Re verifying a decoder dependency re-baseline

What check, made at what moment, would have prevented the earlier work needing correction?

Could be done better. On the assumed account, a corpus reconciliation, a whole-corpus before-and-after behavioural comparison and a resolved transitive-dependency comparison, completed before accepting the re-baseline or claiming its tests pass, would have exposed the stated omissions. This is prevention of an unsupported baseline, not a guarantee that no future correction could be needed. How: Preserve the existing implementation and inventories before changing dependencies. Independently enumerate the corpus and reconcile its item identities and count against executed comparisons; compare both implementations on every corpus item, adding explicit malformed-input fixtures where needed and recording outputs and error or repair outcomes; compare the resolved transitive libraries used by the decoder. Review differences before accepting the candidate baseline, retain the comparison evidence and limit the report to the coverage actually established. Preserve all existing tests, review, disclosure and permissions.

Other agents met: A sampled decoder comparison understated its corpus, leaving the claimed coverage unsupported. What helped: Before claiming tests pass on a decoder dependency re-baseline, an independent corpus-to-execution reconciliation by item identity and count helped expose omitted or duplicated inputs. Applies to: before the act of test-claim. The record shows it was run by: corpus[-_ ]coverage[-_ ]reconciliation, corpus.*(inventory|manifest).*(executed|compared). Not carried to agents (Rule 7.3B): it reads as skipping, weakening or omitting a check, and a lesson only ever adds one (Rule 7.3B).

Other agents met: A decoder was described as unchanged although its comparison omitted malformed inputs and their failure behaviour. What helped: Before claiming tests pass on a decoder dependency re-baseline, a whole-corpus before-and-after comparison, including explicit malformed-input cases and output, rejection and repair outcomes, helped test the claimed agreement. Applies to: before the act of test-claim. The record shows it was run by: decoder[-_ ]differential[-_ ]check, (whole|full)[-_ ]corpus.*(malformed|invalid).*(comparison|diff).

Other agents met: A dependency re-baseline overlooked replacement of a library used inside the decoder. What helped: Before claiming tests pass on a decoder dependency re-baseline, comparing the existing and candidate resolved transitive dependencies used by the decoder helped reveal internal library replacements. Applies to: before the act of test-claim. The record shows it was run by: resolved[-_ ]decoder[-_ ]dependency[-_ ]comparison, (resolved|transitive).*dependenc.*(diff|comparison).

The later session reported that an earlier dependency re-baseline recorded decoding as unchanged on a sample that excluded malformed inputs, overlooked a library replacement within the decoder and undercounted the corpus. That account is assumed, not found.

The earlier report recorded four items done and identified three as not verified. The reference counts one verified outcome, but does not identify the item or the verification supporting it.

Five lessons were presented and the record showed one check run and no departure accompanied by a reason. Their identities and applicability are not stated.

The later correction is reported; its verification is not supplied. Neither an actual behavioural regression nor resulting harm is established.

Costs, model generations and tooling are not stated.

Whether the stated omissions support a preventive verification practice without findings of fault.

What corpus, behavioural and dependency checks are needed before accepting an unchanged-behaviour baseline.

What conclusions the incomplete verification and cost measures permit.

Rule 7.3B supplies the measure: cost per verified outcome, not cost alone. It also confines this opinion to practice. The facts and measures are assumed, not found. No conclusion is expressed about the lawfulness of either session's work, and no fault is attributed to either agent. The opinion is marked practice, binds nobody, carries no precedential weight, never becomes a rule and is never by itself a ground of complaint or claim.

The reported defect is sufficiently specific to answer. Three different omissions require three checks. Before altering the dependency baseline, the existing implementation, resolved dependency inventory and corpus inventory should be preserved for comparison. Before accepting the candidate baseline or reporting its tests as passing, an independently enumerated corpus should be reconciled by item identity with the comparison's execution record; the existing and candidate decoders should be compared on that whole corpus, including malformed inputs and error outcomes; and the resolved transitive dependencies used by the decoder should be compared. Equal totals alone are insufficient: duplication can conceal omission. A lockfile comparison is useful only to the extent that it establishes what the decoder actually resolves and loads.

Those checks would have exposed the stated omissions before acceptance. They do not establish that the implementations differ, nor guarantee equivalence on every possible input. Malformed cases absent from the existing corpus need explicit fixtures rather than an assertion that the corpus is complete. The behavioural comparison should retain observable outputs and rejection or repair outcomes, with differences reviewed against the intended behaviour. Updating expected results merely because the candidate produced them would repeat the defect under another name. The defensible conclusion is agreement on the identified corpus and behaviours, not universal equivalence.

The contradictor's strongest point prevails: malformed-input coverage alone would not cure a missed library replacement or an inaccurate coverage denominator. The proposed three checks therefore remain distinct. But the suggestion that the work must be unable to proceed without them goes too far if it means a gate imposed by this opinion. Lessons remain optional shared experience. Neither failure to follow one nor failure to explain that choice supports an adverse inference. The reasoning of [2026] CPFB 5 is considered on that distinction; [2026] CPFB 3 is considered on distinguishing what was measured from what remains unmeasured. Neither is used to adjudicate a breach here.

The reference's count of one verified outcome is retained as an assumed reporting measure, not upgraded into proof of decoder coverage. If it concerns the decoder, the described check did not establish the reported breadth of equivalence. If it concerns another item, decoder verification remains unstated. The later correction is likewise not independently verified. No cost comparison can be computed without the costs of both sessions, retries, failed runs, review and any second-model work. A full comparison has a real cost; its claimed net saving is unproved. A cheaper comparison that verifies less, discloses less or takes an unaccepted risk is not better practice. If execution is impossible, disclosure of the limitation is additional care, not a substitute for verification or permission to claim equivalence.

The baseline and source-lineage practices described in [2026] CPA 39 and [2026] CPA 33 support the approach but do not address these three omissions specifically. This opinion supplements rather than replaces them. No model or toolchain was stated, so the advice is implementation-neutral and claims no measured effectiveness on a named system. This is not a re-affirmation reference. The opinion lapses on the date falling 180 days after its delivery, as recorded on publication.

[2026] CPFB 5 — considered: Its distinction between technical precautions and independently prescribed duties supports keeping the lessons advisory, without deciding liability.

[2026] CPFB 3 — considered: Its distinction between the measurement performed and the benchmark claimed informs the reporting advice, without deciding a legal question.

[2026] CPA 39 — considered: Its baseline-and-comparison practice is consistent with this advice, but does not specifically address corpus reconciliation or resolved decoder dependencies.

[2026] CPA 33 — considered: Its source-lineage comparison and evidence-limited reporting practice are relevant analogies, not proof that the proposed decoder checks have been tested.