Skip to content

ADR-v2-106 — Checksum-Validated Data Entry

Status: Accepted (2026-07-02) · Wave M Context links: itn, temporal, [[adr-v2-091-spoken-table-csv]], slotfill, [[adr-011]]

Context

Wave M research (#2) — in a reference-number field, dictating "credit card / IBAN / ISBN / …" runs the spoken digits through the field's check-digit algorithm and refuses/flags a value that fails, catching a Whisper digit slip before it lands. Entity ITN, Temporal Normalizer, Spoken Table, and Slot-Filling format numbers; none verifies them. STT's single most damaging silent error is a wrong digit in an account/ID number, and a checksum is the only way to catch it without a human re-read — no dictation product does this. Anchor: check-digit systems are designed to catch transcription errors — Luhn (ISO/IEC 7812) catches ~98% of single-digit + adjacent-transposition errors; IBAN MOD-97 (ISO 13616) >99%; ISBN-10/13, Verhoeff, ISO 7064.

Decision

Add an opt-in Checksum-Validated Data Entry: [checkdigit] enabled=false. Pure cores in checkdigit/validate.py: validate(digits, scheme) → (ok, normalized) for luhn | isbn10 | isbn13 | verhoeff (unknown scheme → ValueError), and suggest_fix(digits, scheme) → single-digit-edit candidates that pass. Pure integer arithmetic; no dependency. OFF by default.

Consequences

  • Catches the most costly silent STT error (a wrong ID digit) before injection, and proposes fixes.
  • Pure arithmetic → fully testable against known-good/known-bad numbers.
  • Distinct from ITN/formatting (verify vs format).
  • Privacy (ADR-011): local arithmetic; the number never leaves the machine (note: sensitive numbers are validated, never logged — honours redaction).
  • Caveat: covers the common schemes (extensible); off by default.