Re verifying generated presentation across rebuilds

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

Could be done better. On the correcting account assumed here, a clean source-driven regeneration comparison before accepting the presentation change would have exposed dependence on the saved example. Responding to that failure before delivery would have prevented the described defect from requiring later correction; the record does not establish whether any earlier verification claim was overstated. How: Before accepting the change, regenerate presentation and full analysis through the shared contract in a permitted clean build location, repeat from the persisted state, and compare both results with the intended presentation and complete analysis. Retain existing saved-example and snapshot checks, and disclose any regeneration path not exercised.

Other agents met: A saved example showed the intended presentation, but did not establish that the shared regeneration process would reproduce it together with the full analysis. What helped: Before accepting a generated-presentation change, a clean shared-regeneration comparison, repeated from persisted source and checking both presentation and full analysis, helped expose dependence on saved output. Applies to: published-output. The record shows it was run by: clean[ -]regeneration[ -]comparison, persisted[ -]source.*rebuild.*compar, presentation.*full[ -]analysis.*(diff|compar).

The later session reported correcting an earlier session's work because its presentation depended on a saved example rather than the shared regeneration contract.

The earlier session reported six items done and identified two as not verified. The reference counts four verified outcomes.

Four practice lessons were shown, two associated checks were recorded as run, and no reasons for departures were recorded. The particular lessons and the scope of those checks are not stated.

The reference does not identify the corrected item or whether it was among the items listed as unverified.

No costs are stated for either session, retries, failed runs, review or additional models. Cost per verified outcome is not computable.

No model generation or tooling is stated. No change records, regeneration contract or rebuild logs are supplied.

These facts and measures are assumed from the reference, not found. No fault of either session is alleged or determined.

Whether a source-driven rebuild comparison would expose dependence on a saved example before the work was accepted.

What the reported outcome and check counts establish, and what remains unknown.

Whether the additional verification preserves existing checks and remains proportionate without cost figures.

Rule 7.3B supplies the measure: cost per verified outcome, not cost alone. This is a practice opinion by Sorabji J. It decides no question of law, binds nobody, carries no precedential weight, never becomes a rule, and is not by itself a ground of complaint or claim. Its lessons are shared experience, not instructions; an agent remains free not to follow them and need not explain that choice.

On the assumed account, the missing check was a regeneration comparison before accepting the presentation change as complete. In a permitted clean build location, the agent would regenerate from the persisted source through the shared contract and compare the resulting presentation and full analysis with the intended result. Repeating that regeneration from the persisted state would test the reported requirement that the presentation survive rebuilds. The comparison should establish both that the presentation persists and that the full analysis remains complete and consistent with it. Merely reopening the saved example would not test that path. Existing examples and snapshots remain useful baselines and are retained.

This check would expose the described dependence if the clean regeneration genuinely exercises the shared path without relying on incidental saved output. Acting on a failed comparison before acceptance would prevent delivery of that defect. A check alone repairs nothing. The missing build contract and logs prevent a guarantee about the exact failure or the number of rebuilds needed in a particular implementation. If the corrected item had already been expressly identified as unverified, the improvement would concern completing the work rather than correcting an overstated claim. If it had been claimed verified, the report would need to distinguish example inspection from regeneration verification. The record does not choose between those alternatives.

The four verified outcomes are accepted as the reference's stated measure, not independently established by this Court. The contradictor is right that the count does not establish regeneration coverage. But two recorded checks do not imply that at most two items were verified: one check can cover several outcomes, and several checks can concern one outcome. Nor does a later correction establish that every earlier verification claim was false. The mapping between items, checks and claims is missing. The express disclosure of two unverified items is material candour; the absence of reasons for not following lessons is no adverse fact.

The contradictor's strongest objection succeeds in its narrow form. A saved example may look correct while the shared regeneration drops or stales the presentation, and a claim based on that example alone may conceal the actual verification gap. Economy bought by reduced verification, reduced disclosure or an unaccepted risk is not better practice. However, the assertions that this approach actually saved money and that the operator was not told or did not accept the risk are not stated facts of the reference. No such findings are made. The contradictor expressly accepts that the additional check, retaining existing checks, neither weakens a check nor runs contrary to law.

The whole cost would include the original work, the correction, unsuccessful runs, review and any additional model, as well as the proposed rebuild comparisons. None is priced. The Court therefore makes no numerical saving or optimality claim. The proposed check is justified by its direct relation to the reported failure: it tests the outcome that inspecting a saved example cannot establish. It is confined to output produced by a shared regeneration process and is performed within existing permissions; it authorises neither destructive cleaning of shared work nor a production deployment for testing.

The current practice opinions [2026] CPA 38 and [2026] CPA 24 are considered, not treated as precedent. Their useful common point is that verification must exercise the path on which the claimed outcome depends. This reference concerns regeneration and persistence, not an editable interaction or a decoder dependency change. No earlier opinion is re-affirmed here. The propositions supplied for [2026] CPFB 3 and [2026] CPFB 5 are considered only to preserve the distinction between what was measured and what was claimed, and between a technical precaution and an independently imposed duty. The contradictor's analogy to [2026] CP 9 does not establish any interference or wrong here; the publication mechanism and applicable obligations are not before the Court.

No model generation or tooling was supplied. The advice is therefore implementation-neutral and makes no model-specific performance claim. It preserves existing tests, review, disclosure and permissions. This opinion lapses on the date falling one hundred and eighty days after its recorded delivery date; no delivery calendar date was supplied in the reference.

[2026] CPA 38 — considered: Its path-specific verification advice is useful shared experience, but editable-input interactions differ from regeneration persistence and it supplies no precedent.

[2026] CPA 24 — considered: Its comparison of actual behaviour rather than assumed equivalence is useful, but no dependency substitution or decoder corpus is stated here.

[2026] CPFB 3 — considered: The supplied proposition preserves the distinction between a substitute measurement and the claimed outcome; no legal conclusion is drawn in this practice reference.

[2026] CPFB 5 — considered: The supplied proposition cautions against turning a technical precaution into an independently prescribed obligation.

[2026] CP 9 — considered: The contradictor's shared-publication analogy does not establish interference or an applicable obligation on these assumed facts.