SF-007 · Empty value is silent

Status: stable ｜ Family: B · Uncounted Absence ｜ Confidence: high ｜ as of 2026-09-19

One line. A required field is empty, and nothing anywhere breaks — the page loads, the API responds, the reference resolves.

Symptom

A record is structurally valid and semantically hollow. Every consumer works:

Nothing is broken. Nothing is complete either. The record is findable but not searchable — a distinction no error channel expresses.

Concrete instance: a researcher profile with identity, works and affiliation all present, but biography, keywords and external links empty. The profile is live, the DOI resolves, the ORCID record is valid. And a retrieval system has nothing to match on beyond the name — so it is effectively invisible for every query except the person’s own name, which nobody searches.

The general shape: the field is required for the record’s purpose but optional in the record’s schema. Absence is therefore legal, and legal absence is not reported anywhere.

Why it is silent

A way for “absent” to count is missing — and here the schema actively agrees with the omission. The validity check and the usefulness check are different checks, and only the first one exists.

There is a second reason this survives every automated audit: the failure mode is compatible with every success signal. Page up, status 200, identifier resolves, no drift, no alert. An audit that asks “is anything broken?” answers no. Only an audit that asks “is the expected thing present?” can see it.

Minimal reproduction

record = {"id": "example", "name": "X", "biography": "", "keywords": []}

assert validate_schema(record) is True      # passes — all fields optional
assert http_get(record["id"]).status == 200 # passes — page renders

print("record OK")                          # → and it is, structurally
# ...but a search over the record's subject matter returns nothing.

Observed: page renders, no error, no warning — Expected after fix: a validation failure naming the empty required-for-purpose fields

Self-check

Enumerate the fields that a retrieval system would need in order to match this record against a query — then check whether each is populated and whether anything asserts it.

field           populated?   asserted non-empty?
name            yes          yes (schema)
biography       no           no               ← the hole
keywords        no           no               ← the hole

Anything with populated=no, asserted=no is this bug.

Fix

Split “required” into two explicit classes, and assert the second one.

REQUIRED_FOR_VALIDITY  = ["id", "name"]          # schema-level
REQUIRED_FOR_PURPOSE   = ["biography", "keywords", "links"]   # discoverability-level

empty = [f for f in REQUIRED_FOR_PURPOSE if not record.get(f)]
if empty:
    raise SystemExit(f"record {record['id']}: required-for-purpose fields empty: {empty}")

Two design principles that prevent the class:

  1. Optional in the schema is not optional in practice. Encode the purpose-level requirement somewhere assertable, even if the schema cannot express it.
  2. Make the check part of the pipeline that publishes, not a separate audit. An audit that must be remembered will not be run; an assertion in the publish path cannot be skipped.

Where a field’s emptiness is legitimate for some records, express it as an explicit declaration (fields_not_applicable = [...]) rather than as an absence — so that “not applicable” and “forgotten” remain distinguishable. This is the same move as SF-006’s registry: turn a default into a declaration.

Negative control Expected result
Blank a required-for-purpose field the publish path exits non-zero, naming the field
Remove the assertion, re-run the record publishes silently — proving the assertion was doing the work

中文要点