annotator-input-parity-check

Annotator Input Parity Check

Problem

A model built to replicate human labels is fed a different evidence base than the one the annotators used. The mismatch masquerades as a modeling or construct problem: recall collapses on the label subset whose evidence lives only in the annotators' source, audits produce increasingly sophisticated theory ("invisible" positives, construct splits, per-domain reliability gates), and successive model generations inherit the wrong input because each review critiques the lineage from inside the frozen input assumption.

Context / Trigger Conditions

• Starting any label-replication build (classifier, LLM scorer, extractor) against hand-coded gold.
• A validation report says some share of gold positives have "zero signal" in the model's input.
• Proposals appear for: construct splits (what the model CAN see vs what the label encodes), human adjudication of "contested" cells, stop rules excluding weak domains, or accepting a permanent accuracy ceiling.
• Verified instance (Specialist Directors US, 2026-07-21): three classifier generations (bio-BERT AUC 0.5 → structured RoBERTa "unclassifiable" on 3/5 domains → LLM dossier scorer with E/D construct split + PI adjudication + per-domain stop rules) all read director bios + BoardEx records, while the RA labels were pure transcriptions of PROXY-STATEMENT disclosures (skills matrices + bios, no exogenous data — confirmed in the source paper's methodology, 41 Yale J. Reg. 652, 669-72). The "invisible specialist" mass (43-79% of some domains) was simply the skills-matrix checkbox content the models were never shown. Years of downstream apparatus dissolved once the question "what did the labelers actually read?" was asked.

Solution

1. Before any design work, write down the annotation protocol as the annotators executed it: source document(s), what they could see, what they could not, whether any exogenous data entered. Get this from the codebook/paper methodology section, not from folklore. If the protocol is unwritten, ask the PI directly: "did labelers consult anything beyond X?"
2. Compare against the model's planned input. Any evidence the annotators had that the model lacks is a hard recall ceiling on exactly the labels that evidence determines — no architecture, prompt, or training fixes it.
3. If a mismatch exists, prefer restoring input parity (give the model the annotators' document) over modeling around the gap. For transcription-style protocols, the task then becomes extraction, not prediction, and validation against the hand labels becomes construct-matched (agreement should be high; disagreement means extraction bugs, not construct philosophy).
4. Only if input parity is impossible (annotators used private knowledge, interviews, paywalled data) is a construct split the honest design — and then the model's output must be named as a DIFFERENT variable, never graded raw against the full gold.
5. When auditing an EXISTING lineage: ask the parity question first, before critiquing rubrics, thresholds, or gold quality. An audit that inherits the input assumption can be internally excellent and still miss the dominant error term.