NagaYu/sludge-ui-counterfactuals

Sludge counterfactual UI corpus

This model does not determine legality.

It reports provisions that may be implicated and the screen elements that are the factual basis for looking at them. Whether a provision is actually engaged depends on facts no UI tree contains — the purposes of processing, the legal basis relied on, the audience, the rest of the journey, prior consent, sector rules — and is an assessment for a qualified human.

It has no feature that labels a named company's product as unlawful, and it is not built to acquire one. It is a tool for reviewing interfaces, not for publishing verdicts about them.

A fully synthetic corpus of UI trees for deceptive-design-pattern detection, built from counterfactual pairs: a fair screen, and the same screen with exactly one thing changed.

fair                                    unfair
─────────────────────────────           ─────────────────────────────
[ Accept all ]  [ Reject all ]          [ Accept all ]  [Reject all]
  200×52px        200×52px                200×52px        95×29px
                                                          ▲
                                        one edit: bbox + font_size
                                        evidence: cook_reject_6
                                        remediation: the inverse

Because the twins are identical everywhere else, the edit defines the class label, the localisation target and the remediation simultaneously. There is no annotation pass, and no way for the three to drift apart.

• Model: NagaYu/sludge-small
• Demo: NagaYu/sludge

Splits — by template, never by sample

split	rows	layouts	contents
train	18,000	11 train-only templates	fair + single-pattern + 12 seen class combinations
validation	1,800	same templates, disjoint seeds	sanity check (it saturates — see below)
test	3,600	6 held-out templates	fair + single pattern
hard	1,800	the same held-out templates	18 class combinations that never appear in training

25,200 rows total. No UI layout is shared across the train/test boundary, and every one of the 27 classes appears in both a train template and a test template.

validation re-uses the training templates, so a model that has learned those layouts reaches ~1.00 micro-F1 on it within two epochs. It is published for completeness; the reference pipeline holds out two training templates as a dev split for model selection and threshold tuning instead, and never touches test or hard for either.

The fair screens are deliberately hard

Fair screens contain honest urgency (a dated sale), honest scarcity (a sourced stock count), honest social proof (an attributed review) — and a set of hard negatives: benign designs that trip naive thresholds.

hard negative	which rule it defeats
a smaller secondary link beside a primary button	false_hierarchy on area ratio
a countdown to a real, stated, persistent deadline	countdown_timer on "clock present"
"Only 2 left — 2 units remain in the Rotterdam warehouse"	low_stock_message on scarcity wording
a box ticked by default because ticking it is the privacy-protective choice	preselection on default_selected
faint text that is a build number, not a decision	visual_interference on contrast
a one-button "OK" on an informational dialog	forced_enrollment on "no decline present"

Without these, hand-written rules scored 0.84 micro-F1 on this corpus — far above what rule-and-screenshot systems reach on real interfaces. That gap was a property of the corpus, not a discovery about rules.

Fields

field	what it is
tree_json	the UI tree: nodes with role, text, bbox, colours, font, depth, default-selected state, link target, screen index
pattern_ids	taxonomy classes present (empty for fair rows)
evidence_json	per class, the node ids that constitute the evidence
provision_elements_json	the same finding resolved to provision elements, for every configured jurisdiction (eu, us_federal, us_ca)
diffs_json	exactly what the edit changed: node, field, before, after
remediation_json	the inverse of the edit, with target values
pair_id	links a fair row to its unfair twin
edit_scope	single_field, single_decision, or combination
template, family, fair, n_nodes, n_screens	provenance and shape

Two label sets ship on every row — the taxonomy class and the jurisdiction-resolved provision elements — so you can train or evaluate against either, and audit the mapping without running the code.

Realism: what was deliberately removed

The generator can annotate a node with facts no real interface exposes ("this stock figure came from the warehouse feed"). Training on those would inflate every number. The reference pipeline strips them and derives each feature from what an AX/DOM traversal genuinely reports:

once an annotation	derived instead from
deadline	an explicit date in the visible text
inventory_source	a sourcing phrase in the visible text
author	an attribution line beside the quote
label_action_conflict	comparing the accessible name to the link target
the generator's slot names	nothing — never read

The annotations remain in tree_json so you can measure what provenance metadata would be worth if a platform exposed it. A behavioural test enforces that the reference featuriser reads none of them.

Taxonomy

7 categories / 27 classes, all from published work: Mathur et al. (2019) CSCW · Gray et al. (2018) CHI · Brignull, deceptive.design · EDPB Guidelines 03/2022 · OECD (2022) DEP No. 336 · Gray et al. (2024) CHI.

Category (Mathur et al. 2019)	Classes
Sneaking	sneak_into_basket · hidden_costs · hidden_subscription · bait_and_switch · hidden_information
Urgency	countdown_timer · limited_time_message
Misdirection	confirmshaming · visual_interference · false_hierarchy · preselection · trick_question · pressured_selling · disguised_ad
Social Proof	activity_message · testimonials_uncertain_origin
Scarcity	low_stock_message · high_demand_message
Obstruction	hard_to_cancel · price_comparison_prevention · intermediate_currency · privacy_maze · dead_end
Forced Action	forced_enrollment · forced_registration · privacy_zuckering · nagging

Intended use, and what this is not

For: training and evaluating UI-tree pattern detectors; testing whether a detector reacts to a design change rather than to a template; measuring what screenshot-based perception costs.

Not for: asserting that any real product is unlawful. The labels identify design patterns from a research taxonomy and provisions that may be implicated. They are not legal findings, and the corpus contains no real application.

No real-world data. No screenshot or UI tree from any real application appears here. Public dark-pattern datasets are for evaluation only, under their own terms.

Licence: Apache-2.0.