[ericvanlaar]

The false positive problem is where I'd push hardest on determinism. In change detection work with Sentinel-2, the most damaging false positives came from seasonal artifacts - snow, agricultural bare soil, cloud shadows - that a well-tuned spectral mask catches reliably before any intelligence layer sees them. SCL cloud masking + NDVI thresholds eliminated a whole class of noise that an LLM would otherwise have to reason its way around.

The principle that worked: anything with a known spectral or geometric signature should be filtered deterministically upstream. Reserve the model for the genuinely ambiguous cases where multiple signals conflict. You probably already know this intuitively from your "most of the system is deterministic plumbing" framing - in my experience the filter boundary can be pushed further upstream than you'd expect.