Deriva — Derive the Algorithm

← Projectsdocument-intelligence / L1Row Contract & Validator~12 min

one move: make one row trustworthy

A document classifier cannot learn from a row with no stable identity, empty text, absent label, or evaluation reservation. Implement the validator that decides whether a raw row may enter the training release. Never silently discard a row.

Required API

def validate_row(row, reserved_ids=None):
    return (ok, reason)

Behavior

row must be a dictionary
id must be a non-empty string
text must be a non-empty trimmed string
label must be present and allowed (bug, question, feature)
reserved evaluation IDs are rejected before training
every rejection has exactly one stable reason
the input row is never mutated

Examples — tap to reveal

Constraints

reasons are fixed strings: malformed, missing-id, missing-text, unknown-label, missing-label, reserved
a valid row returns (True, None)

Opening…