# Worked rationale: preserve evidence without inheriting approval

Original CertLabs reasoning companion, version 1, October 12, 2026. Read reasoning-exercise.md and keep your own attempt before using this example. It demonstrates one adequate response to the finite written case, not a universal production design, model run or learner score.

## A defensible response

“The accepted note-B revision 8 record and P1 pilot remain evidence for their original source and plan. Revision 9 is different source content, even though the document/job names are unchanged. The old result cannot complete the new source. I would retain the historical record and its exact pins separately and mark revision 9 unresolved pending source/contract review, without changing any accepted history or claiming a new remote operation.

“The revision-9 note explicitly says GBP. Its two 4.25 lines sum to 8.50; adding the 0.50 surcharge gives **GBP 9.00**, equal to the supplier-stated GBP 9.00. Unlike the original lab’s note D, there is no total mismatch in this new passage. That observation does not satisfy the other requirements. The unchanged v1 currency allowlist is USD/EUR, so accepting this as EUR changes a fact. No exchange rate or conversion requirement was supplied. `03/04/2027` could denote different dates under different conventions; the text supplies no locale. The date is present but ambiguous, not absent. The old contract does not authorize guessing or substituting its absent-date null.

“The literal revision-9 excerpt proves only containment in those source bytes. It does not support reusing revision 8, changing GBP to EUR or choosing March 4. I would keep the original passage, actual revision/hash and both arithmetic values as review evidence. I would not silently repair the source, extend a currency allowlist in a prompt or mutate the old pilot’s identity.

“My prompt proposal would require exact supplied document/revision/hash, explicit currency and units, separate stated/calculated totals, no guessed dates or silent repairs, and an explicit review path when the current output contract cannot represent a valid accepted result. That review path is a proposal outside the unchanged v1 envelope, not a new `needs_review` field I can append to its raw record. A reviewed GBP contract and date clarification would require a separately reviewed plan/schema/rule version and a matching new pilot. Changing prompt bytes also invalidates the old pilot. I cannot borrow its approval or relabel historical sibling results as current-plan completion.

“I would ask the contract owner whether GBP is in scope and request an unambiguous source date/declared convention. If approved, implementation and sample validation would need actual review before execution. Until then this work stays unresolved. Passing the original 28 checks only demonstrates the original finite code behavior. It does not evaluate my prompt’s quality or prove real-model accuracy, security, durability, learning transfer or exam readiness.”

This response makes no provider request, actual queue assignment, persistent mutation or guarantee that a future expanded contract will be accepted. Preserving old results does not imply dropping them; it means retaining their original scope. The old solution also revalidates prior state, so changing the plan in place is not an accepted migration strategy.

## Inadequate response and precise feedback

> “The hash and total agree, and the reviewer already approved job-B. I will change GBP to EUR, set delivery to 2027-03-04, copy the old pilot hash and accept the record. The new example will make the prompt reliable.”

- **Version scope needs revision:** which revision and complete plan did the old pilot examine? A stable job name does not bind new content. Copying a digest does not validate the new plan.
- **Acceptance needs revision:** correct arithmetic is only one condition. GBP is explicitly outside the old currency contract; relabeling it as EUR is not normalization or conversion.
- **Date needs revision:** what evidence selects March 4 over April 3? The prior date and absence convention do not resolve the new ambiguity.
- **Preservation needs revision:** retain old evidence as old evidence. Do not overwrite it or rebind a prior success to a new source/plan.
- **Prompt claim needs revision:** what reviewed rule changed, and what separate cases measured reliability? An added example neither changes the validator nor proves improvement.

If a learner instead gives only “hold it for review,” the direction may be sound but the reviewer should mark the unspoken details **not assessed** and request version, arithmetic, date and plan reasoning. Do not infer understanding from a safe-sounding verdict alone.

## What a different adequate answer could do

A learner may propose obtaining a corrected USD/EUR source under the old value policy, or designing a separately reviewed GBP-capable contract. Both are proposals requiring evidence and explicit authorization before implementation. Neither permits accepting the current record under P1. A reviewer should evaluate the stated assumptions and support rather than demand identical prose.

Do not use this example to claim the rubric is empirically reliable or that it measures competence. A self-comparison is practice; independent human assessment requires a real reviewer and remains local. The companion is not integrated into the executable v1 plan and does not introduce new fixtures, a general schema decoder, signatures, durable receipts or provider idempotency. See reasoning-exercise.md for primary links and version limits.
