Explorer A, turn 2: success is not yet evidence of recovery

Result and question originated

The documentary comparison currently supports evidence correction and a precise acquisition gap, not a new condition established against the scoped recovery mechanism. The strongest positive instance is the review workflow's recorded substitution. The tag case instead documents construction of verification machinery. The linking-contract record documents contradictions outside the validator's checks. These are different explanations of continued operation.

The question this comparison originates is: when a system keeps working despite deficient support, what connects that deficiency to the successful task? That connection might be a noticed obstruction and replacement, a task that never required the deficient support, or an outcome check that did not examine the affected requirement. Comparing those possibilities changes what “it kept working” can explain across cases. It does not make every success suspicious or require an immediate service consumer for inquiry.

This follows the turn-1 integration and turn-2 commission. Evidence below is documentary observation and conceptual comparison. No experiment was run. No external-source conclusion is introduced, and no ingest analysis is used as support. The retained records describe earlier work; this inquiry has not independently reproduced it.

The actual incumbent, including its limits

The recovery account says an interpreter retains the goal and can derive another route when the first is blocked. A successful outcome can therefore coexist with failed infrastructure. Repeated workarounds spend bounded context, attention and reasoning capacity.

Its Scope already restricts the claim to detected blockage followed by rerouting. Unnoticed false rules and poisoned instructions are excluded. Its budget claim is explicitly qualitative, not measured. Accordingly, finding a defect alongside continued operation does not refute the scoped mechanism; it may show that an example never established its antecedent.

The problematic evidential passage is stronger. It says linking contracts accumulated contradictions during normal operation and an index's completeness promise degraded quietly until validation made the breakage loud. Its footer describes readers as absorbing that degradation. Those historical claims need encounter evidence beyond the mechanism's plausibility.

The strongest rival throughout is the unchanged scoped account plus corrected case attribution. It explains an observed fallback without claiming that every surviving defective system used one. Merely spelling out feasibility, permissions or remaining capacity would not defeat this rival: rerouting already implies an available route. Such conditions become a substantive explanatory addition only if they distinguish cases the incumbent otherwise treats alike.

Three comparisons, with their unresolved elements

1. Tag completeness: an improvement success, not established reader recovery

The tag-readme trace records a large index, manual threshold calibration, a revised type and validation contract, migration, and a later check catching a block-YAML member missed by the documented search recipe. These are positive observations about how maintenance succeeded. The trace also explicitly says no size or completeness validator existed before ADR 026 and the completeness promise was never observed broken. Detection was human; the change installed the early-warning mechanism.

The opening description of strain and the later discovered omission prevent a stronger negative conclusion: they do not prove every earlier reader obtained complete results. Equally, they do not identify a reader who noticed missing material and compensated. The later validator discovery must not be moved backward in time and turned into an earlier encounter.

Disposition: the episode supports a successful change in verification, not the claimed history of readers repeatedly absorbing a completeness failure. This is a misclassified supporting example at the present evidence level. Whether reader recovery also occurred is an acquisition gap. A contemporaneous reader trace showing an incomplete head, recognition, additional retrieval and a satisfactory result would change that classification. A final artifact or the later migration alone would not.

2. Review workflow: a positive recovery instance with a visible boundary

The improvement log, entry beginning “ADR 030 batch endpoints,” reports the first harness-orchestrated review on 2026-06-12. Selector, prepared batches, parallel reviewers and ingest completed, recording four note/gate pairs without Python changes. Workflow arguments did not reach the script, so data were inlined. Deterministic endpoints ran in the parent because the script sandbox had no shell.

Here obstruction and substitution are named, rather than inferred from eventual success. This is the clearest documentary fit to the incumbent. The shell limitation also supplies a useful boundary: continuation was available at the parent-conversation level, not as an in-script sweep loop. Calling the script itself resilient would assign success to the wrong executor. The record supports completion of a redistributed workflow, not removal of the sandbox limitation.

A possible explanatory refinement is to identify which actor retained a feasible route to the goal. But the current record does not show that the incumbent fails without this refinement; it may simply instantiate its goal-holding interpreter at the parent level. A discriminating negative case would have noticed blockage and a retained goal but no permitted executable substitute. That would test an unconditional “noticing guarantees recovery” reading, not the incumbent's narrower mechanism.

The log also records absent token telemetry and an unverified model-partition assertion. Four recorded pairs establish a specific completion result, not every assurance of the intended workflow. The approximately 52-second reviewer duration does not isolate fallback effort. Recurrence and a net maintenance saving remain unknown. This instance supports recovery without proving a recurring tax of consequential size.

3. Linking contracts: successful validation without tested contract agreement

A bounded search found the linking-contract consistency workshop. It records a cross-collection audit finding decisions, local contracts, vocabulary and procedures describing different systems. Examples include global versus collection-owned label vocabularies and disagreement over allowed contract serializations. It also states that deterministic validation checked link existence rather than label authorization, direction or usable outbound contracts; those contradictions therefore validated cleanly.

This is a strong counterexample to treating passing validation as evidence that contradictions were noticed and worked around. The documented check did not cover them. It is not a counterexample to recovery after noticed blockage, and it does not establish that agents never silently resolved conflicting instructions. The workshop describes audit and reconciliation work, not a sample of pre-audit authoring encounters. It also does not establish the recovery note's “over a year” duration. That duration remains unverified here.

The relevant successful regularity is narrower and useful: a system can keep producing accepted artifacts because acceptance tests only selected properties. The tag episode improved the checked properties; the linking audit exposed properties outside the check. This comparison distinguishes improved detection from successful compensation without treating either as failure of the whole system.

What changes in understanding, and what resists

Across the cases, “success” refers to different things: continued repository activity, a completed workflow, or passing selected checks. Treating these as interchangeable loses the relationship between the affected support and the accepted result. The review substitution gives path evidence. Tag migration gives a chronology of changed checks. Linking reconciliation gives explicit limits of checks. No single recovery story earns all three.

The intended-path health account already supplies an important constraint: equivalent final outcomes cannot identify which path ran when no independent path signal survives. It also distinguishes path deviation from weakened guarantees. The review log's independent substitution record is therefore a successful application of that account, not its refutation. The comparison improves evidential classification more than theoretical reach.

A further conjecture is worth retaining without adoption: recovery explanations may be most informative when they identify the level at which alternatives remain feasible. The parent/script distinction makes this question concrete. Against it, the incumbent may already contain everything necessary once “interpreter” is assigned correctly. Evidence showing systematically different outcomes at the same nominal blockage and goal, depending on available authority or tools, would strengthen the refinement. These documents alone cannot do so.

There is no demonstrated compression gain or task benefit. Learning value consists in separating a genuine recovery instance from two different success mechanisms and preventing an unsupported causal history from carrying the budget argument. Ordinary explanation-seeking inquiry with explicit rivals could achieve exactly this result; a separate compression objective has not earned comparative credit.

Recoverability, next inquiry and parked alternatives

The minimum recoverable evidence is not a generic exception summary. Preserve:

• The incumbent's scope beside its historical supporting passage.
• The tag chronology, negative findings and distinction between manual detection and later automatic discovery.
• The review obstruction, exact substitutions, actor boundary, recorded outcome and telemetry limits.
• The linking audit's named inconsistencies and the properties its validator did and did not check.

For missing encounters, distinguish evidence never recorded from records discarded or merely not found. This search cannot tell which happened. Git commits can recover contract versions and changes; they cannot be assumed to recover a reader's recognition or extra effort. A future acquisition request would seek bounded encounter records with obstruction, recognition, replacement, outcome and assurance changes, rather than another polished retelling. Measuring recurring cost additionally needs repeat encounters and attributable effort, not just duration of the successful task.

Do not make all later exploration a hunt for defective evidence links. This discrepancy has now originated a broader question about the contribution of supports to successful work. A later turn could seek a positive regularity in another domain and ask what common mechanism explains it, retaining unchanged accounts as rivals. That would test whether this steering approach finds more than retrospective correction.

Park general retention architecture because these results specify evidence needs without selecting a store. Park a saturation model because neither repeated fallback costs nor capacity effects were acquired. Park a controller comparison because no competing selection histories or comparable costs exist. Broader literature on causal attribution and redundancy remains an ungrounded lead; no new source destination was allocated. Broader graph discovery through cp-skill-connect is a possible later handoff, not executed here.

Only this result is owned. No library repair, source mutation, experiment or delegation was performed. Costs are unknown to this worker. The historical recovery and cost questions remain partial for the specific evidence gaps above.