The Anatomy of a No
A maintenance planner rejects a recommendation because its inspection belongs to a removed assembly. Follow the correction, then try to break it.
An illustrative maintenance-planning case. All records and identifiers are fictional. This is a study of evidence and software design, not an account of a client incident or a maintenance instruction.
The recommendation is ready. Close work order WO-417. The inspection is complete, the equipment tag matches, and the system has put both records in the review packet.
The maintenance planner leaves it open. She is comparing a small plate on the installed assembly with the identifier on the inspection sheet.
“That inspection belongs to the one we took out.”
The system has found a real inspection for a real piece of equipment. It has attached that evidence to the wrong physical history.
- Evidence retrieved
- Inspection A · revision 4
- Match accepted
- Equipment position P-17
- Inference made
- Inspection applies to installed assembly.
- Missing check
- Which assembly did the inspection cover?
“That inspection belongs to the one we took out.”The planner's objection in this fictional case.
The work order remains unresolved.
Follow the tag into the wrong history
In this illustrative case, P-17 names the equipment position. It stays in the register when an assembly is removed and a replacement is installed. The inspected assembly was S-108. The installed assembly is S-224.
A search on P-17 retrieves the inspection. A comparison on P-17 appears to confirm the match. The same convenient identifier has carried the mistake through two steps.
The planner sees the difference because she asks a question the recommendation never resolved: which assembly does this inspection describe?
That question is more useful than “did the AI get it wrong?” It tells the builder where to look. The search may be working exactly as requested. The failure is in deciding that a shared equipment tag establishes applicability.
The evidence / two records, one shared tag
P-17 names a position.
It does not settle which assembly was inspected.
Inspection complete
Assembly identified on the inspection.
Equipment position: P-17.
Replacement installed
Assembly identified as currently installed.
Equipment position: P-17.
ObservationThe equipment tag matches.
P-17 appears in both records. That observation is supported. It establishes that both documents concern the same equipment position, not that both describe the assembly now installed.
InterpretationThe inspection applies to the installed assembly.
This is the unsupported step in the recommendation. The inspection identifies S-108; the replacement record identifies S-224. Applicability needs evidence beyond the shared position tag.
Missing contextThe equipment's replacement history matters.
The planner brings the change of assembly into view. The records owner must resolve which inspection applies, including any approved configuration relationship. The expert's objection is retained as an objection until that evidence settles it.
I would keep the rejected recommendation beside those records while the planner explains the objection. Do not replace it with a cleaner answer yet. The original packet shows which relationship the system accepted, which evidence it displayed, and what the reviewer had to supply from somewhere else.
The planner's objection is a finding to investigate. She could have read the wrong plate, or a controlled configuration record could establish that the inspection still applies. The records owner needs to resolve that question. Accepting the expert's confidence without the evidence would preserve the same design weakness with a different person supplying the answer.
Write down less than you think you learned
The quick repair is a new instruction: every inspection must match a serial number before a work order can close.
That would make the demonstration look better. It would also teach the system something the case has not established.
Some equipment may use a controlled configuration record rather than an individual serial match. Another class of work may require different evidence. The planner has shown that this inspection does not establish what this recommendation claims. She has not rewritten the company's inspection policy.
The first correction should stay close to that finding. Preserve the rejected link to S-108, the replacement record identifying S-224, and the unresolved question about the replacement's inspection. Keep the work order open in the proposed workflow until an authorized reviewer resolves it.
Revision / write the smallest supported correction
Give the correction an address.
Every inspection must match a serial number before a work order can close.
- Rejected link
- WO-417 → Inspection A / S-108
- Supporting record
- Replacement record r2 → installed S-224
- Scope
- This work order and this unresolved relationship.
That is a useful correction even before the team decides whether a wider change is needed. The next run can explain why the case remains unresolved. Another reviewer can inspect the same evidence. The builder has a specific failure to replay.
A broad note in the prompt loses those advantages. It does not show who resolved the objection, which records supported the correction, or when a change in the equipment would make the note obsolete.
Take the correction somewhere awkward
The rejected case is the easiest test to pass. The team already knows the answer it wants to see.
Try the proposed lesson against a case where only one important condition changes. Keep the same kind of work, but give the equipment an approved configuration record instead of the individual serial convention. Or keep the equipment unchanged and let the records owner supply the missing inspection.
Those cases ask whether the system understands the boundary of the correction. A rule that fixes WO-417 by blocking legitimate work elsewhere has exchanged one error for another.
Challenge the scope / an illustrative test bench
Change one condition.
Does the correction still belong?
The recommendation stays in view. Select a case to inspect what changes in its evidence, reach, and next review.
A second look at the unresolved work order
Operating conditionP-17 still contains S-224.
Evidence availableNo applicable inspection has been established.
The local correction still fits its address. It explains the disputed link and the evidence still needed; it does not manufacture a closure decision.
Another asset uses a controlled configuration record
Operating conditionThe approved identity path is configuration-based.
Evidence availableThat record links the inspection to the current configuration.
The broad instruction would block a legitimate identity convention. WO-417's correction has no authority here. Review this case under its own approved evidence path.
The missing inspection is supplied
Operating conditionAn applicable inspection identifies S-224.
Evidence availableThe records owner can now review that evidence.
The local exception may have served its purpose. Retain its history, but do not let an obsolete note continue controlling the recommendation after its supporting conditions change.
The different-condition case is particularly valuable. The broad serial-number instruction would reject evidence that the local process accepts through another controlled identity path. The original correction should have no authority to do that.
The changed-evidence case exposes a different problem. A case-specific hold can become stale. If the replacement's applicable inspection is located, a permanent “do not close” note will keep obstructing the work after its reason has gone away.
Retain the correction's history, but reassess its eligibility to control the next recommendation. The records owner can review the new inspection and retire or revise the local exception. A retired correction remains useful evidence of what happened; it should no longer behave like a current restriction.
A wider rule needs its own owner
The team may eventually find that the same identity mistake affects a whole equipment family. That supports proposing a broader design change: resolve the installed configuration before deciding whether an inspection applies.
The proposal needs more than the first rejected case. Give it examples of a valid older inspection, a replacement, a missing record, conflicting identities, and an approved alternative identity convention. Say where the rule applies and who is allowed to release it.
The outcome of a test need not be closure. An explicit unresolved question can be the correct result when the evidence is incomplete. What matters is whether the workflow distinguishes supported conclusions from missing information, and keeps the decision with the right owner.
The decision record / what survives
Keep the history.
Reassess the authority of the lesson.
- r1Rejected recommendationPreserve the original claim and the inspection it relied on. Do not overwrite the failure.
- r2Local correctionRecord the disputed link, supporting evidence, scope, owner, and review trigger.
- r3A candidate wider ruleRelease only after tests define where it applies and an owner accepts the change.
- ↳Or a retired exceptionWhen evidence or configuration changes, review whether the local restriction is still eligible to apply.
This is where the work becomes reusable. The team has preserved a disputed relationship, the evidence that settled it, and the cases that define the correction's reach. It can change the search or the model and still ask whether that behavior holds.
In The Second Workflow Is the Real Test, I argued for keeping corrections useful beyond the task that discovered them. The address and the review trigger are what let that happen without turning every local lesson into policy.
Make the next disagreement easier
The maintenance planner has a queue. She should not need to write a specification every time a recommendation makes a bad match.
Let her point at the suspect inspection and record the objection while the system retains the work-order identity, proposed action, and evidence versions. Give the records owner a way to resolve it. Then show the planner what changed.
If the same unsupported recommendation returns on Monday, the feedback loop has failed even if somebody saved her comment. If a new rule quietly blocks unrelated work, it has failed in another direction.
I would want the next reviewer to see a smaller problem than the first one did: the disputed relationship already identified, the source close at hand, and the remaining decision clear.
