Field notes

Each note follows the same four beats: what was seen, why it happened, what changed, and what is still not covered. They are short on purpose — a note is one measurement and one consequence.

#A fix that reproduced its own defect

• Observed — a request combining three deliverables produced only two of them.
• Cause — the negation list was matched as a substring. The French word for “a” contains the word for “not”, so “a presentation” read as “no presentation”.
• Fix — negation is matched on whole words, splitting on everything that is not alphanumeric, the apostrophe excepted.
• Test — a combined request must yield three distinct outputs.
• Still open — the parser is lexical. A request phrased as a question rather than an instruction is not covered.

#A test that accused the prose

• Observed — a guard asserted that the workbook tool exposes no file path, and it failed.
• Cause — the guard was reading the tool’s human description, found the word “path” in a sentence explaining that there is no path, and failed on its own explanation.
• Fix — the guard inspects the declared input schema keys instead.
• Test — the schema carries exactly six keys, and none of them is a path.
• Still open — nothing checks that the description and the schema agree.

#A file on disk is not evidence

• Observed — a spreadsheet sat in a data room, and it looked like proof that the engine worked.
• Cause — it had been written by a Python agent in an earlier version, before the engine existed.
• Fix — none needed in code; the reporting rule changed. A produced file counts as evidence only when the run that wrote it is identified.
• Test — the engine’s own tests write, reopen and verify, so the engine’s claim never rests on a file someone found.
• Still open — the engine has still never run inside a real agent session.

#A counter that saw the neighbours

• Observed — a test asserting that a blocked card starts no process failed, apparently at random.
• Cause — the process counter was global and picked up harmless helper processes spawned by tests running alongside.
• Fix — the counter only counts vendor processes.
• Test — the same assertion, now stable.
• Still open — the counter is a heuristic on the process name.

#A request that convened the wrong role

• Observed in a live run — a workbook request went to the fundamental analyst alone. Nobody who owns the workbook format was called.
• Cause — routing keyed on the words of the request rather than on the owner of the requested format.
• Fix — a requested deliverable convenes its owner. One format, one owner.
• Test — a workbook request calls the accounting role even with no accounting keyword in the sentence.
• Still open — tested, not yet observed in a live run.

#A hard kill is not a clean quit

• Observed — one orphan process remained after a session, and had to be terminated by hand.
• Cause — the application was stopped with a signal rather than through its own quit path, so its shutdown routine never ran.
• Fix — none. The behaviour is correct: a brutal kill bypasses cleanup by definition.
• Test — the quit path stops every process it started.
• Still open — this is not evidence that a normal quit leaks; it is evidence that a hard kill does.