# Tracing and Filtering Unsupported Biomedical Extractions

Narendra Kumar · Fliar BI · Preprint v2 · 15 September 2026

This study audits 55,545,265 HCC v1.0 extraction rows and tests an explicit lexical safeguard on newly selected documents. It reports computational evidence, not clinically validated relationships. It has not been peer reviewed.

## Main results

- Fixed source weights reduce the apparent 2019-2020 unspecified-study increase from 22.13 to 1.92 percentage points.
- A new temporal sample contains 300 PMID-linked documents dated 2010-2018. Inputs were recovered for 299; none contains the tested COVID-family terms. Their released outputs contain 380 explicit COVID-family rows, all rejected by the candidate safeguard.
- A retained-checkpoint replay matches complete released tuples for 62 of 70 new documents. Its 30 temporal inputs produce 51 COVID-family rows, all failing the safeguard. This localizes unsupported generation in those experiments but does not prove an identical mechanism for every anomaly.
- General-stratum row retention ranges from 77.2% to 86.3%. Confidence intervals resample whole documents and are conditional on successful input recovery. Retention is not precision or recall.
- One of 829 deliberately mismatched-input controls passes the stricter gate. These easy constructed controls are not a clinical specificity benchmark.
- In the post-2020 comparison, 56 of 529 assessed COVID-labelled rows also lack the tested term family in their inputs. Date filtering alone would miss them.

## Read and inspect

- `output/pdf/HCC_scientific_audit_v2.pdf`: 19-page main paper, including references.
- `output/pdf/HCC_technical_supplement_v2.pdf`: 10-page technical supplement.
- `output/pdf/HCC_arxiv_submission_v2.pdf`: one 29-page PDF containing both documents, prepared for arXiv's PDF-only submission route. Preparation does not establish submission or acceptance.
- `figures/`: eleven PNG/SVG figures, with exact source CSVs named in the captions.
- `results/revision_input_provenance.csv`: sample membership, frame sizes, recovery status and input fingerprints.
- `results/hcc_grounding_audit_sample_v2.csv`: all 1,896 sampled rows with quality flags; not a random row sample or a population-proportional sample.
- `results/hcc_lexically_retained_sample_v2.csv`: 1,174 rows passing the literal-entity rule. The file name is not a claim of semantic validity.
- `results/replay_outputs.csv` and `replay_comparison.csv`: all new replay outputs and discrepancies.
- `results/reviewer_a.csv`, `reviewer_b.csv`, `adjudication_worksheet.csv`: blank human-review worklists. Independent human semantic annotation has not been performed.

Public study materials: https://fliarbi.com/assets/papers/hcc-input-audit-v2-20260915/

Original release: https://doi.org/10.5281/zenodo.22754767

Original release mirror: https://doi.org/10.6084/m9.figshare.33752332.v1

Existing illustrative HCC sample: https://fliarbi.com/assets/papers/samples/hcc/hcc_sample.csv

## What was fixed

The revised paper now links the temporal anomaly to archived inputs and a retained-checkpoint replay, measures a rejection rule on a separate sample, preserves missing cases and replay discrepancies, reports coverage costs, and provides an implementable audit sidecar. It does not silently relabel diseases or change historical dates. Original public HCC v1.0 files and original review PDFs remain unchanged.

The safeguard requires both entity strings as complete normalized token sequences in the input. It abstains when either is unmatched or the input is unavailable. This corrects an unchecked-output pathway in the experimental derivative. It does not resolve aliases, negation, cross-sentence reasoning, study design or clinical truth. Production inference services have not been changed by this study.

## Reproducibility

Python 3.12 and versions in `requirements.txt` reproduce the tabulated release analysis and figures. PDF building also uses ReportLab, DejaVu fonts (`fonts-dejavu-core`) and Poppler (`poppler-utils`). Set up a dedicated virtual environment.

```bash
python3 -m venv .venv
.venv/bin/pip install -r requirements.txt
.venv/bin/python analysis/download_inputs.py --data /your/data/hcc_v1.0
.venv/bin/python analysis/reproduce.py --data /your/data/hcc_v1.0 --work /your/scratch/hcc --pdf
```

Allow at least 20 GB of scratch and sufficient RAM for the 4 GB database limit plus Python and operating-system overhead. The 22 released input files are fingerprinted in `inputs.json`; the full repository deposit has 28 files including documentation and examples. Corpus ingestion is resumable. Use that scratch directory only for the verified fixed release. Network reference checks are archived from 15 September 2026; refreshes are a new audit and require review of reference fields and manuscript dates.

To apply the safeguard to authorized source inputs, supply a JSONL containing `doc_id` and `text` for each document. No source passages are written to the output CSV:

```bash
.venv/bin/python revision/annotate_inputs.py \
  --claims results/revision_claims.csv \
  --inputs /your/authorized_inputs.jsonl \
  --audit /your/new_audit.csv \
  --retained /your/new_retained.csv
.venv/bin/python revision/test_grounding.py
.venv/bin/python revision/check_results.py
.venv/bin/python revision/make_figures.py
```

Raw historical input passages and retained model weights are NOT distributed. Input hashes alone cannot regenerate the grounding and replay experiments. Those experiments require access to the archived texts and checkpoint, whose fingerprints are supplied. Public identifiers permit independent source inspection, but current abstracts may differ from historical extraction inputs. All reported aggregate arithmetic, gate code, sidecar decisions and replay comparison outputs are inspectable in this package. This is partial public reproducibility, not end-to-end reconstruction of collection and training.

The paper does not claim independent human annotation or improved clinical precision/recall. Fliar BI produced the corpus and conducted this audit; the producer's interest is disclosed. arXiv submission still depends on actual author-account access, any required endorsement and the author's attestations.

Optional retained-checkpoint replay requires the exact checkpoint, authorized original inputs, an available CUDA GPU and compatible packages recorded in `results/replay_environment.json`:

```bash
.venv/bin/python revision/replay_holdout.py --inputs /your/authorized_inputs.jsonl --model /your/checkpoint
```

The replay script checks checkpoint and selected input fingerprints before generation. Changing inputs or weights creates a different experiment, not a reproduction of this preprint.