Metadata is the experiment - what conversion destroys
Lesson 5 of 6 · What the model can and cannot see
1 · Learn the move · Fact-check list
The pixel size, the channel-to-fluorophore mapping, the integration parameters, the acquisition gain, the units - that is not paperwork around the data, it is the experiment. Conversion is where it dies: a stack converted without physical pixel size can never again support a distance measurement; peak areas without integration parameters are not comparable across runs; a signal column that lost its units is a column of naked numbers. The discipline is a fact-check list run around every conversion. Before: extract a metadata manifest from the source file - every field that gives the numbers meaning. After: verify each field in the converted output, and give each a state: VERIFIED, it survived and matches; UNVERIFIED, the output format has no place for it and it now lives only in your manifest; FAILED, it changed or vanished. A FAILED field is a stop. An UNVERIFIED field is a file you must now keep forever alongside the data.
I am converting [file] from [format] to [format]. Step 1: list every metadata field in the source that gives the numbers meaning - spatial calibration, units, channel mappings, acquisition settings, processing parameters - as a manifest table. Step 2: for each field, state where it lands in the target format, or NOWHERE if the format cannot carry it. Step 3: give me a verification checklist to run after conversion: FIELD, HOW TO READ IT IN THE OUTPUT, EXPECTED VALUE, STATE (VERIFIED / UNVERIFIED / FAILED). Fields marked NOWHERE go in a sidecar file that travels with the data permanently.
2 · Your turn. You write the prompt
Your five-year-old confocal stacks are being migrated off the vendor's dying software into OME-TIFF, and a labmate's earlier migration produced beautiful images that no longer know their own pixel size - every measurement made on them since is suspect. Write a prompt that builds the manifest and the after-conversion checklist so your migration cannot fail the same silent way.
Remember: the AI sees only your prompt, not this page. If the situation isn't in your prompt, it doesn't exist.
Optional. These shape the output when you run your prompt below, not your score.