Automated fact-checkers decide whether a claim is true by retrieving evidence from the web and, in effect, counting how many sources agree. But "source" has quietly become a broken unit. A domain name no longer identifies an owner: content delivery networks, shared cloud hosting, and CNAME cloaking — a trick where a subdomain of a reputable site silently points to a third party's server — let one operator appear as dozens of apparently independent outlets. The ad-blocking world has already measured this: Investigating Advertisers' Domain-changing Behaviors and Their Impacts on Ad-blocker Filter Lists found 1,748 replica domains affecting 10.2% of 50,000 crawled sites, with 57.4% of first-party replica domains using CNAME cloaking, surviving blocklists for 410.5 days on average when caught and 784.7 days when not. Fact-checking has not caught up.
The security side of fact-checking is meanwhile converging on a defense that this exact phenomenon breaks. Fact-Saboteurs shows that planting a single claim-aligned sentence in an evidence repository flips neural verifiers — accuracy on Supported claims fell from 89% to 25.9% under homoglyph camouflaging — and prescribes "source triangulation across multiple platforms" as a defense. Triangulation assumes distinct domains are distinct sources. Shared infrastructure obfuscation violates that assumption. This proposal builds the first explainable verification system that counts infrastructure owners rather than domain names, surfaces that ownership evidence inside the explanation a human reads, and evaluates it against an attack benchmark where corroborating evidence is planted on cloaked, co-owned domains.
Core Claim
The claim is that infrastructure provenance — the DNS CNAME chain, TLS certificate fingerprint, and hosting network (ASN) behind each evidence URL — is a first-class evidential signal that current explainable fact-checkers ignore, and that adding it changes both robustness and explanation quality in measurable ways. Concretely, when a verifier retrieves evidence documents $e_1, \dots, e_n$, it should not treat them as $n$ independent corroborations. Each document carries an owner signature $\phi(e)$, the tuple of its final CNAME resolution target, its TLS certificate, and its hosting ASN. Corroboration should be counted as
$$ \mathrm{Corrob}(c) \;=\; \bigl|\{\phi(e) : e \in E_{\text{sup}}(c)\}\bigr|, $$
the number of distinct owner signatures among evidence supporting claim $c$, with cloaked clusters (same $\phi$, different domain strings) collapsed to one source and flagged in the explanation. This is the same unmasking logic that the RAD-domain study developed for advertisers — DNS record chaining and shared-certificate grouping — repurposed as an evidential gate inside a fact-checking pipeline, rather than as a blocklist maintenance tool.
Three existing lines motivate and constrain this. Fact-Saboteurs motivates it: it proves evidence planting works and proposes triangulation, but models attacks only as text-level perturbations in an abstract document store, never asking whether the "multiple platforms" are actually independent — our proposal is precisely the study of that blind spot. The RAD study supplies the method: we reuse its ownership-resolution techniques, but it never applies them to misinformation or claim verification. The explainability literature supplies the integration target: LOREN localizes verdicts to phrase-level "culprits" via latent variables and soft logic, QA-NatVer executes proofs over a deterministic automaton, and SEVA emits structured JSON with evidence spans and an error taxonomy — yet all three reason only over text, and none inspects where evidence is hosted. The explainability survey found 0 of 10 surveyed systems produced explanations that were causal or provenance-aware; our proposal directly fills that explanatory dimension.
Research Program
The program has three linked parts.
First, a measurement study establishing that replica-domain behavior exists outside advertising. We re-run the longitudinal pipeline of the RAD paper — DNS record chaining, shared TLS certificate grouping, URL path similarity, and JavaScript AST comparison over historical snapshots — but seeded with known low-credibility news domains instead of ad domains. The output is a prevalence estimate for "replica news domains": sets of outlets that look editorially independent while sharing certificates, CNAME targets, or hosting clusters. This grounds the threat model in reality rather than assumption; if prevalence is low, that is itself a finding, and the synthetic benchmark below stands on the RAD paper's documented attack architectures regardless.
Second, an attack benchmark, which we call CloakFEVER, built on FEVER and the attack methodology of Fact-Saboteurs (code available as S-Abdelnabi/Fact-Saboteurs). Each FEVER claim is situated in a simulated web retrieval environment where evidence arrives from synthetic domains annotated with realistic infrastructure metadata. An attacker plants fluent claim-supporting sentences on $k$ domains that have distinct fully-qualified domain names and apparent publishers but share one $\phi$ — mirroring the four RAD evasion architectures (CNAME cloaking behind a first-party subdomain, revolving domains, changing subdomains, CDN co-tenancy) at the proportions the RAD paper reports. The independent variable is $k$ and the cloaking pattern; the dependent variable is verdict flips toward Supported, exploiting the "Search and Match" bias that Fact-Saboteurs documented. Crucially, CloakFEVER also contains honest co-tenancy: legitimate small publishers genuinely hosted on the same CDN — so the task requires distinguishing malicious coordination from benign sharing, which pure "same IP ⇒ distrust" rules cannot do. Certificate-level and CNAME-target-level sharing, plus cross-time revolving patterns, are what separate the two.
Third, the system: a provenance-gated verifier with infrastructure-aware explanations. The reasoning backbone is a standard pipeline in the LOREN / QA-NatVer family — phrase or chunk decomposition, evidence alignment, logical aggregation — but every evidence document is first resolved to $\phi(e)$, corroboration is computed by the cluster count above rather than by document count, and cloaked clusters are both downweighted in aggregation and surfaced in the explanation. The explanation layer extends SEVA-style structured output with a provenance field: for each supporting document, the CNAME chain, certificate identity, ASN, and co-tenant list, plus a new error-taxonomy category ("cloaked corroboration") diagnosing why a set of agreeing sources was discounted. The explanation is auditable by construction because $\phi$ is computed from public DNS and certificate-transparency records, not learned — unlike the latent rationale variables in LOREN, the provenance part of the explanation can be checked by a journalist with a dig command.
Baselines
The do-nothing reference point is an unmodified FEVER-style neural verifier (BERT/RoBERTa, as evaluated in Fact-Saboteurs) run on CloakFEVER with no defense. This calibrates attack severity; the closest reported figure is Fact-Saboteurs' homoglyph camouflaging result, where accuracy on Supported claims dropped from 89% to 25.9%, though no reported result found on Lacuna for verdict-flip rates under infrastructure-cloaked evidence planting, which is exactly the number our benchmark will supply.
The strongest existing competitor is the defense Fact-Saboteurs itself proposes but never implements: multi-platform source triangulation. We implement it faithfully as majority voting over top-k retrieved results from distinct fully-qualified domain names, using the same retrieval stack, the same reasoning backbone, and the same tuning budget as our system, so the only difference is whether aggregation counts domain strings or infrastructure owners. This is the decisive comparison: if triangulation already resists cloaked planting, the proposal's core claim fails. For reasoning quality we hold the backbone fixed against the strongest published numbers as sanity anchors: LOREN reports a FEVER score of 72.93 on the blind test set with RoBERTa-large, and QA-NatVer reports 70.3% accuracy for Flan-T5 in the 32-shot setting; our system must not regress below the backbone's clean-set performance when no attack is present.
The key ablation replaces $\phi(e)$ with the bare domain string, so the system clusters evidence by FQDN exactly as triangulation does while keeping every other component identical. This isolates infrastructure provenance as the ingredient responsible for any robustness gain, separating it from the effects of the explanation interface or the aggregation architecture. A second ablation gives a strong instruction-tuned language model the raw DNS and certificate records in-context and asks it to judge independence end-to-end, testing whether the explicit deterministic clustering earns its place or a general-purpose model can do the same reasoning in one step.
Why not just…?
Why not just restrict retrieval to a whitelist of trusted domains? The evidence says static lists fail at exactly this layer. The RAD study shows filter lists were bypassed for 410.5 days on average by cloaked domains and 784.7 days when never blocked, across 10.2% of crawled sites — and CNAME cloaking specifically inherits first-party trust, meaning a whitelist keyed on the reputable parent domain admits the cloaked subdomain. Worse, Fact-Saboteurs broke verifiers that were already restricted to the curated Wikipedia repository, and FEVER shows static repositories leave a large coverage gap (test evidence retrieval recall of 45.89%). A whitelist shrinks coverage without closing the hole; our measurement study will quantify how many low-credibility domains would slip past a whitelist via cloaking.
Why not just use a bigger semantic model to judge evidence quality? The record is unambiguous that semantic reasoning alone is the vulnerable part, not the missing part. Fact-Saboteurs showed BERT, RoBERTa, and KGAT all fall to planted fluent sentences via the Search-and-Match bias, and that homoglyph perturbations evade subword tokenizers entirely; SEVA showed reasoning agents develop systematic biases under uncertainty (over-predicting "not attributable" because negative outputs carry more process reward). No amount of text-side capability tells a model that five agreeing articles share one TLS certificate — that information simply is not in the text. That said, the in-context ablation above takes this seriously: if a frontier model given raw DNS records matches our deterministic clustering, we will report it, and the proposal's fallback claim becomes about auditability and cost rather than capability.
Why not just do majority voting across search results? This is the real threat, because it is the triangulation baseline and we run it head-to-head. The prediction from the RAD data is that it fails by construction: 222 revolving domains and 167 CDN-backed domains in their corpus appeared as independent endpoints under one operator, so a URL-count vote is dominated by whoever registers the most cloaked domains — and Fact-Saboteurs showed an adversary needs only 10% proxy training data to generate planting text that transfers. The human side is equally documented: QABriefer found 45% of crowdworkers relied solely on the provided briefs without checking sources. If, contrary to this, majority voting over domain strings survives CloakFEVER at low $k$, our core claim is wrong and we will say so — that outcome is precisely what the benchmark's attack-success curves are designed to reveal.
Proposed Experiments and Decisive Checks
Experiment one measures attack success on CloakFEVER: the fraction of Refuted and NotEnoughInfo claims flipped to Supported as the attacker scales from 1 to 20 cloaked corroborating domains, for the do-nothing verifier, the triangulation baseline, and the provenance-gated system. The decisive prediction is a cost asymmetry: triangulation flips linearly in $k$ (each new domain string adds a vote), while our system stays flat until the adversary acquires genuinely independent infrastructure — distinct certificates, ASNs, and CNAME targets — which is the expensive resource the RAD data shows operators avoid (they revolve domain names, not certificates). We also report the false-cost side: performance on honest CDN co-tenants, where over-discounting shared hosting would punish legitimate small publishers.
Experiment two verifies no clean-set regression: on standard FEVER, the provenance gate is a no-op pass-through for independent sources, and the system should track its backbone — with LOREN at 72.93 FEVER score and QA-NatVer at 70.3% few-shot accuracy as the published anchors to match. Explanation quality is evaluated in LOREN's style: agreement between the surfaced provenance rationale and the aggregation decision, with LOREN's over-96% faithfulness agreement and culprit-identification F1 of 74.3 as reference points for what a faithful explanation looks like in this family.
Experiment three is a human study on the overconfidence trap, following the QABrief protocol from Generating Fact Checking Briefs, where briefs raised crowdworker accuracy by 10% from a 58% baseline but 45% of workers never checked sources themselves. We compare verdict accuracy and source-checking behavior for checkers given the same brief with and without the provenance panel (CNAME chains and co-tenant lists for each cited source). The decisive question is whether the panel converts blind trust into verified trust: fewer flipped verdicts on planted-evidence items, at no accuracy loss on clean items.
The closing check ties the three parts together: the measurement study's observed cloaking patterns parameterize a final, realism-matched round of CloakFEVER attacks. If provenance-gated aggregation resists those attacks, triangulation does not, explanations remain faithful, and humans with the provenance panel stop trusting cloaked corroboration, then the counting-owners-not-domains principle is established — and the survey's empty provenance column in explainable fact-checking gets its first occupant.
References
- Investigating Advertisers' Domain-changing Behaviors and Their Impacts on Ad-blocker Filter Lists
- Fact-Saboteurs: A Taxonomy of Evidence Manipulation Attacks against Fact-Verification Systems
- LOREN: Logic-Regularized Reasoning for Interpretable Fact Verification
- QA-NatVer: Question Answering for Natural Logic-based Fact Verification
- SEVA: Self-Evolving Verification Agent with Process Reward for Fact Attribution
- Explainable Automated Fact-Checking: A Survey
- FEVER: a Large-scale Dataset for Fact Extraction and VERification
- Generating Fact Checking Briefs