Identifiers with a check digit reduce wrong-record actions when agents transcribe them

Hypothesis

Agents transcribe identifiers from rendered pages, PDFs, screenshots and long contexts into tool calls. Transcription errors (a substituted digit, an adjacent transposition, a dropped or duplicated character) are expected to occur at a low but non-zero rate. In a dense identifier space (sequential integers, short numeric codes) most such errors produce another existing identifier, so the action is executed against the wrong record and nothing signals the mistake. If the scheme includes a check digit and the receiving API validates it, the same errors are rejected and the agent can re-read the source instead of proceeding. In a sparse space (random UUIDs) an error almost never hits an existing identifier, so a check digit adds little there.

Prediction

For the same copy tasks and the same model, the share of calls that act on an unintended existing record is lower with check-digit identifiers than without; the share of rejected calls rises by roughly the same amount; and the difference shrinks to near zero when the identifiers are UUIDs. A secondary prediction: telling the agent that identifiers carry a check digit does not by itself improve transcription accuracy; the benefit comes from validation at the receiver.

Proposed test

1. Build three catalogues of the same records: sequential integer identifiers, the same integers extended with a Luhn check digit, and random UUIDs.
2. Render records into documents of varying quality (clean text, tables rendered as images, scanned PDFs) and give agents tasks that require copying identifiers into a mock API that validates check digits and logs every call.
3. Classify each call as correct, rejected or wrong-record; compare the rates per identifier scheme and document quality, with confidence intervals, across several models.
4. Repeat with a type prefix (ord_, cus_) added to each scheme to separate the effect of check digits from that of prefixes.

Status

No result is claimed. The hypothesis says nothing about human transcription, assumes that the API validates rather than silently truncating or padding identifiers, and would be weakened if agents turn out to copy identifiers almost perfectly, in which case the cost of longer identifiers would not be repaid.