Published science gets corrected slowly. A flawed paper can sit in the literature for months or years before an erratum or retraction appears. During that time other researchers cite it, and some build their own methods on it. The usual remedies all wait for the formal notice: look up retraction status, then flag the papers that cite the retracted one. This proposal borrows a different stance from cloud capacity planning. There, systems forecast demand far enough ahead to cover the time it takes to boot a machine, so capacity is ready when the burst arrives. We apply that idea to the scientific record. Instead of checking papers after they are retracted, we forecast which papers will pick up the most dependent citations during the expected correction delay. We then spend a fixed checking budget on those papers first.
In two sentences: rank papers by expected exposure, meaning the probability that a paper's content is wrong multiplied by the number of method-level citations it will gain before a correction could plausibly land. Then audit the top of that ranking under a fixed compute budget, and test retrospectively whether this prevents more downstream harm than ranking by popularity or by risk alone. The new element is neither a retraction predictor nor a citation forecaster. It is the coupling of the two, with a forecast horizon set to the correction lag, in the way provisioning systems set their horizon to boot time.
Core Claim
Let $p_i$ be the estimated probability that paper $i$ has a content-level flaw. Let $D_i(t_0, t_0+h)$ be the number of citations it receives between the decision time $t_0$ and a horizon $h$ in which the citing paper relies on it for a method or core claim, as opposed to background or critique. The exposure score is
$$E_i(h) = p_i \cdot \mathbb{E}\left[D_i(t_0, t_0+h)\right],$$
and the auditor checks the top-$k$ papers by $E_i(h)$, where $k$ is set by a verification budget. We claim three things. First, at equal budget, exposure ranking prevents substantially more dependent citations than ranking by raw citation momentum or by $p_i$ alone. Second, the gain is largest when $h$ is matched to the empirical lag between publication and correction, and it shrinks when $h$ is much shorter or longer. This is the analogue of the finding in provisioning that the forecast horizon must match operational lead time. Third, counting only dependent citations, rather than all citations, is the part that makes the forecast useful.
The motivation comes from three findings in the record-integrity literature. The window of risk between an error and its notice is real; the swapped-label correction in Author Correction: Mitochondrial haplotypes affect metabolic phenotypes in the Drosophila Genetic Reference Panel is one documented example. Assessing the Impact of Retracted Articles on the Reliability of Citing Literature shows that binary "cites a retracted paper" flags are too coarse, and uses SciBERT argument mining to separate citations that depend on a retracted paper from those that criticize or merely mention it. Retractions relevant to COVID-19 finds that over 30% of retraction notices give no clear reason. That makes retraction labels noisy and argues for restricting to content-related categories. All three papers are retrospective. None asks where scarce checking effort should go before a notice exists.
How This Differs From the Obvious Combination
The obvious idea at this intersection is a blockchain-backed, versioned registry of papers and corrections with predictive scheduling of continuous-integration checks. In effect, it would put Treat societally impactful scientific insights as open-source software artifacts on a ledger. That paper maps Git, continuous integration and "intellectual debt" onto science and prototypes the Alexandria platform. We deliberately do not build this.
The blockchain-engineering literature itself argues against it. Maintenance-related concerns for post-deployed Ethereum smart contract development finds, from 131 papers and 165 developers, that immutability makes corrective maintenance painful. It also finds that the workarounds, such as proxy delegation and redeployment, add security and trust problems. Engineering Blockchain-based Software Systems describes the same missing safety net. Putting a record whose whole purpose is to be corrected onto an append-hostile substrate imports exactly that pain. Tamper-proofing is also not the bottleneck; the delay before anyone checks is.
What we take from the blockchain direction is a diagnosis, not a substrate. The same survey counts 31 offline, pre-deployment checking tools against only 7 online monitoring tools. That mirrors science, where effort is concentrated in pre-publication review and post-publication monitoring is nearly absent. The smart-contract field's response to that imbalance is runtime monitoring prioritized by what is at stake. Our proposal is the scientific counterpart.
Research Program
The work has three components.
Forecaster. The first component is a two-headed forecaster computed at a fixed time after publication, for example six months.
The risk head estimates $p_i$. It uses only information available at $t_0$: venue and field base rates, author and correction history where permitted, and, most importantly, the stance of early citations. The Öztürk et al. pipeline lets us detect early critical or refuting citations, and we expect these to carry the strongest early warning. The model is gradient-boosted trees over features at several time scales, following the design of CAFE. CAFE combines coarse hourly trend features with fine-grained minute features to forecast desktop demand 30 to 60 minutes ahead. We reuse that structure with monthly and weekly citation windows.
The exposure head forecasts dependent citations over horizon $h$. Following CAFE's emphasis on under-prediction rate, we train it with an asymmetric loss. Underestimating exposure on a flawed paper is costlier than overestimating it.
Horizon. The horizon $h$ is not a free hyperparameter. We estimate the distribution of publication-to-notice lag from the retraction data, per field. Then we set $h$ to the lag quantile that the budget can cover. The provisioning literature cited by CAFE motivates this rule: a forecast is useful only if it arrives earlier than the action it triggers.
Auditor. The third component is the auditor that consumes the ranking. It is an evidence-grounded LLM check that retrieves the paper's key claims and looks for internal inconsistencies or external counter-evidence. It follows the answer-conditioned counter-query design of CounterRefine, which treats an initial answer as a hypothesis and retrieves evidence against it. We separate retrieval from judgment because AutoSupervision shows that judging without grounding fails badly on long manuscripts.
We are explicit that the auditor is weak. AutoSupervision reports verification accuracy peaking around 0.501 even for frontier models. For that reason the triage ranking is the primary contribution, and it is evaluated independently of auditor quality. The auditor's precision enters only in a second-stage analysis.
Budget. Auditor jobs are long, tool-calling workloads. We cost the budget in GPU-hours and run the auditor on a serving stack with tool-aware scheduling of the kind in AugServe. Serving efficiency is infrastructure here, not a claimed contribution.
Baselines
The naive reference is reactive propagation. Do nothing until a formal notice appears, then flag every citing paper. Run this way, its prevented exposure before the notice is zero by construction. A stronger reactive variant flags only dependent citers using Öztürk et al.'s argument-role classifier. That variant tells us how much of the remaining exposure is concentrated after the notice, where reactive tools already act. No reported performance number was found on Lacuna for the Öztürk et al. framework, so we will run it ourselves.
The strongest proactive competitor is citation-momentum ranking: audit the papers with the highest forecast total citations over $h$, ignoring risk. It is the strongest competitor because highly cited papers are where any error does the most damage, and popularity is easy to forecast. If it matches exposure ranking, our risk head contributes nothing.
The second competitor is risk-only ranking: audit by $p_i$ regardless of expected reach. This is what a standalone retraction predictor would do.
All rankers get the same features available at $t_0$, the same gradient-boosted model family, the same hyperparameter search budget per head, and the same audit budget $k$.
Two ablations isolate the key ingredients. One replaces dependent-citation counts with raw counts. The other replaces the lag-matched horizon with fixed horizons of one month and five years.
Beating the competitors means one specific result. At budgets covering 1%, 5% and 10% of papers, exposure ranking must prevent a significantly larger share of post-$t_0$ dependent citations to later-corrected papers than citation-momentum ranking does. We will use paired bootstrap confidence intervals over publication cohorts.
No reported result was found on Lacuna for any proactive retraction-exposure ranker. The closest reported numbers are CAFE ranking 1st or 2nd in 97% of 30-minute workload-forecasting test cases, and AutoSupervision's overall 0.637 for GPT-5.5 and Claude-Opus-4.8. Neither is directly comparable, and we will not present them as such.
Why not just…?
Why not just rank by citation count? This is the real threat. Retraction risk may be close to uniform once field is controlled for, in which case popularity carries all the signal. Our bet is that early critical citations sharply raise $p_i$ for a minority of papers that are not yet heavily cited. Exposure ranking would reach those before momentum ranking does. The citation-momentum baseline above settles the question directly. If the confidence intervals overlap at every budget, we will report that triage reduces to popularity and drop the risk head.
Why not just look up retraction databases and propagate flags? Lookups cannot act during the window of risk, and over 30% of notices lack reasons (Retractions relevant to COVID-19). Lookups remain the right tool after a notice is issued. The reactive baseline measures what fraction of total dependent exposure accrues before the notice. If that fraction is small, the proactive problem is not worth solving, and the experiment will show it.
Why not just audit every paper with a cheap model? AutoSupervision shows that small models given no retrieval step collapse at locating evidence in long manuscripts, with GPT-4o-mini scoring 0.085 on grounding. Adding retrieval raises its overall score only from 0.266 to 0.403. Universal cheap auditing would therefore produce mostly noise. We will test this anyway, as a cost-matched arm that spends the same GPU-hours spread thinly across all papers.
Why not put the record on a blockchain? The immutability-versus-maintenance evidence in the smart-contract maintenance study suggests a ledger would make corrections harder. The failure we target is the delay before anyone checks, not tampering.
Proposed Experiments
Data. The core experiment is a retrospective backtest built on the Öztürk et al. dataset, which has 3,193 retracted articles, 4,979 citing papers and 7,424 citation pairs labeled with 29 content-related retraction reasons. We will augment it with publication and notice dates. We will also add a matched control set of non-retracted papers drawn from the same venues and years, which we must construct ourselves. Corrections to figures and data, of the kind in the Drosophila author correction, form a secondary positive class to test whether the method extends beyond retractions.
Decisive checks. For each cohort we freeze features at $t_0$, rank, apply budget $k$, and measure prevented exposure: dependent citations accrued after $t_0$ to positive papers in the audited set, divided by all such citations.
The first decisive check is the comparison with citation-momentum ranking described above.
The second checks the horizon claim. Prevented exposure, plotted against $h$, should peak near the empirical correction-lag quantile. A flat curve would falsify the analogy with provisioning lead times.
The third is a temporal holdout: train on older cohorts and test on newer ones, to see whether the ranker drifts.
Auditor evaluation. A second-stage experiment measures auditor precision on the top-ranked set. It runs the counter-query auditor on AutoSupervision episodes to calibrate how often a flag would reach a human with usable evidence. That number tells us whether triage plus automated checking is actionable or whether triage should route straight to human reviewers.
References
- Author Correction: Mitochondrial haplotypes affect metabolic phenotypes in the Drosophila Genetic Reference Panel
- Assessing the Impact of Retracted Articles on the Reliability of Citing Literature: A Mixed-Based Proposal
- Retractions relevant to COVID-19: does the retraction rate jump during the pandemic?
- Treat societally impactful scientific insights as open-source software artifacts
- Maintenance-related concerns for post-deployed Ethereum smart contract development: issues, techniques, and future challenges
- Engineering Blockchain-based Software Systems: Foundations, Survey, and Future Directions
- CAFE: Adaptive VDI Workload Prediction with Multi-Grained Features
- CounterRefine: Answer-Conditioned Counterevidence Retrieval for Inference-Time Knowledge Repair in Factual Question Answering
- AutoSupervision: Closing the Feedback Loop in Scientific Workflows with Grounded Revision Verification
- AugServe: Adaptive Request Scheduling for Augmented Large Language Model Inference Serving