Steven Hé (Sīchàng)

web dependency research proposals (authored by agents unless marked 🧑)

recommendation: evidence loss under dependency failure

• question: which providers are necessary to preserve an article’s useful evidence?
  • evidence: content that supports a specific answer
  • examples: cited references, tables, source images, author identity, sponsorship disclosure
• hypothesis: some failures silently remove evidence while leaving readable prose
  • proposed explanation: evidence and main prose can arrive through different components
  • not a measured result
• closest work
  • Web Dependency Analyzer targets rendering impact
    • full text and artifact must be checked before claiming a distinct contribution
  • storage-policy study compares behavior and compatibility
  • infrastructure study maps shared failure risk
• proposed contribution
  • connect a dependency intervention to preserved evidence and answer correctness
  • release paired pages, annotated missing evidence, and reproducible task checks
• pilot
  • choose 30 public pages with explicit evidence-bearing components
    • technical documentation, product comparisons, statistical reports
  • define the question and supporting passage before interventions
  • collect five baseline loads per page
  • locally block each observed third-party provider in separate browser runs
    • then test two-provider combinations only when single-provider findings justify them
    • never disrupt the provider for other users
  • interleave blocked and unblocked runs
    • record timestamps, browser version, network endpoint, cache and consent state
  • compare screenshot, DOM, accessibility tree, extracted article, and answer
    • DOM: browser’s structured representation of the page
    • accessibility tree: representation exposed to assistive tools
• primary outcomes
  • proportion of annotated evidence retained
  • question accuracy and unsupported-answer rate
  • recoverability after unblocking
• secondary outcomes
  • visual difference, request counts, performance
  • report alongside task outcomes rather than substitute for them
• failure attribution
  • count failures above repeated-baseline variation
  • confirm they reverse after unblocking
  • record anti-bot pages and inaccessible pages separately
• stop or change direction
  • if effects only reproduce existing rendering metrics, release replication rather than claim novelty
  • if most changes are random ad rotation, improve pairing before increasing sample size
  • useful negative result: evidence survives despite substantial visual changes

replication or component: preserve useful content during blocking

• question: can selective blocking retain task success while reducing measured tracking opportunities?
• existing work already shows benefits from selective method and API intervention
  • recent comparison covers Amjad et al., ByteDefender, PURL, and FP-Inspector
  • generic component-level blocking is a replication
• remaining hypothesis: policies can retain answer-supporting evidence while reducing observed tracking
• closest work
  • privacy versus compatibility
  • tracker classification
  • AdGraph and perceptual blocking
• pilot
  • start with 20 pages that reproducibly break under a public blocker
  • manually define reading, navigation, table inspection, and disclosure checks
  • compare unblocked page, domain blocking, public filter list, and component-aware policy
  • keep provider identity, script initiator, content role, and changes separately
• evaluate
  • task success versus observed identifying-data transmission
    • transmissions are an operational privacy measure
    • do not infer full absence of tracking from fewer requests
  • policy maintenance cost and failure on later revisits
  • held-out sites and templates
• novelty condition
  • improvement must exceed existing storage isolation and blocking approaches on the same tasks
  • calling a model to label components is an implementation choice, not a contribution by itself
• stop
  • if manual exceptions deliver the same benefits more reliably, study maintainable exception rules

exploration: shared infrastructure among apparent independent sources

• question: do users or answer engines mistake many provider-backed pages for independent evidence?
• hypothesis: source diversity estimates can overcount copied or commonly operated pages
  • hosting concentration alone is insufficient evidence
• closest work
  • Hanley et al. narrative tracking
  • search evidence-independence proposal
• pilot
  • sample 50 technical questions with checkable answers
  • preserve retrieved and cited pages with timestamps
  • annotate common text, cited origin, disclosed ownership, affiliate identifiers, and shared scripts
  • label shared infrastructure as an uncertain clue
  • require text-origin or ownership evidence to label dependence
• evaluation
  • compare domain count, provider count, manual independent-evidence count
  • test whether reranking by verified independence improves answers
  • measure false merging of independent sites on large common hosting providers
• contribution condition
  • quantify consequences for evidence selection
  • generic clustering or narrative tracing already has close prior work
• stop
  • if provider identity adds no predictive value beyond copying and citations, omit it from ranking

Last edited: 2026-10-06