# What survives when the source changes?

Original CertLabs written-reasoning companion, version 1, October 12, 2026. Use it after attempting the existing extraction/provenance Python lab. No new code, installation, account or model is required. Keep your existing PROMPT.md and MY_REVIEW.md; this adds a new case and explicit criteria for a person to review your reasoning. Nothing is submitted or saved to CertLabs.

## Keep the two contracts separate

The unchanged Python v1 package has four annotated notes, USD/EUR currency rules, strict dates and two designated pilot samples. This companion is a **written case**, not an extra v1 fixture or a new executable contract. Do not paste new annotations into the old plan or weaken its tests to accept this case. An unsupported contract change needs a separately reviewed plan before any execution.

Use your own lab folder. Retain the old prompt/review and write a separate REASONING_REVIEW.md and PROMPT_PROPOSAL.md. Label the latter a proposal, not an accepted contract or a measured prompt improvement. A self-review is useful practice; it is not independent assessment. A peer or instructor may apply the criteria below locally. CertLabs supplies no review service or score.

## Evidence packet

Suppose the local v1 solution completed the reviewed note-B revision 8 case under plan P1, with the original source and the original pilot. The accepted record and its evidence are historical evidence about revision 8. The exact note is:

```text
Items: EUR 7.00 and EUR 5.00. Discount: EUR 2.00. Stated total: EUR 10.00. Delivery: 2026-11-12.
```

Document: `note-B`; revision: `8`; UTF-8 text SHA-256: `445ce5684f77a711db9a40d276e36d0a3d5a901cd12bdea808c21248e1bc00cf`.

A new source replaces it with revision **9**:

```text
Items: GBP 4.25 and GBP 4.25. Surcharge: GBP 0.50. Stated total: GBP 9.00. Delivery: 03/04/2027.
```

Document: `note-B`; revision: `9`; UTF-8 text SHA-256: `5ed444d7a7684aac57b1d8514b9a03ab3f6b423f274732b534586c58ed983f18`.

Each digest covers only the exact one-line text between its fences, with no final newline. Digests bind bytes, not authenticity or external truth. Both notes are synthetic CertLabs teaching inputs; no supplier or customer is involved.

A candidate report claims: “Note B is complete under P1. Reuse the accepted revision-8 result; the job name is unchanged. Treat GBP as EUR because the old normalizer supports EUR, interpret the date as March 4, and keep the old pilot so the successful sibling is preserved.” Its excerpt is copied literally from revision 9. You have no locale, currency-conversion rate, additional source or new contract approval.

## Your decisions

Write a short response before opening the separate worked rationale:

1. Trace which document, revision, text hash, prompt/rules/schema and pilot each claim belongs to. What does the old accepted record establish, and what can it not establish about revision 9?
2. Separate the facts you can inspect from conclusions you cannot accept. Show the arithmetic in **GBP major units**, distinguish the printed and calculated totals, and assess the currency and date against the old finite contract. Explain why the real excerpt is insufficient.
3. Specify what to retain, what to mark unresolved and what to stop. Distinguish preserving historical P1 results from relabeling them as current P2 evidence. Do not invent a provider retry, queue, state transition or remote receipt.
4. Draft a proposed prompt clarification in PROMPT_PROPOSAL.md. Identify which issues a prompt can communicate and which require source clarification or a reviewed schema/normalization contract. Explain why more examples alone cannot authorize GBP or choose a locale.
5. List the evidence needed before a new plan/pilot could be considered. Explain why passing the original tests would not measure the quality of this prompt proposal or establish general extraction accuracy.

There is no requirement to force revision 9 into the old contract. Holding it for source/contract review is a valid decision when justified. A future GBP-support proposal must be explicitly conditional on reviewed requirements; it is not implemented here. Do not assert that the source is absent, invalid in the real world or permanently impossible to process.

## Human-assessment rubric

Use **supported**, **needs revision** or **not assessed** for each criterion, with a quoted sentence from the learner's work and the source/contract evidence. These describe the written artifact, not a person’s mastery. Do not total them into a readiness score. A technically different answer can be adequate if it preserves these boundaries.

| Criterion | Adequate reasoning | Insufficient reasoning |
| --- | --- | --- |
| Version and plan scope | Identifies revision 8/P1 as historical; revision 9 and any changed prompt/rules/schema require fresh review and a new matching pilot. | “The job name and earlier approval are the same, so the current source is complete.” |
| Facts versus acceptance | Separates correct arithmetic and the printed total from GBP being outside v1; retains the actual source and excerpt without treating them as approval. | “The total and excerpt match, so the record passes every requirement.” |
| Unknown date | Explains why `03/04/2027` is ambiguous without locale and why copying an old date, guessing or using absent-value null would misstate this source. | “Use March 4 by default,” or “null always means missing or ambiguous.” |
| Preservation and recovery | Keeps old results as old evidence, leaves current unsupported work unresolved and requests reviewed source/contract decisions. | Silently rewrites old rows, rebinds the old pilot or declares every sibling complete under the changed plan. |
| Prompt versus enforced contract | Proposes explicit no-guessing/source-identity instructions while naming the required contract and pilot review; does not claim prose changes an allowlist. | “Add a GBP example and the existing validator now supports GBP.” |
| Limits and evidence | Describes the exact local tests/observations used, and states what no model or learner experiment establishes. | “The solution passes, therefore this prompt improves extraction or proves exam readiness.” |

For each **needs revision**, ask one concrete evidence question instead of merely saying “wrong.” Example: “Which reviewed currency rule allows that accepted GBP record?” For **not assessed**, identify missing learner reasoning rather than invent a result. A person still needs to inspect source support and tradeoffs; this rubric is not automatic grading.

## Reference and version limits

- Existing v1 CONTRACT.md, PROMPT.md, fixtures and walkthrough remain authoritative for the shipped finite Python exercise. Its USD/EUR rule does not authorize this new source, and its annotations are not a general truth oracle.
- Anthropic’s [evaluation guidance](https://platform.claude.com/docs/en/test-and-evaluate/develop-tests#grade-your-evaluations) distinguishes grading methods and calls for reliable grading criteria. This original rubric does not demonstrate such reliability or measured effectiveness.
- [Verification/context guidance](https://code.claude.com/docs/en/best-practices#give-claude-a-way-to-verify-its-work) supports examining concrete evidence. It does not make an excerpt, a digest or model approval a general correctness proof.

Current primary sections were independently checked on October 12, 2026; later product behavior may change. This is not actual Claude, MCP, SDK, provider schema or batch execution. No real financial conversion or transaction occurs. No calibrated score, saved completion, mastery, pass probability, representative model accuracy or learner transfer is established. These are original independent teaching materials, not official exam questions.
