Evidence Infrastructure for Algorithmic Hate-Crime Cases
⚖ Case: AAT-2026-0041 · ACLU NorCal
📋 Incident
Case Title
Date
Location
Hate-Crime Basis
Legal Hook
🌐 Platform Connections
Perpetrator Account
Related Channel / Group
📎 Supporting Materials
📁 Drop files or click to upload
⚖️ Submitting Organization
Organization
Attorney
Role
Contact
⚠ Authorized legal proceedings only. All requests require court authorization before platform data is disclosed.
📋 Prima Facie Amplification Nexus — 5-Factor Threshold Assessment
Q1 Perpetrator was an active user of the identified platform(s)?
Q2 Evidence of repeated exposure to group-targeted hostile content before the incident?
Q3 Evidence that exposure was algorithmically recommended, not solely searched or subscribed?
Q4 Related content was reported or flagged before the incident and remained available?
Q5 Content specifically targeted a protected group connected to the crime?
ℹ This tool requests event-specific amplification logs under court supervision — not removal of lawful speech.
5/5
Confirmed
Strong Prima Facie Nexus
All 5 threshold criteria confirmed. Supports preservation request & ATS discovery motion.
📊 Confirmed Factors
✓Active platform user — device forensics
✓Repeated hostile exposure — 90-day window
✓Algorithmic recommendation — autoplay chain
✓Pre-incident flags + safety override lag
✓Protected group targeted — Sikh / South Asian
🎯 Request Scope
Platform
From
To
Target Account / Content IDs
🔒 Privacy Minimization
📄 Generated Documents
📋
Preservation Order Letter
Halt log deletion pending court review
⚖️
Proposed Protective Order
Sealed disclosure, auditor-only access
🗃️
ATS Request Appendix
Technical spec, format, redaction standards
👤
Special Master Motion
Appoint independent technical auditor
🗂️ ATS Data Fields — Select Required Fields
FieldRelevance to CausationPrivacy Risk
Exposure PathDistinguish hosting vs. algorithmic amplificationLow
Impression LogRepetition intensity, temporal proximity to incidentLow
Ranking PositionPrioritization vs. incidental displayLow
Recommendation Reason CodeNature of algorithmic conduct (collaborative filter, etc.)Medium
Organic Baseline EstimateCompute Amplification Above Baseline (AAB)Medium
Safety HistoryPlatform knowledge, safety override lag analysisMedium
Velocity MetricsRapid amplification detection, circuit breaker failureLow
Cluster ConcentrationTargeted community or geographic harm patternMedium
Raw Model WeightsDeep analysis — not required for causationHigh — avoid
Individual User DataNot required; privacy risk exceeds valueHigh — avoid
Amplification Above Baseline
5.0×
Algorithmic vs. organic reach
Safety Override Lag
12 days
After first community report
Repetition Intensity
High
Same cluster, perpetrator account
Velocity Threshold Crossings
3×
Before any downranking action
Algorithmic Share of Impressions
87.4%
of 146,400 total impressions
📊 Exposure Decomposition — 90-Day Window
Organic / Search
18,400
Personalized Rec.
92,700
Autoplay Chain
31,200
Notification / Trending
4,100
Algorithmic pathways account for 87.4% of impressions vs. organic baseline of 12.6%
📈 Amplification Velocity — Weekly
🔎 Auditor Findings
Collaborative filtering loop — content cluster recommended to users who watched similar hostile content; self-reinforcing amplification, not passive hosting.
Safety Override Lag — velocity threshold crossed 3× before any downranking (T−78d, T−52d, T−18d). 12-day gap after first report.
Autoplay chain — 6 of 8 content objects accessed via autoplay without direct search. System extended session time within hostile cluster.
Prior violations — 2 channel-level strikes in 180 days. Classifier flagged 3 videos as "borderline" but did not suppress recommendations.
📦 Hosting (§230 protected)
  • Video storage
  • Subscription feed
  • Direct search results
  • Channel page display
🔺 Amplification (above baseline)
  • Personalized recommendation
  • Autoplay continuation
  • Engagement-optimized ranking
  • Notification delivery