Publish blind hard-negative audit and holdout limitations

Summarizes independent EVREN model votes; labels remain unreviewed. Clarifies that the public challenge set is not a hidden benchmark holdout.

PROGRESS.md

@@ -9,6 +9,14 @@ Updated: 2026-09-25
 - Added `tests/test_dataset_validator.py` and `scripts/validate/validate_dataset.py` to validate generated main and hard-negative JSONL drafts locally. Dataset artifacts remain excluded from Git, so CI uses small synthetic fixtures; the full local datasets were also checked directly.
 - Latest checks: 8 pytest tests passed; ontology validation passed (40 concepts, 40 relations, 80 lexical mappings); main draft passed (1,000 rows, balanced 10 intents × 4 languages, 840/80/80 group-safe splits); hard-negative draft passed (80 rows).
 
+## Hard-negative audit
+
+- Audited all 80 contrast examples with blind, shuffled intent classification from the three models currently listed by EVREN (`deepseek-v4-flash`, `glm-5.3`, `qwen3.8-flash-next`). The models did not see the current label or contrast target.
+- The 2-of-3 vote matched the current label on 80/80 examples; all three models agreed with the current label on 74/80. Six rows had one dissenting model. No labels were changed; model agreement is not human ground truth.
+- The TF-IDF baseline classified 31/80 contrast examples incorrectly, while the independent EVREN majority supported their current labels. This points to difficult distinctions in the challenge set, not automatic proof of bad labels.
+- The challenge set and labels are public in this draft, so it must not be treated as an undisclosed benchmark holdout. A future benchmark needs a new, properly reviewed private/held-out challenge set.
+- Full local report and raw model audit are under ignored `data/generated/evren-v0.1/hard_negative_blind_audit.*` files.
+
 ## Completed
 
 - Expanded the hospitality ontology from 4 to 20 intents, with department/action relations and TR/EN/DE/RU lexical mappings.
@@ -34,6 +42,8 @@ The EVREN dataset page currently disables the **Metin** modality for this accoun
 - `data/generated/evren-v0.1/candidates.jsonl`, `model_calls.jsonl`, `multilingual_review_calls.jsonl`, `hard_negative_calls.jsonl` — ignored local generation/review audit trails.
 - `scripts/build_multilingual_pilot.py` — resumable generation, extraction, validation, and group splitting.
 - `scripts/generate_hard_negatives.py` — contrast-set generation and validation.
+- `scripts/audit_hard_negatives.py` — resumable blind second-opinion audit across available EVREN models.
+- `docs/hard-negative-audit-summary.md` — vote totals, dissenting records, and provisional label boundaries.
 - `scripts/evaluate_tfidf_baseline.py` — development-only baseline.
 - `scripts/validate/validate_dataset.py`, `tests/test_dataset_validator.py`, `requirements-dev.txt` — pytest and JSONL validation harness.
 - `huggingface/dataset-card.md` — dataset card source.

README.md

@@ -22,6 +22,7 @@ This repository contains synthetic hospitality-intent data generated for HOSIA.
 - `hosia_v0.2_hard_negatives.jsonl`: separate 80-record contrast challenge set with `split: hard_negative_eval`; excluded from the main training data.
 - `tfidf_baseline_report.md`: development-only word/character TF-IDF + LinearSVC report. Main test accuracy 0.988, macro-F1 0.987; hard-negative foil accuracy 0.613. These results describe only this synthetic pilot and are not generalizable benchmark claims.
 - EVREN models assisted with intent/entity checks and translation validation/repair. No human review has been completed.
+- In a blind second opinion, the three currently available EVREN models had 2-of-3 majority agreement with existing labels on 80/80 challenge examples; all three agreed on 74/80. This is model agreement only, not human ground truth; six examples need adjudication.
 - Every record remains `reviewed: false` and `allowed_for_training: false`.
 - A basic email and phone-pattern scan found no matches. It is not a full privacy review.
 
@@ -35,6 +36,8 @@ The planned research concerns multilingual hospitality intent classification, en
 
 Synthetic LLM output and model review can contain wrong labels, awkward or culturally mismatched language, annotation omissions, semantic duplicates, and personal or sensitive information. Basic pattern screening cannot detect every privacy risk. The baseline uses a small synthetic test split and should not be read as evidence of real-world performance. All examples, entities, translations, splits, privacy, and rights need human review before any training or benchmark use.
 
+The hard-negative file and its labels are public in this draft; it is not an undisclosed holdout and must not be used to claim benchmark performance. A future benchmark needs a separately reviewed, properly held-out challenge set.
+
 The EVREN dashboard currently does not enable text dataset upload for this account; EVREN was used for generation and validation, not as a dataset host.
 
 ## License and provenance