Privacy law is fragmented. The EU and UK require opt-in consent before tracking, California works mostly through opt-out ("Do Not Sell or Share"), and Brazil's law (LGPD) differs again. Websites respond by showing different consent interfaces depending on where the visitor appears to be. Automated audits of whether sites honor consent work well today only for the EU case, and they rely on hand-written rules that break on unfamiliar interfaces. We propose to crawl the same websites from several geographic vantage points, which gives matched pairs: one site, several legal regimes, several interfaces. These pairs let us train and test a small fine-tuned web agent (a model under about 10 billion parameters that operates a real browser from screenshots) whose job is to execute consent choices and report what the site actually did. The legal judgment about whether that behavior is lawful is kept in an explicit, swappable rule layer, separate from the model.
The central claim is that cross-jurisdiction failure in automated compliance auditing is mostly a perception and navigation problem, not a legal-reasoning one. Interfaces shift from opt-in banners to "Do Not Sell" footer links, and the meaning of a click shifts with them. If that is right, a model trained only on the jurisdiction-neutral task of doing and describing consent interactions should transfer to a held-out jurisdiction far better than models trained end to end to output "violation / no violation." That should hold even when the end-to-end models get jurisdiction-specific adapters. The paired crawl makes the claim testable in a way that static legal text corpora cannot, because case law and privacy policies never come in site-matched parallel versions.
Why this, and not the obvious combination
The obvious idea at this intersection is to take progressive modular adapters from ProMALex and bolt them onto a screenshot-based agent like those in BrowserForge. ProMALex shares low-rank adapter layers (small trainable weight updates) across jurisdictions in the lower network layers and branches them per jurisdiction higher up; BrowserForge fine-tunes 4B and 9B screenshot agents on 203k generic web trajectories. The combined auditor would branch per jurisdiction and learn to flag violations. We think this is the wrong default.
ProMALex was evaluated only by perplexity on court text. Its gains come from jurisdictions sharing legal vocabulary, which is not the bottleneck when a model has to find a hidden revocation button. Asking one set of weights to learn both "where is the reject control" and "is persisting this cookie lawful here" also entangles two things that should be separable. Beyond Borders constrains the design here: forcing cross-jurisdiction invariance in summarization stripped a specialized legal model of useful knowledge, which they call representation erasure. The lesson we take is to keep jurisdiction-specific knowledge in a place where it cannot be erased. For privacy auditing, much of that knowledge can be written down as explicit rules.
The measurement literature motivates the gap. User Tracking in the Post-cookie Era automated Accept/Reject/No-Action choices with Puppeteer and the rule-based Consent-O-matic extension across 27,953 sites. They found over 75% of first-party ID leaks occurred before consent or after "Reject All." Johnny Can't Revoke Consent Either scripted multi-step revocation journeys and found 74.2% of sites failed to forward revocation to third parties, but needed heavy manual work where scripts failed. Both are essentially GDPR-only, and both depend on rules that encode known consent-platform layouts.
The modeling literature supplies the method but stays static. Chen et al. showed that fine-tuning small models with low-rank adapters (LoRA) beats prompting on privacy-policy text across US, EU and Chinese taxonomies, but never checked whether sites behave as their policies say. AutoBot distilled dark-pattern detection into small models, but works from single screenshots under one generic taxonomy. GraphCompliance showed that explicit policy graphs with exception handling beat raw LLMs on GDPR enforcement cases, but builds its context from curated text, not live sessions. Our design borrows from GraphCompliance the idea that law should be an explicit structure, and from BrowserForge the screenshot agent recipe. It then replaces the agent's generic task distribution with consent journeys whose outcomes the browser itself can verify.
Core Claim
Consider an auditor factored into two parts. The first is a small screenshot agent trained on instrumentation-verified consent trajectories. Its output is a consent trace: which choices were offered, how many actions each took, what was executed, and what cookies, storage values and network requests followed. The second is a per-jurisdiction rule table evaluated over that trace. On a jurisdiction held out from training, this factored auditor will lose much less accuracy in detecting violations than (a) an end-to-end violation classifier and (b) the same agent with ProMALex-style jurisdiction branches.
We will report the transfer gap for each system. For a held-out jurisdiction $j$, this is $\Delta_j = F_{\text{in}} - F_j$, where $F_{\text{in}}$ is violation-detection F1 averaged over training jurisdictions and $F_j$ is F1 on the held-out one. The claim predicts the smallest $\Delta_j$ for the factored system. It is falsified if the branched end-to-end agent matches it.
Research Program
Geo-paired crawl. We draw sites from the Tranco list used by Papadogiannakis et al. and seed it with the consent-platform sites in the Johnny DS1–DS3 lists and the 680-site UK CMP corpus from Dark Patterns after the GDPR. We then visit each site from vantage points in an EU member state, the UK, California, and Brazil, using fresh browser profiles. Each visit runs the instrumented stack those papers used, recording HTTP traffic, cookie set and delete calls, localStorage, IAB consent strings via __tcfapi, and OneTrust variables. We will report the fraction of sites whose consent interface differs across vantage points, since that fraction alone determines how much interface shift exists to study.
Verified trajectories. A large teacher agent proposes and executes consent journeys in a Proposer–Solver loop like BrowserForge's: accept, reject, opt out via footer link, revoke after accepting, and send a Global Privacy Control signal. The browser state labels each trajectory automatically. If the agent claims it rejected optional cookies, the consent string and the subsequent cookie and request log either reflect rejection or they do not. This replaces BrowserForge's model-based judge with a ground-truth check. It also directly penalizes the click loops and aimless navigation that cause 29% and 42% of BrowserForge's failures. Surviving trajectories are rewritten into step-by-step rationales and used to fine-tune 4B and 9B screenshot agents.
Rule tables. With legal annotators, we encode for each regime a small set of predicates over traces. Examples: "optional tracking cookie set before any choice," "revocation requires more actions than acceptance," "opt-out link absent," and "third-party requests continue after opt-out." Following GraphCompliance, each predicate carries exceptions, such as strictly necessary cookies.
Gold evaluation set. Annotators label a stratified sample of site-by-vantage-point visits per jurisdiction for both trace correctness and violation status under the local rules. Johnny's 20.5% "more actions to revoke" and 57.5% "cookies persist after revocation" rates give us an EU sanity check on whether our labels are calibrated.
Baselines
The naive reference point is Consent-O-matic with jurisdiction rule tables. This is the rule-based pipeline of Papadogiannakis et al., with our same rule tables applied to the traces it produces. It isolates how much a learned agent adds over hand-written interface rules. We expect it to be strong in the EU and to lose coverage at non-EU vantage points, since its rules target known consent platforms. It is beaten if the factored agent produces correct traces on clearly more visits outside the EU at equal or better EU accuracy.
The strongest existing agent is the BrowserForge-style generic fine-tuned agent: the 4B/9B recipe trained on generic web episodes, prompted with consent tasks, and run through the same rule tables. Its reported strength is 38.00% task success on Online-Mind2Web for the 9B model (BrowserForge). We found no reported result on Lacuna for any agent on consent navigation or revocation, so all consent-task numbers will be new. To keep the comparison fair, it gets the same base model, the same number of fine-tuning steps on a size-matched mix of its generic data plus our trajectories, and the same step budget per site. Beating it means higher verified action success and trace accuracy on held-out jurisdictions.
The frontier zero-shot agent is a large proprietary vision-language model given screenshots and the rule text in its prompt, and asked for traces or verdicts. Its cost per site is reported alongside its accuracy. The nearest reported figure for a frontier model in this space is AutoBot's Gemini pipeline at 0.90 F1 on D3 single-screenshot dark-pattern detection (AutoBot). That task is easier than multi-step consent journeys, so it is a ceiling hint, not a direct comparison.
The end-to-end branched agent is the key ablation. It uses the same trajectories and the same base model, with ProMALex-style shared lower adapters and per-jurisdiction upper branches, under the same total rank budget. It is trained to output violation verdicts directly. A second variant, a pooled end-to-end model with no branching, matches ProMALex's "Shared" baseline. In ProMALex's multilingual text setting, the shared baseline reached perplexity 7.104 versus ProMALex's 5.468. The factored design wins if its $\Delta_j$ is smaller on every held-out jurisdiction, with confidence intervals bootstrapped over sites.
Why not just…?
Why not just prompt a frontier model with the local law? This is the most serious threat. Frontier models are good at reading banners, and the legal predicates here are simple. Two pieces of evidence point against it. Chen et al. found prompting confused marketing language with legal requirements on privacy text. BrowserForge's fine-tuned 9B agent competed with GPT-5 plus a grounding model on live navigation. Neither result is on consent journeys, though, so we run this baseline at full strength. If it matches the factored small agent, the contribution reduces to the paired benchmark and the factoring result, applied to a frontier model. That outcome is still useful, and we would report it as such. Cost and reproducibility across hundreds of thousands of site visits remain the argument for small models.
Why not just extend Consent-O-matic's rules to new regimes? Rules are cheap and auditable. But Johnny needed manual intervention exactly where interfaces deviated from known layouts, and 19.87% of sites changed interface modality to make revocation harder. The Consent-O-matic baseline measures this directly. Its coverage drop from EU to non-EU vantage points, compared against the learned agent's, settles the question.
Why not just give the end-to-end model jurisdiction adapters, ProMALex-style? That is the obvious idea, and we run it as the main ablation instead of assuming it fails. Its weakness should show in a rule-swap test. We alter one predicate in a rule table, for example counting a pre-checked box as no consent. The factored auditor follows the change with no retraining; the branched model cannot without new labels.
Why not just pool all jurisdictions into one fine-tune? Pooling helps in distribution but says nothing about a regime absent from training. Leave-one-jurisdiction-out evaluation answers that question directly.
Decisive Checks
Three results decide the claim. First, leave-one-jurisdiction-out $\Delta_j$ must be smallest for the factored agent across all four regimes. Second, among paired sites, the end-to-end models' errors should concentrate on sites whose interface differs across vantage points. If errors instead spread evenly over identical-interface sites, the failure is legal entanglement rather than perception, and the claim needs revising. Third, the rule-swap test must show that the factored auditor tracks rule changes exactly while the end-to-end systems do not.
Secondary checks will report click-loop rate and verified action success against BrowserForge's failure taxonomy, plus contamination controls in the spirit of Sensitivity of Small Language Models to Fine-tuning Data Contamination: no gold-set site may appear in training trajectories from any vantage point.
The main known risk is anti-bot blocking, which accounted for 9% of BrowserForge failures and may be worse for proxied traffic. We will report how often each vantage point is blocked and exclude blocked visits symmetrically across all systems.
References
- ProMALex: Progressive Modular Adapters for Multi-Jurisdictional Legal Language Modeling
- BrowserForge: Scaling Web Episode via Parallel Browser Sandboxes
- Beyond Borders: Investigating Cross-Jurisdiction Transfer in Legal Case Summarization
- User Tracking in the Post-cookie Era: How Websites Bypass GDPR Consent to Track Users
- Johnny Can't Revoke Consent Either: Measuring Compliance of Consent Revocation on the Web
- Using LLMs for Automated Privacy Policy Analysis: Prompt Engineering, Fine-Tuning and Explainability
- Automatically Detecting Online Deceptive Patterns
- GraphCompliance: Aligning Policy and Context Graphs for LLM-Based Regulatory Compliance
- Dark Patterns after the GDPR: Scraping Consent Pop-ups and Demonstrating their Influence
- Sensitivity of Small Language Models to Fine-tuning Data Contamination