# Recorded contract-editing results

Vespper ran this fictional service agreement through its hosted DOCX MCP Server
on 29 September 2026. One edit replaced the bold payment term **30** with **45**.
The tracked run recorded the replacement under **Contract assistant**. The
tracking-disabled run changed the text directly. Both retained the two pending
confidentiality revisions belonging to **Counterparty reviewer**.

| Mode              | Output                                                    | Fixture checks | Result                    |
| ----------------- | --------------------------------------------------------- | -------------- | ------------------------- |
| Tracking enabled  | [Redlined DOCX](results/service-agreement-redlined.docx)  | 15 passed      | Passed with normalization |
| Tracking disabled | [Direct-edit DOCX](results/service-agreement-direct.docx) | 13 passed      | Passed with normalization |

Read the [tracked check report](results/service-agreement-redlined.checks.json)
and [direct check report](results/service-agreement-direct.checks.json). Each
includes input/output SHA-256 hashes, dependency versions, tracking mode, the
API result timestamp, the revalidation timestamp and every check result.

## What changed outside the payment paragraph

These outputs are not byte-identical outside the requested edit. The service
added four font properties in unedited locations:

- `w:bCs` in the two bold table-header runs, matching their existing `w:b`.
- `w:szCs` in the header and footer, matching their existing `w:sz` of 18
  half-points (9 pt).

These properties describe complex-script text formatting. The affected runs in
this fixture contain only ASCII text. The table, header and footer retained
their visible layout in the inspected render. This observation does not prove
equivalence for documents containing complex-script text or inherited styles.

The first verifier required exact unedited body structure and byte-identical
headers/footers, so both initial runs failed those two checks. Those original
reports are retained as [tracked initial checks](results/service-agreement-redlined.initial-checks.json)
and [direct initial checks](results/service-agreement-direct.initial-checks.json).

Verifier version 2 identifies and reports only the matching additions described
above; it does not ignore arbitrary formatting differences. Fifteen offline
regression tests cover ordinary passes, unchanged output, wrong author, removed
revisions, lost bold, missing image, unrelated table changes, wrong tracking
mode, and allowed versus conflicting font additions. The API output files were
not edited to make them pass.

## Reproduce the checks without credentials

From the extracted example directory, using Python 3.10 or later:

```bash
python verify_contract.py service-agreement.docx results/service-agreement-redlined.docx
python verify_contract.py service-agreement.docx results/service-agreement-direct.docx --tracking-off
```

To make a new API call, follow `README.md`. Keep recorded results separate from
your new output. The result can differ between service runs; compare your actual
output and check report rather than copying these recorded outcomes.

## Scope of the visual check

The input and both outputs each rendered as one page using the bundled
LibreOffice renderer. The payment term, existing revisions, table, list, image,
header, footer, spacing and page boundary were inspected. The guide's preview
is cropped from those actual renders.

Native Microsoft Word accept/reject interaction was not verified in these
recorded results. For your application, open the output in Word, accept and
reject only the payment revisions on separate copies, and confirm the unrelated
confidentiality revisions remain pending. Package checks and a LibreOffice
render do not establish every aspect of Word interoperability.

No separate agent model was used in the recorded runs. The optional OpenAI
Agents SDK path is supplied but is not presented as a live-tested agent run.
This is one synthetic fixture with two modes, not a comparative benchmark or a
general formatting-preservation guarantee.