is a possible separately authorized follow-up, not part of this turn.