Validating model output before it touches a record

Between the parse and the DML. Everything that arrives is a proposal, and a proposal is not a value.

1. VideoAllow-listed enumerations, and no free-text statusPreview6 min
2. VideoRange-checked scores, and the confidence that was 1.46 min
3. VideoCapped arrays, and the response that wanted 400 children6 min
4. VideoLength-capped strings, and the field that silently truncated6 min
5. ReadingEverything that arrives is a proposal, not a value5 min
6. LabLab: build the validator15 min

An agent looping single updates wrote a misspelled picklist value and blanked stages across records. Fixed with validation, batching and a backstop rule.

1. VideoAn agent looping single updates, one record at a time6 min
2. VideoA misspelled picklist value, and what it did to every stage it touched6 min
3. VideoServer-side validation against the live schema6 min
4. VideoBatched writes, so one bad value cannot cascade6 min
5. ReadingA validation rule as the backstop nothing else caught5 min

The rule that matters most: an unlabelled claim is the dangerous one, so default to the weaker provenance.

1. VideoAn unlabelled claim is the dangerous oneA value with no source reads as fact. Label every written value, and default to the weaker label when unsure.6 min
2. VideoStoring provenance beside the value, not in a log6 min
3. VideoShowing provenance to the person reading the record6 min
4. VideoDefaulting to the weaker claim when the source is ambiguous6 min
5. ReadingGeneral model knowledge against internal data: telling them apart5 min
6. LabLab: add provenance to three written fields15 min

The enum-constrained single write in the field triage agent, and the ritual that resolves, maps, states and echoes before it commits.

1. VideoThe single write, enum-constrained server-side6 min
2. VideoAn endpoint that rejects anything off-enum, so a wording slip cannot corrupt it6 min
3. VideoResolving the exact record before anything is written6 min
4. VideoMapping the user’s natural phrasing to the enum, and stating the mapping6 min
5. VideoEchoing it back and waiting for a yes5 min
6. LabLab: build the confirmation ritual end to end10 min

A remediation decision costs six figures. The engine only nominates candidates, and a data gap must never read as confirmed.

1. VideoData gaps read amber, never silently green6 min
2. VideoWhy a six-figure decision cannot rest on a candidate flag6 min
3. VideoThe null hypothesis: assume the model is wrong until proven6 min
4. ReadingA human verification queue as the calibration signal5 min
5. LabLab: design the amber state for a missing field14 min

Dropping is the default. A review queue is the exception, and it needs an owner or it becomes a graveyard.

1. VideoDrop, do not store with a warning6 min
2. VideoA review queue with an owner and a decay policy6 min
3. LabThe test that proves an invalid response never reaches the database15 min
4. VideoWhy a swallowed validation error is worse than a loud one5 min
5. ReadingWriting the validator’s test cases before the validator5 min
6. QuizCheck: output validation10 min