Explorer B: investigate successes whose explanation does not survive comparison

Result and evidence boundary

A promising question is when does continued successful work show useful recovery, and when does it merely leave the role of the supposed support unknown? The KB already distinguishes final success from mechanism health. The worthwhile investigation is not to rediscover that distinction. It is to determine whether its explanation of successful operation actually fits its concrete cases.

A focused search found a consequential mismatch: a theory note presents the tag-README episode as quiet completeness degradation absorbed during operation, while its linked evidence explicitly says the completeness promise was never observed broken. This is documentary friction inside an existing explanation, not a newly demonstrated runtime phenomenon. Beside a genuine recorded workaround and a source-supported memory intervention result, it suggests a useful investigation into different reasons that work can succeed without establishing the adequacy of its named support.

I searched theory descriptions and passages about success, surprise and workarounds, then followed the relevant evidence and log entries. This is a selected comparison, not a corpus survey. I did not read the other explorer's result or return. No experiment, source mutation or library repair was performed. External empirical claims below use retained Quotes only; library arguments and historical records remain different kinds of evidence.

Three cases, with the actual prior accounts

1. A successful workflow with a recorded substitution

The improvement log records the first harness-orchestrated review run on 2026-06-12. It completed selector, prepared batches, parallel reviewers and ingest, recording four note/gate pairs with no Python changes. It also records that workflow arguments did not reach the script, so data were inlined as a workaround. Deterministic endpoints had to run in the parent because the script sandbox lacked a shell. Token telemetry was absent, and the recorded model partition was an unverified orchestrator assertion.

Documentary observation: the run record contains both success and a specific substitution. It therefore supports the existence of recovery without establishing how much recovery cost, how often it recurred, or whether guarantees were preserved. The approximately 52 seconds reported for parallel reviewers is not a measurement of the workaround tax.

The relevant existing account is final task success does not establish intended-path health. Its claim is conditional: when primary and fallback paths share the same terminal outcome, that outcome alone cannot identify path health. The log does not refute it; the independent path record supplies exactly the missing distinction. The open question is why this run remained feasible and whether the substitution was incidental or a recurring dependency. Ordinary problem-led inquiry could fix argument delivery and stop. Curiosity about shared understanding asks what made continuation possible and whether that explanation applies elsewhere.

2. A successful KB operation retrospectively described as recovery

A goal-holding interpreter fails soft argues that agents can retain the goal and reroute around blocked paths, paying a recurring cost. It says an index's completeness promise degraded quietly until validation made the breakage loud, citing the tag-README trace.

The evidence note is more discriminating. Its verified timeline says no completeness validator existed before ADR 026. It explicitly states that the completeness promise was never observed broken. Detection was human; the early-warning signal was installed by the change, not responsible for originating it. The trace also records a real later discrepancy: the new symbolic check found a block-style YAML member that the documented search recipe missed. That finding changed the recipe.

Documentary observation: these accounts cannot jointly support the specific historical story that an observed broken completeness promise was being absorbed by recovery. Continued operation is recorded; a relevant sequence of compensating actions is not. Nor was the improvement simply a clean separation of two artifact jobs: the trace says the README still introduced the tag and could enumerate membership. What changed was verification.

This matters beyond correcting a sentence. The explanatory account may be taking continued operation as evidence for the mechanism it proposes to explain that operation. The strongest live rival is simpler: the deficiency was not encountered as a blocking condition, so no compensation was necessary. Another possibility is that users independently searched around it, but the cited record does not establish that. I cannot supply their missing history.

The case resists a broad recovery explanation while supporting a narrower account of human detection followed by changed checking machinery. Its current historical detail is therefore learning material, not an inconvenient exception to summarize away.

3. Weak dependence on a supplied memory is a different condition

The retained Quotes in Large Language Model Agents Are Not Always Faithful Self-Evolvers report controlled interventions across four frameworks, thirteen backbones and nine environments. They report larger behavioral changes under raw-experience perturbations than condensed-experience perturbations, with weak dependence on condensed experience persisting when raw experience was absent. The retained methods distinguish removing semantic content while preserving formatting from omitting an experience section entirely.

Source claim: this supports a bounded raw-versus-condensed behavioral-dependence contrast. It does not establish that every successful trajectory ignored memory, that summaries faithfully preserved the relevant information, or that condensation generally reduces task performance. It supplies no evidence about what caused the Commonplace episodes.

The existing activation account already separates presence, exposure, behavioral uptake and benefit. This is not a surprising discovery relative to that account. It becomes useful here as a limit on unification: weak dependence is not the same as detecting an obstruction and repairing it. A capable system may solve a task without needing the supplied material; alternatively, the material may have lost what would make it useful. The retained Quotes do not decide those explanations.

Candidate shared understanding, and its strongest rival

Conjecture: apparently successful operation with questionable support should originate questions about which contribution the support actually made. Three distinctions must remain separate:

• A support failed on an encountered path, and another path replaced it.
• The support's inadequacy was not encountered or did not matter for that task.
• The support was supplied, but its semantic content had weak measured influence.

These are not exhaustive or mutually exclusive states. A run could ignore one support while repairing another. Their shared structure is only that success leaves an explanatory dependency unresolved. It does not establish a single psychological or computational mechanism.

The strongest rival is that this is just ordinary causal diagnosis under a broad heading. The existing non-identifiability and activation notes already contain the main constraints. If comparison adds only a new umbrella phrase, no improved understanding has been gained. The concrete opportunity is narrower: the tag-README mismatch exposes an overextension of the recovery account and asks which other supporting cases actually contain noticed blockage plus substitution. A finding that few do would narrow the account even if it produced no reusable mechanism.

The next discriminating evidence would be an existing record connecting obstruction, recognition, changed route and result. For the tag case, contemporaneous use records showing readers encountering missing members and compensating would support recovery; records showing no encounter would support the nonbinding-deficiency rival. Absence of either is unresolved, not evidence of absence. For the memory case, task-relevant content preserved across the compared forms would be needed before attributing weak dependence to presentation rather than information loss. No such new acquisition is authorized here.

How this could steer inquiry

Improvement of shared understanding: originate a question where one explanatory claim is carrying several concrete cases but the cases disagree about its necessary conditions. Select the case that could force the largest consequential revision of that claim, including restriction rather than unification. Here, preserving the actual tag timeline changes which observations can count as evidence of recovery. That learning value does not depend on an immediate service consumer.

Ordinary problem-led inquiry: start from the broken argument channel, the missed YAML member, or an underused memory. Seek a repair that restores a specified outcome. This can also discover general mechanisms; curiosity is not its opposite. Its distinctive stopping opportunity is successful local repair. Shared-understanding inquiry may continue because the explanation connecting cases remains unsettled, even when all current tasks succeed.

Novelty seeking: select an unfamiliar mechanism, new system or striking analogy. It could find valuable evidence, but unfamiliarity alone would favor another success story over re-reading an old timeline. The selected discrepancy shows why an old case can have more learning value than a new topic: it constrains an already-used explanation.

This proposal offers no compression score. Any economy would be explanatory: fewer unsupported causal assumptions while preserving distinct routes to success. It might require longer accounts and more retained differences. Demonstrated task benefit remains unknown. A new documentary comparison would demonstrate only whether the account fits the retained cases, not whether a curiosity controller selects better investigations.

Recoverability, alternatives and stopping

Retain the workflow's substitutions alongside its completion claim; the tag trace's negative findings, chronology and commit identifiers; and the source's intervention definitions alongside its reported asymmetry. Keep intended route, actual route, and unchanged outcome distinguishable. A summary saying successful systems route around defects would destroy the very contrast under investigation. The log lacks a complete operational trace; its omissions cannot be reconstructed from the general theory. Commit references offer recovery routes, not a guarantee that every relevant encounter was recorded.

I prioritize the recovery-account boundary because it supplies a concrete tension between an explanation and its own evidence. Park the condensed-memory mechanism question because the retained extracts do not separate content loss, irrelevance and failure of use. Park general success-audit instrumentation because this commission does not authorize implementation and explanatory discrimination should precede a new logging design. Park a broader literature search on redundant causes and causal credit assignment as an ungrounded lead, not an established account or novelty claim.

The useful stopping result may be that the shared account adds nothing beyond existing distinctions, with one historical overstatement identified. That would still improve the evidence boundary. It would not justify a generalized memory architecture, a new controller, or retrospective claims that the first exploration should have selected this direction.