Systems that judge how trustworthy a source is usually reward the source for agreeing with whatever answer currently looks most likely to be true. That reward is the same whether the source agreed on "Paris is the capital of France" or on an obscure flight's exact departure gate. A source can therefore build a high trust score cheaply by repeating easy, popular facts, then spend that trust on false rare facts, which are the ones that matter. We propose to fix this by borrowing a tool from database query optimizers. A cardinality estimator is the component that predicts how many rows a filter like attribute = value will return, which is the same as predicting how common that value is. We use this estimate of how likely a value is to be asserted by chance, and we credit a source only for agreement that beats chance.
The analogy is closer than it sounds. Database research has spent two decades learning that frequency estimates go wrong in two systematic ways. They assume values are spread evenly inside histogram buckets, and they assume different attributes are independent. Both failures are worst for rare values and for correlated attribute combinations. Those are exactly the cases where a credibility score would be laundered. So the proposal has two parts. It gives source-trust estimation a measure of how much information each agreement carries, and it brings along the database field's hard-won knowledge of when frequency estimates mislead.
Core Claim
Truth discovery means inferring the true value of a disputed fact while jointly estimating each source's accuracy. In standard truth discovery and Knowledge-Based Trust, every confirmed claim adds equally to a source's credibility. We claim that weighting each confirmation by how unlikely it was to occur by chance, using a correlation-aware estimate of how often that value is asserted, has two effects. It closes a concrete attack in which a source "launders" its credibility through easy facts. It also raises accuracy on rare, high-information facts, without losing to majority voting on clean data. The second part of the claim is that 1D-histogram estimators, which assume independence, recover only part of this gain. The remaining gain requires joint estimation over each entity's attribute bundle, mirroring the database finding that independence assumptions drastically underestimate correlated cardinalities (Li et al., Rabbani et al.).
Where This Sits
The motivating evidence comes from Knowledge-Based Trust. That work jointly estimates true values, source accuracy and extractor reliability by expectation-maximization over 2.8 billion triples. The research file records that it treats facts uniformly and does not distinguish being right about common facts from being right about rare ones. We keep the expectation-maximization structure and change only how each claim is weighted. The constraining evidence comes from Truth Discovery Algorithms: An Experimental Evaluation, which compared 12 algorithms and found that unweighted majority voting is a "formidable baseline." It also found that dependency (copier-detection) models scale quadratically and become numerically unstable beyond about 5,000 sources. This rules out a design that detects laundering by modeling pairwise source copying. Our weights are computed per claim from summary statistics, so cost grows linearly with the number of claims.
The closest bridging work is Query Perturbation Analysis. It uses database machinery to test how fragile a data-driven claim is when query parameters shift, but it assumes the database is uncorrupted ground truth and never scores sources. We do the reverse: database frequency estimation serves source scoring. RA-RAG estimates source reliability for retrieval-augmented generation by cross-checking sources and then taking a reliability-weighted majority vote. It shows that plain majority voting collapses when 90% of sources are adversarial. However, its reliability update also counts each agreement equally, so it is exposed to the same laundering. The method we reuse comes from cardinality estimation: histograms and most-common-value lists, and index-based join sampling, which estimates the frequency of correlated combinations by sampling rather than multiplying marginal frequencies.
Method
Let a claim $c = (e, a, v)$ say that entity $e$ has value $v$ for attribute $a$. Let $\hat p(c \mid S_e)$ be the estimated probability that a source with no real knowledge of $e$ would assert $v$, given the other values $S_e$ it asserted about $e$. This is a conditional selectivity, read off the claim store. Let $\hat y_c \in [0,1]$ be the current probability that $c$ is true. The trust update for source $s$ with claim set $C_s$ is then
$$ t_s = \frac{\sum_{c \in C_s} w_c \, \hat y_c}{\sum_{c \in C_s} w_c}, \qquad w_c = -\log \hat p(c \mid S_e). $$
The weight $w_c$ is the information, in nats, carried by the assertion. It is near zero for values nearly every source asserts and large for rare ones, because self-information grows as a value's probability shrinks. Conditioning on $S_e$ is what defeats bundled laundering. Suppose a source copies a popular record wholesale. Its agreement on the record's attributes is then mostly one piece of evidence, not many, and the conditional estimate discounts the redundant attributes.
Truth inference stays the same as in the base algorithm, using trust-weighted voting or KBT-style expectation-maximization. The two steps alternate until they converge. We will compute $\hat p$ with three estimators of increasing strength. The first is per-attribute 1D histograms with most-common-value lists, which assume uniformity and independence, as PostgreSQL does. The second is a correlation-aware estimator built by sampling entity records under a fixed lookup budget, as in Leis et al.. The third is a small learned joint model per attribute group.
The analysis component asks how estimator error propagates into trust error. We will measure the q-error of $\hat p$ (the factor by which an estimate is off) on held-out claims. The literature uses this error measure, and Perron et al. use a q-error of 32 as their re-planning threshold. We then relate it to error in trust rankings. Underestimating how often a rare value appears inflates $w_c$ and overrewards sources that are merely lucky. The estimator literature predicts this underestimation is systematic for correlated values.
Learned estimators can be poisoned (PACE, algorithmic complexity attacks). An adversary could therefore flood the store with rare-looking claims to manipulate $\hat p$. To guard against this, we cap each source's contribution to the statistics used to compute $\hat p$, and we test against this second-order attack explicitly.
Baselines
The naive reference point is unweighted majority voting over the same claims. It needs no tuning, and the evaluation in Truth Discovery Algorithms found it matches or beats complex models on the Book, Flight and Weather benchmarks. The file gives no per-dataset accuracy figure for it (no reported result found on Lacuna for majority-voting accuracy on Flight or Book), so we will rerun it ourselves. Beating it on clean data is not required. Not losing more than one point of accuracy on clean data while winning on tail facts and under attack is the bar.
The strongest existing competitors are the best-performing iterative and graphical truth discovery methods from that evaluation, together with KBT-style expectation-maximization. We include TruthFinder, 2-Estimates/3-Estimates and the latent truth model, initialized at the 0.8 source-trust value that evaluation identified as the best starting point. For the retrieval setting, the strongest competitor is RA-RAG's weighted majority vote with $\kappa$-reliable source selection (Hwang et al.). That paper reports a 99.1% token reduction with 1,000 sources. The file gives no exact accuracy value for it, only that it "maintained high accuracy" under adversarial injection, so we report our own rerun. All methods run on identical claim sets and splits. Each gets the same hyperparameter search budget on a development split and the same iteration cap, and runs on the same single machine.
Two ablations isolate the key ingredient. The first, inverse-popularity weighting, sets $w_c$ to minus the log of the fraction of sources asserting $c$. It uses raw counts, with no estimator and no conditioning. The second is our method with the independence-assuming 1D estimator. The full method beats inverse-popularity weighting only if it improves tail accuracy and suppresses laundered trust beyond what raw counts achieve. The gap between the 1D and joint estimators then tests whether correlation-awareness matters.
Why not just…?
Why not just use majority voting, since it is already so strong? It is strong exactly when errors are independent and minority, and we will report that it wins or ties on clean data. It cannot handle a source that is correct on popular facts, because it never scores sources at all. Under widespread misinformation it falls into the "majority vote trap" documented by Ge et al. and collapses under the adversarial-majority setting in Hwang et al.. Our attack experiments measure this directly on tail facts.
Why not just down-weight popular claims by counting how many sources assert them? This is the real threat, because it captures most of the intuition at almost no cost. It has two weaknesses. Raw counts are poor estimates for the rare values where weighting matters most: a value seen once may be common in reality and absent from the sample. Counts are also blind to bundles: a copied record yields many individually uncommon attribute values that are jointly one piece of evidence. If the bundled-laundering experiment shows no gap between this ablation and the joint estimator, the honest conclusion is that the database machinery is unnecessary. We would then report the simpler method as the result.
Why not just estimate a separate accuracy per source per attribute? Per-attribute trust removes laundering across attributes, for example being right on capitals and lying on vote counts. It does not remove laundering within one attribute, where common and rare values of the same field coexist. It also multiplies the number of parameters and makes sparse sources noisier. We include it as a baseline and stratify results by within-attribute versus cross-attribute attacks.
Why not just whitelist reputable sources? Dong et al. found that 85% of high-trust, low-PageRank sites they checked were truly trustworthy, so reputation is a poor stand-in for accuracy. Ge et al. found that source filtering underperformed generation-stage reasoning on CONFACT. Static lists also cannot react to a newly laundered source.
Proposed Experiments
The primary testbed is the real truth discovery benchmarks used by Waguih & Berti-Équille: Book, Flight, Weather, Population and Biography, about 10.8 million claims in total. On clean data we report overall accuracy and accuracy stratified by deciles of the true selectivity of the gold value. The second stratification is where we expect gains.
We then inject laundering adversaries. These synthetic sources copy the majority value on high-frequency attribute values and assert false values on a chosen set of rare facts. We vary the adversary share from 5% to 90%, matching the range of the adversarial-majority setting in RA-RAG, and run two variants. In the per-claim variant, adversaries agree attribute by attribute. In the bundled variant, adversaries copy whole popular records. Metrics are accuracy on the targeted rare facts, the trust rank of adversaries compared with honest sources, and rank correlation between estimated and gold source accuracy.
A transfer experiment uses RA-RAG's question-answering setting on NQ, TriviaQA and HotpotQA. Here simulated sources answer easy questions correctly and lie on hard ones. Answer frequency across sources stands in for selectivity, and we compare against weighted majority voting. A third experiment runs the poisoning attack on $\hat p$ itself, with and without the per-source contribution cap.
The decisive checks are three. First, under bundled laundering, the joint-estimator method keeps adversary trust below honest-source trust where every baseline, including inverse-popularity weighting, does not. Second, tail-decile accuracy improves on clean data while overall accuracy stays within a point of majority voting. Third, the q-error of $\hat p$ on tail values predicts the size of trust-ranking error, which would show that estimator quality, not just weighting, drives the effect. If the first check fails against inverse-popularity weighting, we drop the database component and report the simpler result.
References
- TKHist: Cardinality Estimation for Join Queries via Histograms with Dominant Attribute Correlation Finding
- Optimizing SPARQL Queries using Shape Statistics
- Knowledge-Based Trust: Estimating the Trustworthiness of Web Sources
- Truth Discovery Algorithms: An Experimental Evaluation
- Query Perturbation Analysis: An Adventure of Database Researchers in Fact-Checking
- retrieval augmented generation with estimation of source reliability
- Cardinality Estimation Done Right: Index-Based Join Sampling
- How I Learned to Stop Worrying and Love Re-optimization
- PACE: Poisoning Attacks on Learned Cardinality Estimation
- Algorithmic Complexity Attacks on All Learned Cardinality Estimators: A Data-centric Approach
- Resolving Conflicting Evidence in Automated Fact-Checking: A Study on Retrieval-Augmented LLMs