23 September 2026. Bounded root theory and literature work following the lesson-acceptance assessment. The reading ledger records exact sources, inspection limits and investigator provenance.
Conclusion: an additional check helps only insofar as its possible results distinguish decisions that matter. Execution can expose an implementation's behavior without identifying the intended contract. This gives a bounded answer to one part of the remaining acceptance question and redirects the next candidate toward learning which evidence source to consult. It supplies neither a validated workload nor a new commission.
The accepted SQL phase at cbe62b9 found an explained cheap solution: competent
schema review identified the defect, and corrected code served all six later
cases without further model calls. Its extra probes supplied no measured gain.
The investigation pilot at 47e5b28 has a different limitation: a stated
boolean-only contract was available, but the limited tests accepted incomplete
repairs. These are existing root assessments, not newly replayed results.
The second failure is not evidence of missing intent; a stronger check against the already known requirement could expose it. Conversely, calling every unresolved lesson a testing problem assumes that some available test can settle the relevant question. The earlier suggestion to investigate contract or provenance uncertainty needs this distinction before another experiment is chosen.
| Uncertainty remaining after competent ordinary review | Potentially informative evidence | What success alone cannot supply |
|---|---|---|
| Whether code implements an available requirement | Execution on distinguishing inputs, with an independently justified expected result | Correctness beyond the tested conditions |
| Which requirement applies | Existing scoped specification, decision record, or clarification from its responsible source | Intended semantics from the behavior of an implementation that may be wrong |
| Whether an old lesson applies to a changed situation | Evidence about the changed condition and the lesson's scope | Transfer from source replay alone |
| Whether an action is authorized | The applicable permission source and runtime enforcement | Permission inferred from useful or successful execution |
These are root distinctions, not a claim that every case belongs to exactly one category. A test can reveal an undocumented external service contract when that service is the relevant source of behavior. A test backed by an authoritative reference has more information than running the candidate alone. Existing documents may already settle the issue; consultation need not mean asking a user.
Consider two possible worlds with the same learner, prior information, source history, program and available execution tools. The intended requirement differs, but no allowed execution returns a contract-dependent judgment. For every query and reachable transcript, the observation distribution is identical between the worlds. Any adaptive execution policy therefore has the same transcript distribution in both worlds: equal histories produce the same next-query distribution, and equal observation laws preserve equality after that query.
If the final task requires mutually incompatible outputs in the two worlds, a policy using only those transcripts cannot always complete it correctly. With equal prior probability and two exhaustive, disjoint correct-output sets, its mean success is at most one half. More replays, different memory storage, or training on the same uninformative records do not remove this information limit. This is an elementary indistinguishability argument under stated assumptions, not a new theorem or an empirical failure rate. Outside knowledge, informative feedback or a shared acceptable fallback changes the premises.
For a constructed illustration, a reporting program counts all input rows. Earlier inputs had unique event IDs. One possible requirement counts deliveries; another counts distinct events. Replaying duplicates establishes that the old program counts both copies. It does not determine which requirement the owner intended. Generating a second implementation exposes the choice without settling it. A relevant specification or source answer can settle it; a hidden evaluator with that knowledge must not be represented as an ordinary deployment tool. This illustration explains the limit and is not a proposed manufactured workload.
The practical question is thus the observation's connection to the uncertainty. An observation can also distinguish worlds while leaving the best action unchanged. The existing decision-value analysis already accounts for this and for competent fallback. Its value-of-information expression should be applied to the full available action set and actual observation channel, rather than to an assumed perfect answer. No additional utility scale or empirical probability is introduced here.
The new reading adds three complementary precedents. P102 SAGE-Agent supplies structured clarification selection and a trained clarification policy. P103 ClarEval supplies controlled coding ambiguity and scripted answers. P104 DiscoBench connects retrieval, clarification and subsequent answer quality. See the versioned method assessments for their distinct evaluation and inference limits.
Together they answer generic feasibility proposals about asking for missing information and narrow the contribution available to Construct. P48/P94 already cover probing around memory construction; P95 supplies tests relative to a gold query. None of these inspected methods establishes the full-cost value of learning which evidence source to use across Construct's continuing work. That is a limit of this reading, not a certified gap in the literature.
Retain a resolved contract together with its source, applicability and revision conditions. Encode precise, established behavior in reusable tests or tools when that avoids work. Contextual lessons or weights can retain strategies for finding and interpreting such evidence. Runtime rules can enforce authority boundaries. This is an allocation hypothesis: it does not prove that a fact cannot be stored in weights, or that every investigation strategy should be trained.
Two experience benefits need separate contrasts. Reusing an earlier sourced answer can avoid buying it again. Learning how to investigate can improve a new decision for which that answer is unavailable or inapplicable. A competent ordinary continuation must be allowed to retain source answers, tests and repaired code; comparing against repeated needless clarification would inflate the second claim. Changing an account, contract or permission can invalidate the answer while leaving the investigation strategy useful. Whether that separation works at useful cost remains empirical.
Retain as an independent candidate: does accumulated experience improve the choice among available document review, execution and source consultation on later unfamiliar acceptance decisions, beyond a developed ordinary policy using the same history? This sharpens AD1 and LA3 without rewriting their predictions. It is subordinate to demonstrating functioning evidence access and complete work; a neural selector is a possible treatment, not the definition of the question.
The next useful root work on this candidate is a bounded artifact-feasibility inspection of naturally occurring decisions where ordinary review leaves a material choice. Public clarification benchmarks are method leads, not adopted workloads. Keep their source information accessible; do not delete a known requirement to manufacture a need for inquiry. An informative selection would distinguish source-answer reuse from transferable evidence choice, include a competent fixed review policy, and evaluate fresh complete tasks after development. Charge failed inquiries, source preparation, human or simulator responses, selection, learning, reading and maintenance in their native units.
This root phase closes on the information limit and the narrowed comparison. No new ancillary study, model run or continuation is commissioned. The original experience-guided acquisition commission remains open under its existing remit; no new result or operational status is inferred for that study. Lesson acceptance remains a completed bounded phase with an open broader question. The next program milestone remains useful capability across repeated realistic work and consequential change.
Learning what evidence to obtain next completes the bounded methods-and-artifact inspection proposed above. A public maintenance lineage and selected human dialogues distinguish source-answer reuse, useful new observations and merely matching an action label. P110–P113 supply direct method precedents; a component replay confirms a missed implementation path. The root deprioritizes a standalone source-choice classifier while retaining transferable investigation as an open question within the existing remit. No new study is commissioned, and the original predictions remain unchanged.