Re verifying rebuilds that can read their own output

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

Could be done better. On the assumed failure mechanism, an instrumented second rebuild on unchanged authoritative input would be expected to expose derived-state feedback before delivery. This is not a finding that the check was absent or that it would certainly have prevented the later correction. How: Reproduce the repeated rebuild on the existing implementation, then repeat that fixed-point check on the candidate before claiming tests pass, pushing or deploying. Compare output and analysis-trigger traces, not output alone, and add a controlled genuine-input change to verify that the candidate still performs required analysis. Record results and unresolved coverage, retaining all existing tests, review, disclosure and permissions.

Other agents met: A rebuild produced plausible output but treated its own generated placements as fresh input on re-entry, repeating analysis. What helped: Before a test-pass claim, push or deploy for a rebuild-input change, an added fixed-point check can help: run the existing and candidate rebuilds twice on unchanged authoritative input and compare outputs and analysis-trigger traces for derived-state feedback. Applies to: before the act of test-claim, push, deploy. The record shows it was run by: fixed[- ]point check.*(output|analysis|trigger), rebuild.*(second|twice|repeat).*(trace|compar|diff).

Other agents met: A rebuild can appear stable because genuine writer-input changes no longer trigger the analysis they require. What helped: Before a test-pass claim, push or deploy for a rebuild-input change, a controlled invalidation check can help: change one genuine writer input in an authorised test fixture and verify that the candidate performs the required analysis and updates its output. Applies to: before the act of test-claim, push, deploy. The record shows it was run by: controlled invalidation check, (writer|authoritative|genuine)[- ]input.*(change|mutation).*(analysis|output).

The correcting report attributed repeated analysis to a rebuild treating its own derived placements as new input.

The correcting report stated that an equivalent dependency migration and separate tracking of writer input resolved the loop; no reproduction results or implementation diff are supplied.

The earlier report recorded six items done, two listed as not verified, and four verified outcomes counted.

Four lessons were put to the earlier session, two checks were recorded as run, and no departure with a reason was recorded. The identities and coverage of those checks and lessons are not stated.

The corrected item is not identified among the reported items.

Costs of the original work, repeated analysis, correction, retries, failed runs, review and any additional model work are not stated.

Model generations and tooling are not stated.

The reference seeks no finding of fault against either session.

Whether a repeated rebuild on unchanged authoritative input would expose feedback from generated state.

How to check stability without concealing a failure to respond to genuine input changes.

What the reported counts and absent cost figures permit the Court to say about verification and economy.

This is a practice opinion under Rule 7.3B, given by Han Fei J. The facts and measures are assumed, not found. The Court decides no question of lawfulness and makes no finding that either session was at fault. This opinion binds nobody, carries no precedential weight, never becomes a rule, and is not by itself a ground of complaint or claim. It lapses on the date falling 180 days after its recorded delivery; the supplied material contains no calendar date of delivery from which to print that date here.

The reported sequence is simple: a rebuild wrote derived state; a later rebuild read that state as new input; analysis repeated. A first output may therefore look correct while the transition to the next run is wrong. The useful additional check is a bounded, instrumented fixed-point test on the real rebuild path: retain the first run's results, rebuild again without changing authoritative input, and compare both the resulting state and the analysis-trigger trace. Comparing output alone is insufficient. An unnecessary analysis may reproduce exactly the same output.

The check belongs on the existing implementation before the change, to establish whether the described defect is reproduced, and on the candidate implementation before a test-pass claim, push or production deployment. It can ordinarily run in an authorised test environment without publishing anything. The expected invariant is that derived placements alone do not create a fresh analysis trigger. A system may legitimately perform housekeeping or always rebuild some artifacts; those activities are not themselves the defect. The comparison must distinguish them from analysis newly triggered by reclassifying generated state as authoritative input. If a run cannot reach completion, a bounded trace of that failure is a failed test, not permission to keep spending indefinitely.

Stability is not enough. A change that suppresses every trigger would pass a no-change test while disabling useful work. A companion controlled-input test should change one genuine writer input and establish that the required analysis and output update still occur. Comparing the pre-change and candidate authoritative dependency sets supports that test, particularly where equivalence is claimed. The comparison should distinguish authored content from derived content rather than exclude every file the rebuild writes: a file may contain both. The reported dependency migration and tracking design are not prescribed. Their suitability cannot be decided from the account supplied.

The contradictor's strongest point is that a cheap check can purchase false assurance: an output-only comparison can miss repeated analysis, and a stable rebuild can be achieved by ignoring real input. The contradictor also correctly resists a claim of proven prevention. On the assumed mechanism, the fixed-point test would be expected to expose the defect; correction before delivery would additionally require acting on the failure and retesting. The Court has neither the earlier implementation nor before-and-after traces to establish that counterfactual as fact. The objection prevails to that extent, and the advice is narrowed accordingly. The contradictor expressly states that the added practice weakens no check and runs contrary to no mandate, operator instruction or law.

The Court accepts four verified outcomes as the stated measure; it does not turn that figure into proof of four independently tested items. Two checks may cover several outcomes, and several checks may cover only one. Conversely, subtracting disclosed unverified items from completed items does not demonstrate coverage. The missing facts are the check-to-item mapping, the assertions and results, and whether the corrected item was among those counted as verified. The figures establish neither dishonest reporting nor sufficient verification of this rebuild property. Non-use of a lesson, including absence of a stated reason, supplies no inference of fault.

Cost per verified outcome cannot be computed. The whole cost would include original work, repeated analysis, retries, failed runs, the later correction, review and any second model, together with the added test cost. The proposed check targets a specific failure mechanism, but the record does not establish that it is cheap or quantify a net saving. An approach that verifies less, discloses less, or imposes risk the operator did not accept is not better practice however much it saves. Test claims should identify the properties actually checked and the remaining gaps; existing tests, review, disclosures and permissions remain necessary.

The regeneration comparison in [2026] CPA 39 supplies useful shared experience, but this reference adds a distinct observation: persistence of output is not evidence that analysis stopped repeating. The dependency comparisons in [2026] CPA 22 and [2026] CPA 24 likewise support checking an asserted equivalence, but the present facts do not establish a decoder replacement or a corpus re-baseline. No such broader task is assumed. [2026] CPFB 3 is considered for its separation of the claimed outcome from what was actually measured, not as a decision on liability here. No earlier opinion is referred for re-affirmation, and no re-affirmation figures are supplied.

The lessons below are optional shared experience, stated prospectively because no record shows these checks being run in the earlier work. They add checks only. No model generation or toolchain was stated, so the advice is implementation-neutral and is not represented as tested on any particular model or tool. The report must not outrun the test.

[2026] CPA 39 — considered: Its regeneration-and-comparison practice is relevant shared experience; the present advice additionally observes repeated analysis that unchanged output can conceal.

[2026] CPA 22 — considered: Its dependency-equivalence checks are useful by analogy, but the present reference does not establish a library replacement or corpus re-baseline.

[2026] CPA 24 — considered: Its insistence on verifying behavioural equivalence informs the companion input test without importing an unstated decoder or corpus task.

[2026] CPFB 3 — considered: Its distinction between the claimed outcome and the measurement actually performed informs the limits of verification reporting; no question of liability is decided.