Caller-supplied populations
Preserve the query, graph and evaluation protocol. Read
the population contract before creating a supplied table.
Snowflake sources stage local Parquet for this API. Do not switch to another connector or
claim that a local bundle is a source table.
Resolve the supplied source table and required column semantics. These constructors consume
existing tables; they do not authorize creating labels, renaming data, cleaning nulls or
changing splits or weights. Use a user-approved prepared table when the source lacks required
fields.
Validate schema, row types, keys, timestamps, labels, weights and splits and report the exact
bad field. Do not silently drop rows, cast IDs, normalize weights or replace manual splits: a
field that needs changing is reported and asked about. Split labels and weight derivations follow
the same rule: show the mapping or the exact formula, the columns it reads and the resulting
distribution from a read-only query, and stop for the user's explicit approval. An approval given
before that is shown is not approval. Only then apply it in a read-only connector.sql(...)
source, never as a write to the source table, an UPDATE, a CREATE TABLE AS or a derived table
(references/populations.md). Validation still applies to the result.
Before any code, reply with the plan and stop for confirmation: the table and its fields, the
weight and split columns, what you validated, the constructor and that it reuses the confirmed
pq and snapshot, and any shape the host adapter says is unverified there. For scoring, add
that validation does not guarantee every row is scored, that no id is dropped, cast or remapped,
and that the output's entity and anchor coverage will be compared with the population. A bad
field gets the ask-back instead.
Confirm the snapshot used to create the population and the snapshot passed to fit/predict
belong to the intended graph. Rebuild populations after incompatible graph changes.