Sound Provenance Filtering for Literature-Derived Training Data

Pipelines that turn scientific papers into training data for language models now run at the scale of millions of papers. SPARK alone draws on 3.36 million open-access articles, and VeriSciQA builds figure-based questions from figure-citing text. These pipelines treat every published sentence, figure label and number as fixed truth. The scientific record does not work that way. Papers are retracted (more than 10,000 in 2023, by the count cited in Assessing the Impact of Retracted Articles on the Reliability of Citing Literature). Figures are corrected after publication; the Bevers et al. author correction, for example, fixes swapped labels. And later work builds on findings that have since been withdrawn. None of the current quality filters can catch this kind of error. Model judges, committee votes and uncertainty scores only check whether a generated item agrees with its source. A fluent question built on fabricated data passes every one of them.

We propose that each synthetic item carry a record of exactly which source spans, figure elements and cited works it was built from. A filter then keeps an item only when the correction record shows that none of those inputs has been invalidated. When it cannot tell, the filter says "unknown" and leaves the item out. This borrows a stance from automated planning and database repair. A sound but incomplete approximation never certifies something false, but it may fail to certify some true things. The approach replaces a judge that is known to be weak with a guarantee that is conditional and can be checked: if the recorded inputs cover the item's true inputs, and the correction record is right, nothing kept depends on invalidated content. Where a correction states the fixed value, as with an erratum that swaps axis labels, we patch the item instead of deleting it.

Core Claim

The central hypothesis has a measurable part and a methodological part. The measurable part is that a nontrivial share of literature-derived synthetic data inherits invalidated content, and that this share is not confined to the retracted papers themselves. Much of it arrives through two other routes: element-level errata, and citing papers whose methods or results take a retracted finding as their basis. The methodological part is that a filter which reasons with three values (definitely safe, definitely invalid, unknown) removes this contamination more reliably than the alternatives, at a yield cost we can report. The alternatives are judging by a model, consensus voting and uncertainty scoring. Patching corrected values, rather than deleting items, should recover much of that yield cost.

We state the measurable part first because it can falsify the whole project. If nearly all contamination comes from directly retracted sources, a blacklist of retracted DOIs is enough, and the right paper is a short audit, not a method. The research program is ordered so that this question is settled before anything else is built.

Research Program

The method rests on two lines of prior work. The first is the 0-approximation of Baral et al.. It tracks only fluents (state facts) that are definitely true or definitely false, which makes planning tractable while every produced plan stays correct. The second is the update-based universal repair of Greco, Molinaro and Trubitsyna. It answers queries over an inconsistent database soundly in polynomial time by changing conflicting values instead of deleting whole tuples. We do not claim these theorems transfer formally to unstructured text, since Greco et al. prove theirs only for acyclic equality-generating dependencies. What we take from them is the design pattern: a lower bound on certainty, an explicit unknown, and repair by updating values instead of deleting rows.

The first component is a provenance-recording generator. We re-implement the SPARK recipe (SPARK) on an open-access biomedical corpus. SPARK extracts claim–evidence–derivation skeletons from papers and generates question–answer pairs from them. We change one thing: every skeleton node is bound to character offsets in the source, and to every reference cited inside those spans. We make the same change to the VeriSciQA pipeline (VeriSciQA), recording the figure, the panel and the extracted atomic claim. This provenance is deliberately over-inclusive. An item depends on everything it was generated from and on everything those spans cite. Over-inclusion is what makes the soundness argument work, and it is also the main source of incompleteness.

The second component is a correction knowledge base built from the Retraction Watch categories. Öztürk et al. reduce the 108 categories to 29 content-related reasons, which separate administrative notices from scientific ones. We add publisher errata and correction notices, parsed into structured edits where possible, such as "label A should read B" or "value x should read y". The Informational Scope of Scientific Correction Notices reports that many notices have only administrative scope. This matters directly: an affiliation fix must not invalidate anything, which is exactly what a DOI blacklist gets wrong.

The third component is a citation-function step. It decides whether a citing span relies on a retracted work or criticizes it. We follow the argument-mining scheme of Tailoring Scientific Argument Mining for Scientific Literature Correction, which labels argument units and citation functions and marks a citing result as built on "feet of clay" when it uses a retracted claim as its basis. Öztürk et al. call the opposite case the criticism exception: citations that refute a retracted work are part of the correction itself. Soundness here depends on one-sided errors. Calling a "basis" citation a "critique" would let contaminated items through. The classifier therefore accepts "critique" only above a high-precision threshold calibrated on the Nanobiology argument-mining data, and sends everything else to unknown. Öztürk et al. note that critique citations are rare relative to background citations, which makes this threshold the hardest part of the system to get right.

The keep rule is short. Let $P(s)$ be the recorded input set of synthetic item $s$, and let $\sigma(e) \in \{\text{safe}, \text{invalid}, \text{patched}, \text{unknown}\}$ be the status the correction base and citation-function step give to input $e$. Then

$$ \text{keep}(s) \iff \forall e \in P(s):\ \sigma(e) \in \{\text{safe}, \text{patched}\}, $$

where items with any patched input are regenerated from the corrected value before release. Status propagates along "basis" edges to a fixed depth. A cited-work input inherits the invalid status of a retracted paper unless it is classified as critique. Because the provenance lives with the data, re-auditing when a new retraction arrives is an incremental query, not a rerun of the pipeline. This is a concrete version of the continuous-integration analogy proposed in Treat societally impactful scientific insights as open-source software artifacts. It also shrinks the "window of risk" between an error's publication and its correction.

Baselines

The naive reference point is the unfiltered pool. This is the SPARK-style or VeriSciQA-style pipeline as published, with its own quality filters but no awareness of corrections. On standard benchmarks, SPARK reports an average of 60.83 across GPQA, SciBench and MMLU-Pro for Qwen3-8B fine-tuned on Spark-234K. VeriSciQA reports a +2.05% average gain across CharXiv, MMStar and MathVista. Our filters must not push these numbers down by more than noise. Above that floor, the main measure of success is the retraction-sensitive measure described below.

The direct-source blacklist drops every item whose source paper appears in the retraction database, whatever the retraction reason. It is the strongest cheap competitor and the one most likely to be "good enough". We run it on the identical pool and evaluate it identically. The proposal beats it only if the full filter leaves strictly fewer contaminated items, as judged by human audit, at comparable yield. It must also leave a lower rate of the model asserting retracted claims on our probe set.

Transitive deletion extends the blacklist to every paper that cites a retracted work. It tests whether the criticism exception and citation-function reasoning earn their complexity. It should catch about as much contamination as our filter but cost more yield, especially on the critique-bearing papers that Öztürk et al. argue are the corrective part of the literature.

The strongest learned filters are three. The first is an LLM judge given the item, its source and any retraction or correction notice. This mirrors SPARK's GPT-OSS-120B audit with added correction context. The second is heterogeneous committee voting in the style of Ensemble-Instruct, which filters instructions by agreement across three different open models. The third is CoCoA uncertainty filtering from Stoisser et al., which ranks traces by self-consistency and perplexity. CoCoA reports student accuracy of 0.57 on PerturbQA with 10% filtering, against 0.52 unfiltered, 0.48 for a random 10%, and a ground-truth upper bound of 0.62. All baselines keep the same number of items as our filter, so the comparison is about which items survive, not how many. All fine-tunes use the same base model (Qwen3-8B), token budget and hyperparameter search.

The ablations isolate the key ingredients. The first treats "unknown" as safe, giving a two-valued filter; this tests whether the soundness stance matters. The second removes value patching, so corrected items are deleted instead; this measures what update-based repair recovers. The third removes the critique threshold, accepting the argmax citation function; this measures how fragile soundness is to classifier errors. We have no reported result on Lacuna for any correction-aware filter on literature-derived synthetic data. The comparison numbers above are the only published anchors.

Why not just…?

Why not just blacklist retracted DOIs? This is the real threat, and the first experiment exists to test it. If the audit shows that transitive and erratum-level contamination is negligible next to direct contamination, we will report that and stop. A blacklist also has two known failures it cannot fix. It invalidates content behind purely administrative retractions, and it is blind to corrected-but-not-retracted figures. The audit measures both effects separately.

Why not just have a frontier model audit each item against the correction notice? The evidence is that these models characterize critiques well but verify poorly. AutoSupervision reports 0.754 on characterization but verification peaking around 0.501. Models also tend to accept authors' own assertions rather than checking the evidence. Over 30% of retraction notices omit a clear reason (Cheng et al.), which leaves a judge with little to go on. We still run the judge as a baseline, because a model that sees an explicit notice might do better than on AutoSupervision.

Why not just vote across several models? Voting checks agreement among generators, not facts. On SciFact-Open, different systems retrieve and verify such different evidence that their agreement falls below 0.2 Jaccard, and about 20% of well-studied claims have conflicting evidence. Models that agree with a confidently written retracted paper will vote it through. The matched-retention comparison shows directly whether committee voting removes contaminated items more often than chance.

Why not just use uncertainty filtering? Fabricated results produce confident, low-perplexity text. CoCoA was designed to catch reasoning errors made by the generator, not errors in the external premise. We expect it to remove contaminated items at roughly the base rate. If it does much better, that is evidence that contamination correlates with surface signals, which would be a finding worth reporting in its own right.

Proposed Experiments

The first experiment is the contamination audit, and it decides the project's direction. We join the released VeriSciQA dataset and our provenance-recording SPARK re-implementation against the correction base. We oversample the 3,193 retracted papers and 4,979 citing papers in the Öztürk et al. dataset so that the tails are well estimated. For each item we label the contamination route as direct retraction, transitive basis, or element-level erratum. Annotators confirm each label on a stratified sample, which also yields precision and recall estimates for every filter. The decisive number is the share of contamination that the DOI blacklist misses. We commit in advance to pivoting to an audit paper if that share is small.

The second experiment tests whether contamination matters to trained models. We build a probe set from content-related retractions and parsed errata. Each probe pairs the withdrawn claim, or the pre-correction value, with a matched control claim from an unretracted paper on the same topic. We measure how often a fine-tuned model asserts or prefers the withdrawn version. To separate effect size from detectability, we first inject items from retracted papers at controlled rates. This tells us whether natural contamination rates can move the probe at all, before we compare filters at those rates.

The third experiment is the filter comparison. Every baseline and ablation is trained at matched retention on Qwen3-8B. Each is reported on probe assertion rate, audited residual contamination, yield, GPQA, SciBench and MMLU-Pro (anchored by SPARK's 60.83), and CharXiv, MMStar and MathVista (anchored by VeriSciQA's +2.05%). As a check on the downstream use of science knowledge, we also report SciFact-Open, where MultiVerS reports 52.4 F1.

The fourth experiment is temporal. We build the dataset using the correction record as of a cutoff year. We then measure how many items later corrections invalidate, and how cheaply incremental re-audit finds them compared with regenerating the dataset.

The proposal succeeds if three things hold together. The audit must show material non-direct contamination. The full filter must beat the blacklist on residual contamination and probe assertion rate at no measurable cost on standard benchmarks. And value patching must recover a clear share of the items that deletion would lose.