A model-params row is one model's parameter counts,
edited a row at a time. The rows are the model
parameter table. A scoring pass stores a row's counts on the warehouse rows it
writes, and a filing warns uncatalogued_model for a model with no row.
A result row has a null model_params_total_millions if the scoring pass
ran before its model had a row. A plan's jobs page
(GET /api/v1/plans/{plan_id}/jobs?results=true) and results file
(GET /api/v1/plans/{plan_id}/results.csv) then take both counts from the row
that its descriptor resolves to. A stored count wins. The fix-model-param
command rewrites the warehouse.
A row is {model_key, total_millions, active_millions, note, updated_at, updated_by};
updated_by is null on a seeded or imported row, and note on a row without
one. An omitted active_millions is the total. note says where the counts
come from: it is trimmed, and an omitted or blank one leaves the row without a
note.
Reading the rows takes read_any_plans; each write takes manage_orgs and a
person, never a service account.
A row keyed by a vault resource id holds the counts of one private model, and
is written by the administrators of the organization that owns the file:
manage_orgs and membership, the line the vault's delete draws
(who may do what). Every other row is server-wide,
since two organizations' plans share one result of a public model, and takes
manage_any_orgs, which writes any row. So does the row of a file a person
owns. A write outside that reach is a 403 with one message, whether the key is
a name, another owner's file or no file.
A vault model's record
(GET /api/v1/vault/resources/{resource_id}) carries
its row and whether the reader may write it, so its page sets the counts.
No new permission names: a row is its organization's administrators' or the server's, and reading all of them is for an account that reads every plan.