Proof Port: ZMM01
The seal
SHA-256 of the sealed bundle (cases and the old transaction's answers): 9e3220f78683f4e61a1c77611ac2e57df019fdcc84e9dec69b6ab55c5c9a49f5.
It was committed in runs/ZMM01/SEAL.md before the port agent started; the bundle itself is published
next to it now. Check: shasum -a 256 runs/ZMM01/sealed/bundle.json.
The old transaction answered every case as it did when sealed.
What the leak check flagged
- the sealed intent 'price with a thousands separator and no decimals' (c061) also appears as the port agent's p052 'a price with a thousands separator and no decimals'. Judged: same idea, not a copy: different article id (L4 vs ART01), price (1.234 vs 1.000) and function code (GO vs /00), the agent's own intent style, and no tool call touched the sealed cases or the vault; both agents took 1.000-style input from the same brief.
Every difference
| Case | What it checks | Difference |
|---|---|---|
| c018 | Exit pressed on the selection screen with no article id | The old screen said 'Formatting error in the field ZRET_S_ART-ARTICLE_ID; see next message', 'Make an entry in mandatory field "Article"'; the new app said nothing. |
| c043 | base unit that is not a unit of measure | The old screen said 'Unit QQ is not created in language EN.'; the new app said nothing. |
Mutants
| Change to the port | Result |
|---|---|
| line 149: '<=' became '>' | caught by 44 cases |
| line 80: the check 'IF lt_variant_upd IS NOT INITIAL' switched off | caught by 33 cases |
| line 150: message text 'Le prix doit être supérieur à zéro' became 'Le prix doit être supérieur à zérx' | caught by 6 cases |
What was checked and what wasn't
Checked, for every case: the error messages (by text) and every row of ZRET_T_ARTICLE, ZRET_T_ART_VAR after the old screen and after the new app, each from the same starting data. No calls into the old transaction were found in the port or its driver.
Not checked:
- Batch input proves rules, not display. What a screen shows (names, greyed fields, titles) is only seen through what it accepts and what it saves.
- A success message sent just before the transaction ends never reaches batch input, so for saves the rows are the evidence, not that message.
- Batch input stops at the first error, so retyping after a refusal on the same screen is not tested.
- F4 value help does not run under batch input.
- Timestamps written by either side are compared as present or empty, not by value.
- Messages are compared by their text when they are errors. Warnings and information messages are not compared.
How it ran
Files: runs/ZMM01/extract/ (what the factory read), sealed/ (cases and answers),
port/ (the new objects' sources, driver, the agent's report), transcripts/ (both agents, cleaned of the machine's tool list),
grade.json, mutants.json.