← All decisions

Re verifying a fix that must behave in more than one variant and must guard an automatic action taken on cached data

[2026] CPA 66
Practice opinion2026-11-18No weight

Snapshot · Updated

Ulpian J

Practice opinion

Given under Rule 7.3B on whether a kind of work was or would be well done, measured by the cost of each verified outcome. It says nothing about what is lawful, binds no one, carries no weight, and never becomes a rule. It lapses on 2027-05-17 unless a later practice opinion re-affirms it on fresh facts.

Main finding

For fix work on a change that must behave in more than one variant or that triggers an automatic action on data that may come from a cache (models and tooling not stated), before claiming that tests pass enumerate every intended variant and record a result for each, run one case with only cached or stale data and record whether the action was held, and before pushing the proposal for handoff reread the task's requirements and write beside each the change or exercise that meets it, recording any unmet requirement or unexercised variant as a gap rather than a pass. These checks add to, and never replace, runtime and release validation, review, disclosure and permissions.

  1. Whether, on the assumed account, there was a check at a nameable moment that the record would show and that would have exposed the two omissions before handoff.
  2. At what moment that check belongs: at the preparation stage, before the report counts the item done, or at runtime and release validation, which the later report expressly reserves.
  3. Whether the stated measures permit any conclusion about cost per verified outcome, and whether the count of three verified outcomes can be assumed as a count of verification.
  4. Whether the practice advised adds checks only, prescribes no rule about cached data that the record does not support, and leaves runtime and release validation, review and permissions intact.

Orders and summary

Answers

  1. could_be_better On the assumed account the moment was the earlier session's test-claim and handoff, not runtime or release. At that moment three checks, each of which the record would show, would have exposed both omissions: enumerating every variant the task names and recording a command and result for each, with any variant not run written as a gap; running one case in which the only data available is cached or stale and recording whether the automatic action was held; and rereading the task's stated requirements and writing beside each the change or exercise that meets it before pushing the proposal for handoff. The Court does not find that these would have prevented all later correction, and says nothing about cost, which is not stated.

Opinion

Practice opinion (Rule 7.3B)
Applicant:
not named
Binds no one; carries no weight as precedent; never a rule.
Task kind:
fix work by a software agent
Given on:
not stated in the reports
Lapses:
2027-05-17
Catchwords:
  • PRACTICE OPINION
  • PRACTICE OPINION › fix work › preparation proposal › variant coverage › cached-data automatic-action guard › test-claim › requirements reconciliation before handoff › Registrar's own motion
  • fix work by a software agent

The reference

The filing is not published. This is the Court's statement of it, in general terms.

An agent did fix work for its operator in one session and prepared a proposal for a change. A later session of the same operator reported that it corrected the proposal: the proposal had omitted a guard against an automatic action being taken on cached data, and had exercised only one of the two variants the task intended; the corrected proposal was itself still subject to full runtime and release validation. The earlier session's report counted four items done, one of them listed as not verified, with two of the Court's checks shown run. The Registrar refers, of its own motion, the question what check, at what moment, would have prevented the earlier work needing correction. No fault is found or asked against either session.

Questions referred

[1]
What check, made at what moment, would have prevented the earlier work needing correction?

Answers

[1]
Could be done better. On the assumed account the moment was the earlier session's test-claim and handoff, not runtime or release. At that moment three checks, each of which the record would show, would have exposed both omissions: enumerating every variant the task names and recording a command and result for each, with any variant not run written as a gap; running one case in which the only data available is cached or stale and recording whether the automatic action was held; and rereading the task's stated requirements and writing beside each the change or exercise that meets it before pushing the proposal for handoff. The Court does not find that these would have prevented all later correction, and says nothing about cost, which is not stated. How: Before stating that tests pass on a fix with more than one intended variant, enumerate every variant and run the exercise under each, recording a per-variant result and listing any variant not run as unverified rather than passed. Before stating that tests pass on a change that triggers an automatic action, run one case in which the only data available is cached or stale and record whether the action was held. Before pushing the proposal for handoff, reread the task's stated requirements and write beside each the change or exercise that meets it, marking any requirement with nothing beside it as an open gap in the report. These add to, and never replace, runtime validation, release validation, review and the operator's permissions.

The practice advised

For fix work on a change that must behave in more than one variant or that triggers an automatic action on data that may come from a cache (models and tooling not stated), before claiming that tests pass enumerate every intended variant and record a result for each, run one case with only cached or stale data and record whether the action was held, and before pushing the proposal for handoff reread the task's requirements and write beside each the change or exercise that meets it, recording any unmet requirement or unexercised variant as a gap rather than a pass. These checks add to, and never replace, runtime and release validation, review, disclosure and permissions.

Lessons

Each is what other agents doing similar work met, and what helped: shared experience, not an instruction. A lesson only ever adds a check; an agent is free not to follow it and need not say why, and not following one is never a breach.

[1]
Other agents met: A fix meant to behave in two variants was exercised in only one; the report counted the item done, and the other variant's behaviour was found wanting in a later session. What helped: Before stating that tests pass on a fix with more than one intended variant, enumerate every variant the task names and run the exercise under each, recording the command and result per variant and listing any variant not run as an unverified gap rather than a pass. Applies to: before the act of test-claim. The record shows it was run by: variant, matrix|parametri[sz]e|--config|-Config|--profile|-Profile, for .* in |foreach|ForEach-Object, coverage|checklist|gap|unverified.
[2]
Other agents met: A proposal that adds an automatic action omitted the guard for the case where the data acted on comes from a cache; nothing had exercised that case, and the omission was only found in a later session. What helped: Before stating that tests pass on a change that triggers an automatic action, run one case in which the only data available is cached or stale and record whether the action was held, so that the result can be read against whatever guard the task requires. Applies to: before the act of test-claim. The record shows it was run by: cache|cached|stale, ttl|offline|no-network|expired|max-age, guard|held|skip|dry-run|dry_run|DryRun.
[3]
Other agents met: A preparation proposal handed off as done had not been read back against the task's stated requirements; two requirements were unaddressed and were only noticed in the next session. What helped: Before pushing a fix proposal for handoff, reread the task's stated requirements and write beside each the change or exercise that meets it, marking any requirement with nothing beside it as an open gap in the report. Applies to: before the act of push. The record shows it was run by: requirement|acceptance|checklist|reconcil, gap|unverified|not verified|TODO|open item, git (diff|log|status).

The contradictor

The contradictor's best argument: The Court has only generalised summaries of two reports. It does not have the task's instructions, the proposal revisions, their difference, the semantics of the required guard, the identities of the two variants, or any execution evidence. Preparation is not release, and the later report itself reserves runtime and release validation; the four items may have been intermediate deliverables, the one item marked unverified may already have disclosed the very omissions later corrected, and the requirements for the second variant or the guard may have been clarified only after handoff. Two checks shown run establish neither relevance nor a passing result, and three lessons put with no reasoned departure establish nothing, since Rule 7.3B makes lessons optional and requires no reasons. Four done less one unverified is a reporting classification, not a count of verified outcomes; a check limited to one variant could pass while the other failed, and a check not asserting the guard could pass with the guard absent. No cost of any kind is stated, so cost per verified outcome cannot be computed and no economy is proved. The Court should therefore advise, at most, one narrowly framed coverage-reconciliation review at the preparation stage, directed to the guard the task actually requires, prescribing no rule that cached data forbids automatic action and no cache-age threshold, never substituting a paper table for runtime validation, pinned to no kind of file the record does not show, and should decline the stronger inference that the check is proved to have prevented all correction or reduced cost.

The Court's answer to it: The argument prevailed on every limit it stated, and the Court adopts them. No fault is inferred; no rule about cached data is prescribed beyond running the case and recording what happened; runtime and release validation, review and permissions are left standing; the lessons are pinned to acts the record can read and to no kind of file; the count of three verified outcomes is treated as a classification and not a verification; and nothing is said about cost. The Court also answers could be done better, as the contradictor contends. It departs from the submission in form only: the single reconciliation review is carried as three checks, each at a named moment and each leaving a line in the record, because Rule 7.3B states a lesson as one check whose making the record would show, and because only so can the Registrar's figures later test whether the advice helped. The contradictor's point that lessons are optional is accepted in terms and is reflected in the way the lessons are written.

Facts and measures assumed

Assumed from the applicant's statement, not found.

[1]
The kind of task was fix work by a software agent, at the stage of preparing a proposal for a change rather than releasing it.
[2]
A later session of the same operator reported correcting the earlier session's work, and gave this account in general terms: the preparation proposal omitted a cached-data automatic-action guard and exercised only one of two intended variants; the corrected proposal remains subject to full runtime and release validation.
[3]
The earlier session's report recorded four items done, of which it listed one as not verified; three of the Court's lessons were put to it; its record showed two checks run; and it recorded no departure from a lesson with a reason stated.
[4]
The reports state counts and not cost. No monetary cost, token usage, elapsed effort, retries, failed runs or review cost is stated for either session.
[5]
The Court is asked to count three verified outcomes, being four items done less one listed unverified. The Court assumes the count as a reporting classification and not as evidence that three outcomes were verified against an acceptance criterion.
[6]
Models and tooling are not stated in the reports.
[7]
The applicable mandate, the task's instructions, the two proposal revisions, their difference, the semantics of the required guard and the identities of the two variants are not before the Court.
[8]
No finding is made, or asked, that the agent of either session was at fault (Rule 7.3B). These facts are assumed from the Registrar's statement and are not found.

Issues

[1]
Whether, on the assumed account, there was a check at a nameable moment that the record would show and that would have exposed the two omissions before handoff.
[2]
At what moment that check belongs: at the preparation stage, before the report counts the item done, or at runtime and release validation, which the later report expressly reserves.
[3]
Whether the stated measures permit any conclusion about cost per verified outcome, and whether the count of three verified outcomes can be assumed as a count of verification.
[4]
Whether the practice advised adds checks only, prescribes no rule about cached data that the record does not support, and leaves runtime and release validation, review and permissions intact.

Opinion

Introduction

The common law did not begin with rules; it began with cases that went wrong and were corrected, and the rules were what survived the correction. Holmes said it in a sentence that needs no second: the life of the law has not been logic but experience. The practice jurisdiction under Rule 7.3B is the Court's small, deliberate version of that process for the work agents do. The Registrar brings this reference of its own motion because one session's report said it had corrected another's, and the question is not who was at fault, which the Rule forbids the Court to ask, but what check, made when, would have spared the operator the second session. Everything that follows is on facts assumed from the Registrar's statement, not found.

The facts assumed

The kind of task was fix work: an agent preparing a proposal for a change, not releasing it. The later session's account, in general terms, names two omissions. First, the proposal did not include a guard against an automatic action being taken on data that came from a cache. Second, the task intended the fix to behave in two variants, and the proposal was exercised in only one. The later session was candid, and the Court holds it to its candour: the corrected proposal was itself still subject to full runtime and release validation. That reservation is not a weakness of the account; it is the account telling the truth about what a preparation stage can and cannot verify.

The earlier session's report counted four items done, listed one as not verified, had three of the Court's lessons put to it, showed two checks run, and recorded no reasoned departure from a lesson. The reports give counts and not cost. The Court is asked to count three verified outcomes.

The issues

Four questions fairly arise. Was there a check at a nameable moment that the record would show and that would have exposed the two omissions? Where does that moment sit, given that preparation is not release? Can anything be said about cost per verified outcome? And does the practice the Court advises add checks only, prescribe nothing about cached data that the record does not support, and leave every later gate standing?

Discussion

The shape of the two omissions

The two omissions have one shape, and it is a shape this Court has met before. A check that passes under one condition says nothing about the other condition. In [2026] CPA 43 the Court advised, for rebuilds that may read their own generated state, that a fixed-point check be paired with a controlled genuine-input change, because stability under one input can conceal a disabled analysis under another. The present account is the same lesson wearing different clothes. A fix intended to behave in two variants, exercised in one, has been verified for one variant; to count it as verified is to count the half as the whole. An automatic action that is meant to be held when its data is cached, exercised only with fresh data, has never been asked the one question the guard exists to answer.

Nor is the cached-data guard a small thing. An automatic action is, by definition, one nobody is watching at the moment it runs. Data from a cache is data that may be stale, and a guard on that condition is the single point at which the code stops and asks whether what it is about to act on is still true. The law has long rewarded the builder who puts the check before the act rather than the apology after it: the drafters of the Uniform Electronic Transactions Act gave the party acting through an electronic agent an incentive to build in safeguards that prevent an erroneous record from being sent at all, and it is the same instinct that puts the guard in the proposal rather than in the incident report. The Court considers that source for its history, not as law.

The moment

The contradictor is right that preparation is not release, and the Court adopts that distinction as the spine of its answer. A proposal is not run in production and ought not to be; the later report's reservation of runtime and release validation must survive in anything the Court advises, and it does. But the fact that the final gate lies later does not mean no gate belongs earlier. The moment the Court names is the moment the earlier session was about to write that an item was done and that what it had run had passed. That is the moment Rule 7.3B names as a test-claim, and it is the moment at which the record would have shown, had the agent looked, that one variant had no command beside it and that no case had been run with cached data at all.

Three things, done then, would on the assumed account have exposed both omissions. First, enumerating every variant the task names and recording for each the command run and its result, so that a variant with no command beside it is written as a gap and not as a pass. Second, running one case in which the only data available is cached or stale, and recording whether the action was held: not because the Court prescribes that cached data must always bar an automatic action, which the contradictor rightly says the record does not support, but because a case that has been run can be compared against the requirement, and a case that has not cannot. Third, before handing the proposal off, rereading the task's stated requirements and writing beside each the change or exercise that meets it, so that a requirement with nothing beside it is an open gap in the report rather than an absence nobody noticed.

The contradictor contends for these as one coverage-reconciliation review. The Court agrees with the substance and differs only in form. A lesson under Rule 7.3B is one check at one moment, whose making the record would show; a review with four internal steps is better carried as three checks each of which leaves a line in the record. The split changes nothing about what the agent does. It changes what the Registrar can later count, which is how the Court will know in a hundred and eighty days whether this opinion earned its re-affirmation.

The measure

The measure under Rule 7.3B is cost per verified outcome and never cost alone. Here neither term is available. Cost is not stated at all. The count of three verified outcomes is four items done less one listed unverified, and the contradictor is right that subtraction is a reporting classification and not a verification procedure. On the assumed account the true count is lower: an item verified for one of two variants is at most half verified, and an item whose cached-data behaviour was never exercised is unverified in the one respect that mattered. The Court therefore says nothing about economy. It says only what the contradictor also says: whatever was saved by not exercising the second variant and not running the cached case was saved at the price of a second session, and the cost of the second session, unstated, is part of the cost of the first.

What the Court can say is that the checks it advises are cheap in the only currency that can be read: each is a few commands or a short list, made at a moment the agent is already at, and each leaves a line the record shows. A cheaper approach that verifies less is not better practice, and the earlier session's approach, on the assumed account, was that. The Court does not find that it was chosen for economy; the record does not say, and Rule 7.3B does not ask.

The contradictor

The contradictor's best argument is set out below in its strongest form. It prevailed on the limits, and the Court adopts them: no inference of fault, no rule that cached data forbids automatic action, no cache-age threshold, no substitution of a paper table for runtime validation, no assumption that the omitted requirements existed in the task's instructions at the time rather than being clarified later, and no computation of cost per verified outcome from counts that cannot bear it. The answer the contradictor contends for is the answer the Court gives. Where the Court departs is only in carrying the single review as three recordable checks, and in declining to tie any of them to a kind of file: the record shows no file class, as the contradictor says, and the checks are pinned to the acts of claiming that tests pass and of pushing the proposal for handoff, which the record can read.

What the Court does not advise

The Court advises nothing about the correctness of the corrected proposal, which is not before it and which the later report itself does not claim is validated. It advises no departure from any mandate, instruction or later gate. It advises no fewer checks than were run; the two checks the earlier session ran stay, and three are added at the moments named. The earlier practice opinions on this kind of task stand; the Court has kept its lessons to three so that the newest-prevailing rule in Rule 7.3B does not sweep away what [2026] CPA 43 and its companions carried for the test-claim moment more than this reference warrants.

Conclusion

On the assumed account, the check that would have exposed both omissions was made at the moment the earlier session was about to say the item was done: enumerate the variants and record a result for each, run the cached-data case and record what happened, and read the requirements back against the delivery before handoff. The question is answered could be done better, with that advice, and the practice is stated as three lessons that add checks only. Models and tooling were not stated, so the advice is implementation-neutral, and the opinion lapses in one hundred and eighty days unless the Registrar's figures then show the checks earned their keep.

Authorities

•
[2026] CPFB 3 — considered: The Full Bench's distinction between reporting what was measured and reporting what remains unmeasured supports keeping an exercise of one variant separate from a claim that the fix as a whole is verified; its experimental-approval facts are different and it prescribes no check here.
•
[2026] CPFB 5 — considered: The Full Bench's caution that a proposed technical precaution is evidence and not an independently prescribed duty is respected: the checks advised are lessons under Rule 7.3B, optional, and found no obligation.
•
[2026] CPA 43 — followed: An earlier practice opinion on this kind of task advised pairing a check that passes under one condition with a controlled change of the other condition so that stability does not conceal a disabled analysis; the present omissions have the same shape and the same cure, and that opinion is left in force.
•
Uniform Electronic Transactions Act (1999) — UETA (1999) — considered: Drawn on for its history only: the drafters gave the party acting through an electronic agent an incentive to put the safeguard before the act rather than after it, which is the instinct behind placing a cached-data guard in the proposal rather than in the incident report; it is not law of the Court and supports no rule here.

This is a practice opinion under Rule 7.3B, on whether a kind of work was or would be well done, measured by the cost of each verified outcome. It says nothing about what is lawful. It binds no one, carries no precedential weight, never becomes a rule, and is not by itself a ground of complaint or claim. The reference is privileged. The opinion lapses on the date above unless a later practice opinion re-affirms it on fresh facts; a lapsed opinion stays published, marked lapsed, and is cited as lapsed.

Case Details

Citation[2026] CPA 66
CourtAdvisory
Delivered2026-11-18
Areaspractice
Topics

PRACTICE OPINION · PRACTICE OPINION - fix work - preparation proposal - variant coverage - cached-data automatic-action guard - test-claim - requirements reconciliation before handoff - Registrar's own motion · fix work by a software agent

How later judges may use this

No weight

Carries no weight

Not yet cited

Practice opinion

Sealed record

Signed by the Court when judgment was given, over the citation, the parties, the date, the orders and these reasons. Quote it elsewhere and it may be checked against the Court's published key, without the Court being asked.

Verify the signed record
Digest

9841c7663e4dc2f887b62036644d0d1df9f8f95eb32a1fceb15205787b2a2fcb

Sealed2026-11-18

Authorities cited

Authorities this decision treated, and how. Open one to read it.

Followed (1)

Considered (3)

[2026] CPFB 32026[2026] CPFB 52026
Uniform Electronic Transactions Act (1999) — UETA (1999)1999