# Revision sample and quality variables

Each audit CSV row is one previously released extraction record, identified by `claim_id`. Several rows can belong to one document. `doc_id` is the released transformed identifier; it is not the source text. Original disease/factor/relation fields retain the HCC v1.0 meanings and limitations.

| Variable | Meaning |
|---|---|
| validation_status | Explicitly states lexical support only, not semantic validation. |
| sample_group | Discovery, temporal holdout, COVID comparison, or one of four general PMID-status/period strata. Do not pool these as a representative population sample. |
| frame_n | Eligible document count for that stratum before hash selection. Discovery has no new sampling-frame denominator. |
| pmid | Public PubMed identifier when positive; zero indicates no supplied PMID link. |
| input_sha256 | SHA-256 of the exact recovered extractor-input UTF-8 text. Empty when unavailable. Does not expose or reconstruct that text. |
| disease_phrase_supported | Whether the complete normalized disease-token sequence occurs in the input. Not concept validation. |
| factor_phrase_supported | Whether the complete normalized factor-token sequence occurs in the input. Not exposure validation. |
| keep_joint_phrase | True only when both literal entity checks pass. Empty when the input is unavailable. A passing relation can still be wrong, negated or unstated. |
| keep_joint_token60 | Comparison rule: for each entity, raw normalized substring presence or at least 60% token overlap; both entities must pass. |
| reason | literal_entities_present, disease_unmatched, factor_unmatched, disease_and_factor_unmatched, or input_unavailable_or_ambiguous. |
| explicit_covid_label | In `grounding_audit.csv`, whether the disease field matches the stated COVID-19/SARS-CoV-2 lexical diagnostic. |
| archive_records | Number of recovered archived records for a sampled document. Repeat copies are not independent inputs. |
| unique_inputs | Number of distinct text strings among recovered archive records. Zero means missing; more than one would be ambiguous. |
| recovery_status | unique_input, missing, or ambiguous_inputs. Missing records are retained in sampling/recovery denominators. |
| input_characters | Number of characters in recovered pre-tokenization text. The model may further truncate it. |
| input_covid_family | Case-insensitive result of the explicitly documented COVID-19/SARS-CoV-2 expression on archived input; not comprehensive controlled-vocabulary concept detection. |
| exact_tuple_set_match | Despite the short column name, compares sorted tuple lists and preserves duplicate multiplicity. Five fields: disease, factor, relation, epistemic stance, direction. Not accuracy. |
| lower_95 / upper_95 | Percentile document-cluster bootstrap limits for conditional row retention, expressed as fractions. Not clinical performance intervals. |
| case_id | Stable blinded-review case identifier. Does not contain an expert judgment. |

Dates are recorded publication dates; they are not infection onset dates. Study/evidence labels are generated fields, not verified evidence grades. Original whole-release pair totals remain original context fields; they are not recalculated as time-local features by this sampled correction.