Which Scientific Corrections Need Agreement? Measuring the Consensus Cost of Fixing the Literature

When a published paper is corrected or retracted, someone has to decide which other papers are affected. Today the default is either to do nothing or to treat every citing paper as suspect. Distributed computing has a precise vocabulary for this choice. Some updates commute: they can be applied in any order, by anyone, with no coordination. Others conflict: they require the participants to agree on an order, and in an unreliable network that agreement can stall indefinitely. That stalling result is the Fischer–Lynch–Paterson (FLP) impossibility, which says no deterministic protocol can both stay safe and always finish when messages can be delayed arbitrarily and even one participant may crash.

This note proposes to measure how much of the scientific correction workload falls on each side of that line. It then proposes to build a classifier that sorts corrections onto that line conservatively. The proposal is not to put science on a blockchain. The claim is that a correction can be released to downstream papers without coordination exactly when it is independent of them, in the formal sense used by Byzantine-tolerant replicated data structures. The core research problem is therefore learning an independence relation between corrections and citing papers that errs only toward caution.

Core Claim

Process-commutative distributed objects shows when an object can skip consensus entirely. Operations from different processes must be independent, and applying two valid independent operations in either order must give the same valid state (the "I-diamond" condition). When that holds, a 14-line algorithm over Byzantine reliable broadcast suffices. Byzantine reliable broadcast is a primitive that guarantees all honest parties eventually deliver the same message, even if some parties lie.

The consensus number of a cryptocurrency gives the complementary fact. Updates with a single owner need no consensus, while objects with $k$ independent writers need agreement among those $k$ writers.

We transpose both results to the scientific record. A correction notice and a citing paper are independent if the citing paper's claims stay valid whether or not the correction is applied. Examples are a fixed affiliation, an updated grant number, or a citation used only for background. They are dependent if the correction touches something the citing paper builds on. This is the "feet of clay" situation identified by Tailoring Scientific Argument Mining for Scientific Literature Correction, where a retracted citation in a methodology role invalidates the citing paper's own result.

Separately, a paper whose record is touched by more than one authority behaves like a multi-writer object. Concurrent author corrigenda and editorial retractions are the typical case, and resolving them genuinely requires agreement.

The new contribution has three parts. The first is an empirical census of this structure, with the measurement unit chosen so the theory applies. The second is a learned independence relation whose false "independent" verdicts are the safety violation and are explicitly bounded. The third is a safety–automation frontier that plays the role of the safety–liveness trade-off. On that frontier, liveness means the fraction of correction–citer pairs released without human review. Safety means the rate at which truly dependent pairs are wrongly released.

Relation to Prior Work

Two lines of work motivate the gap, and both stop short of it.

Treat societally impactful scientific insights as open-source software artifacts recasts papers as versioned software with a main branch of consensus. It assumes a trusted maintainer who sequences all merges, and never asks which merges actually need sequencing. We ask exactly that question.

Öztürk and colleagues supply the closest method. Their mixed scientometric–NLP proposal classifies citation contexts with SciBERT over 7,424 citation pairs linked to 29 content-related retraction reasons. It also identifies the "criticism exception": a citing paper that critiques a retracted work is helping correction, not spreading the error. Their work frames the output as an editor-facing analysis of retractions. It does not cover ordinary corrections, does not have an asymmetric safety objective, and does not tie the classification to a propagation rule. We reuse their citation-function features, and we differ in what is predicted (pairwise independence), what is optimized (bounded missed dependencies), and what the output enables.

The distributed-systems papers constrain the design rather than supply it. Gafni and Whittaker et al. show that insisting on one total order invites stalls. Randomized protocols such as HoneyBadgerBFT and No Waitin' HotStuff escape FLP but pay cryptographic and messaging costs. Those costs are only worth paying for the dependent, multi-writer remainder. The census is designed to show how small that remainder is.

Finally, the record-integrity literature explains why the stakes are real. Retractions relevant to COVID-19 reports that over 30% of retraction notices gave no clear reason. That means the independence classifier must often work from thin notice text plus the citing context. Li Shen's correction note shows corrections failing to reach downstream copies.

Research Program

The first component is a correction census. We would assemble correction and retraction notices: Retraction Watch entries already linked in the Ozturk (2024) dataset, plus ordinary errata and author corrections of the kind indexed in The Informational Scope of Scientific Correction Notices. That index ranges from affiliation fixes to swapped figure labels in Bevers et al.. Annotators label each notice twice. The first label records what object it modifies: metadata, a figure or number, a method, or a headline claim. The second records which authorities issued notices on the same paper, and whether those notices were concurrent and conflicting.

From this we report the fraction of notices that are single-owner and the fraction that are multi-writer. In the theory's terms, these correspond to consensus number 1 and consensus number $k$.

The second component is a pairwise independence dataset and classifier. For a sample of (notice, citing paper) pairs, experts label the pair as independent, dependent, or corrective. "Corrective" covers the criticism exception, which is independent in the sense that matters: the citing paper remains valid.

The model sees three inputs: the notice text, the citing sentence with its argument role, and the citing paper's claim that depends on it. The model is built from SciBERT-style citation-context encoders following Öztürk et al. and an LLM judge. Its score is calibrated on held-out data. We pick a threshold $\tau$ so that the empirical miss rate stays below a target $\alpha$:

$$ \Pr(\text{released} \mid \text{dependent}) \le \alpha . $$

Here "released" means the score is below $\tau$, so the pair skips review. The quantity reported is the release rate at that $\alpha$.

The third component is a propagation simulation. Several independent repositories exchange notices with random message delays. Some repositories are adversarial: they withhold notices or send different notices to different peers. Independent pairs propagate by reliable broadcast, following the Frey et al. algorithm. Dependent pairs and multi-writer conflicts go to a review queue, which is resolved either by a single trusted editor or by an asynchronous agreement step. We measure the window of risk, defined as the time from notice issuance until every affected citer is flagged.

This component is deliberately secondary. It tests whether the classification is operationally meaningful; it is not offered as a deployment claim.

Baselines

The naive reference point is flag-all: every citing paper of a corrected work is sent to review. It has zero missed dependencies by construction and zero automation. Any method is judged by how much review load it removes while matching flag-all on safety up to $\alpha$. The opposite naive point, flag-none, approximates current practice. We report its miss rate to show what is at stake.

The strongest existing method on this exact question is the citation-context classifier of Öztürk et al. (SPIRE 2024 proposal). It is a SciBERT model that classifies argument units and citation contexts, and here its outputs are mapped to dependent versus independent. We will reimplement it on the same training split, with the same calibration procedure and the same hyperparameter search budget as our model. No reported result was found on Lacuna for this classifier on any dependency or independence task, so its numbers will be produced here.

The second strong competitor is a frontier LLM judge given the notice and the citing context. The closest reported number is on revision verification in AutoSupervision, where GPT-5.5 peaks at about 0.501 on verification. On SciFact, VERISCI reaches 75.7% label accuracy when rationales are correctly identified. Neither number comes from our task. Both suggest that unconstrained judges will be noisy on directional, evidence-dependent decisions.

Beating the baselines means a strictly higher release rate at $\alpha = 0.02$ and $\alpha = 0.05$. Confidence intervals for this comparison come from bootstrap resampling over citing papers.

Two ablations isolate the key ingredients. The first removes the correction-side input (notice type and modified object) and keeps only citing context, which tests whether reasoning about what changed matters beyond how it was cited. The second removes calibration and uses a plain 0.5 threshold, which tests whether the safety bound, not raw accuracy, drives the operational gain.

In the simulation, the baseline is a single central registry. Central-registry runs both with and without an independence filter are scored with the same window-of-risk metric as the decentralized runs.

Why not just…?

Why not just keep a central registry like CrossRef and skip the distributed-systems framing? For the commutative majority, a central registry is a perfectly good transport, and we expect the census to confirm this. The proposal does not depend on decentralization. The independence relation is equally useful to a central registry deciding whom to notify. What the theory adds is a principled answer to which cases a central party must actually arbitrate. Those are the multi-writer conflicts, and the census counts them. If they turn out to be negligible, we will report that coordination machinery is unnecessary for science, and that is itself a finding.

Why not just use the Öztürk citation-function classifier as is? This is the real threat, because the core signal may live mostly in citation context. The two ablations and the head-to-head frontier settle it. If removing correction-side inputs and calibration leaves the release rate unchanged at fixed $\alpha$, then the contribution reduces to the census and the safety framing. In that case we would say so and drop the claim of a better classifier.

Why not just flag every citing paper? It is safe but does not scale. Over 10,000 retractions in 2023 alone, each with its own list of citing papers, produce a review load that is already not being handled. We report the absolute number of reviews each method saves at matched safety, so that a reader can judge the trade-off directly.

Why not just ask a strong LLM per pair? It is included as a baseline. The AutoSupervision verification ceiling near 0.5 and the evidence-manipulation attacks in Fact-Saboteurs suggest that its errors will be uncalibrated and exploitable. We test this directly by feeding adversarially rewritten notices to every method and measuring how far each one's miss rate moves.

Decisive Checks

The census succeeds if it gives a reproducible split of notices into single-owner and multi-writer classes, with inter-annotator agreement reported. The headline result is the release rate at $\alpha = 0.05$ on held-out citing papers from unseen fields, including nanobiology papers in the style of the Öztürk SAM annotations. That result must beat the Öztürk-style classifier and the LLM judge.

The idea fails if three things happen together: dependent pairs cannot be separated from independent ones at a useful release rate, the correction-side ablation shows no effect, and the simulation shows no shortening of the window of risk beyond what a central registry with flag-all achieves. In that case the honest conclusion would be that the consensus lens describes the problem well but does not change how corrections should be handled.