System Architecture & Evolution
SentinelCore
Chain of Thought Documentation
A living architecture ledger documenting the evolution of SentinelCore — from a policy playbook prototype to a federated, AI-augmented Trust & Safety operating system. Designed for both high-level scanning and deep technical review.
Ahmed Fathy · Internal Document · June 2026
Contents
- Executive Design Philosophy
- The Agent Architecture
- Orchestration Harness
- Policy Evaluator
- Contextual Enrichment
- Guardian Gatekeeper
- Research Foundation
- Regulatory Challenges
- Human Challenges
- Technical Challenges
- Architecture Index
- The Evolution Ledger
- V0 — Policy Foundations & Unified Queue
- V1 — Urgency-Based Queue Segmentation
- V2 — Media Queue & Analyst Wellness
- V3 — SentinelCore Federation
- V4 — Global Linguistic Intelligence
- V5 — Intelligence-Led Investigation
- V6 — Emergency Auto-File & Retraction Queue (Planned)
- V7 — Role-Based Access Control (Planned)
- Known Policy Gaps
- Design Decisions & Discarded Paths
- Future Roadmap
1. Executive Design Philosophy
Trust & Safety tools traditionally optimize for one of three forces: policy fidelity (applying rules consistently), operational velocity (resolving cases before harm escalates), or statutory compliance (meeting legal reporting obligations across jurisdictions). SentinelCore was designed to synthesize all three through AI-augmented decision architecture.
Machine learning classifies content before human eyes see it. Database hashing confirms known violations without analyst involvement. The enforcement ladder is policy-derived — not hardcoded — with AI confidence scores gating which actions are available. When confidence is high and hash matches confirm, the system auto-acts. When it is not, it surfaces exactly what the analyst needs to decide.
The philosophy is simple: let automation do what it can prove, let humans do what requires judgment, and let the system enforce the boundary between the two.
2. The Agent Architecture
SentinelCore was engineered using an orchestrated multi-agent system — not a traditional development team. Seven specialized AI agents operate under a coordinator that delegates tasks, verifies outputs, and maintains state across sessions. This section documents how the development harness, core reasoning agents, and safety gates were architected.
2.1 The Orchestration Harness — How SentinelCore Was Built
The development stack uses a coordinator/delegator pattern: a single Orchestrator agent maintains the architectural vision and delegates implementation to seven specialized sub-agents. No code changes, reviews, or analysis are performed by the orchestrator directly — each task is routed to the agent designed for it.
| Agent | Role | Pattern |
|---|---|---|
| codebase | Implementation — writes all code, runs builds, fixes errors | Plan → Build → Verify |
| planner | Read-only analysis — audits, research, architecture proposals | Research → Audit → Recommend |
| review | Code review — security, performance, best practices | Inspect → Critique → Suggest |
| brutal-critic | UX/content review — scoring against proven frameworks | Score → Critique → Recommend |
| explorer | Codebase search — pattern matching, file discovery | Search → Map → Report |
| legal-advisor | Regulatory research — jurisdiction-aware analysis | Research → Map → Advise |
| docs | Documentation — structured knowledge generation | Structure → Write → Validate |
State & Memory: Context persists across sessions via a tiered system. TheAGENTS.md build history acts as long-term memory — every phase is timestamped with agent identity, files changed, line counts, and build results. Thestate/session-state.json holds active working state. Thehandoff/latest.md provides continuity when sessions span multiple days.
Workflow Patterns: Four standard coordination patterns standardized across all phases — the Implementation Cycle (plan → delegate to codebase → validate build), Documentation Refresh (audit current state → delegate to docs), Full Feature Delivery (sequential phase delegation across planner → codebase → review → critic), and Legal Review Cycle (research → advisory → documentation).
Observability: Every agent task returns structured output with files modified, line counts, and build verification. Build compilation (npm run build) gates every single change — zero errors required before proceeding. The dashboard's AuditTrail component mirrors this same philosophy for the investigation workflow: immutable time-stamped logs of every automated and analyst action.
2.2 Policy Evaluator — Clause Mapping & Decision Routing
The Policy Evaluator is the system's reasoning engine for content classification. It takes raw incident data as input and outputs: the matched policy, the applicable enforcement ladder, and the decision tree path that produced the recommendation.
Input: Incident metadata (content preview, detection source, AI confidence scores, jurisdiction, policyId)
Processing: Queries the policy-playbook-data.ts registry (26 dossiers, each with decision trees, context rules, enforcement guidance, and edge cases). The enforcement-ladder.ts derives available actions from the playbook's decision tree branches — not from a static list.
Output: Connected display showing the playbook decision tree alongside the derived enforcement options, with branch-to-action mapping visible to the analyst. The step-by-step guidance (Step 1 → a/b → Step 2) is rendered inline in the InvestigationPanel.
Key design decision:The enforcement ladder was rebuilt (V4.1) to derive directly from playbook decision trees rather than using a separate static configuration. For 10 key policies, enforcement options are hand-mapped from tree branches (e.g., CSAM: hash-match → block+NCMEC, AI-suspected → SAG escalate, non-sexual → remove+CPS). For the remaining 16, options are auto-derived from the playbook'senforcementGuidance arrays.
2.3 Contextual Enrichment — OSINT, Attribution & Risk Scoring
The Contextual Enrichment layer aggregates external signals to inform the analyst's decision without overwhelming them. Data is surfaced contextually — always available but never competing with the reported content.
UniversalThreatContext
Identity graph (IP, VPN detection, connected alt-accounts, email enrichment), behavioral velocity (account age, posts/24h, anomaly detection), and trust score gauge. All data simulated deterministically from incident ID hash — consistent per case, no re-randomization.
OffenderGraph
Cross-account linking via device fingerprint, IP subnet, email naming pattern, payment instrument, and stylometric similarity. 2-4 linked accounts surfaced with match confidence scores. Currently simulated; roadmap plans live device attestation and behavioral ML integration.
SignalsPanel
AI classification (confidence + tier label), database match status (PhotoDNA, NCMEC, INTERPOL), risk assessment (score gauge + auto-action eligibility), and queue information (SLA, severity badge). Compact left sidebar — always accessible, collapsible.
AuditTrail
Immutable timeline of automated and analyst actions — Case Created → AI Classification → Database Check → OSINT Enrichment → Offender Graph Queried → Case Assigned → Resolution. Collapsible, always accessible. Mirrors the development observability philosophy.
2.4 Guardian Gatekeeper — Enforcement Safety & Compliance
The Guardian Gatekeeper is the final safety layer. It prevents unsafe enforcement actions before they execute — independent of the analyst's discretion. Every outbound action passes through at least one gate; critical cases pass through multiple.
Second-Reviewer Gate
Critical-tier cases (severity=critical, human_trafficking, CBRN) and special cases (CSAM, MED-999, Arabic translation) require a second reviewer's approval before the Resolve button is enabled. The analyst requests approval → 2-second simulated review → the gate unlocks. In production, this would route to a pooled reviewer queue.
Forced Mandatory Reporting
LEA and NCMEC reporting checkboxes are force-checked and disabled for legally mandated categories. Human trafficking → LEA cannot be unchecked. CSAM → NCMEC cannot be unchecked. DMCA → LEA checkbox is hidden entirely (civil process, not criminal). The checkbox label dynamically renders the actual agency name — "Generate NCMEC CyberTipline Report" not "Generate LEA Report."
Safe Option Gating
The "Safe — No Policy Violation" option is hidden when a hash match confirms the violation (PhotoDNA/DNAHash). For all other cases — even zero-tolerance policies — Safe remains available because reporters and AI classifiers can be wrong. Only mathematically verified violations lock the Safe option.
Emergency Auto-File Protocol (Planned — V6)
When a mandatory-reporting case approaches its statutory deadline without analyst resolution, the system auto-files an emergency submission. A dedicated Retraction & Correction Queue surfaces these for human review within 24 hours. Gates on hash-match confirmation only — AI-only detections never auto-file.
3. Research Foundation
Before writing a single line of code, we conducted research across three dimensions: regulatory requirements (what must we do legally?), human factors (what breaks analysts?), and technical gaps (what do attackers exploit that current tools miss?). Every feature in SentinelCore traces back to a finding in this section.
3.1 Regulatory Challenges
Trust & Safety operates under a patchwork of global regulations. Each jurisdiction mandates different agencies, different reporting windows, and different evidence preservation requirements. Getting this wrong is not a policy violation — it is a criminal liability.
| Regulation | Jurisdiction | Mandatory Agency | Deadline | Trigger |
|---|---|---|---|---|
| 18 U.S.C. § 2258A | US | NCMEC CyberTipline | Immediate | CSAM detection |
| FOSTA-SESTA | US | FBI + National Human Trafficking Hotline | 15 min | Human trafficking indicators |
| EU Directive 2011/36/EU | EU | Europol + National Police | Immediate | Trafficking / exploitation |
| UK Modern Slavery Act 2015 | UK | NCA + Modern Slavery Helpline | Immediate | Slavery / trafficking |
| DMCA § 512 | US | Lumen Database (transparency) | 24 hrs | Valid takedown notice |
| GDPR Art. 33 | EU | National DPA | 72 hrs | Personal data breach |
| UK DPA 2018 | UK | ICO | 72 hrs | Personal data breach |
| EU DSA Art. 34-35 | EU | European Commission / DSA Coordinator | Annual | Systemic risk assessment |
| UK Online Safety Act | UK | Ofcom | Ongoing | Illegal content duties |
| PACT Act (US) | US | FBI + ASPCA / Local LE | 4 hrs | Animal cruelty content |
The regulatory table above shows the complexity. An analyst cannot be expected to memorize 10+ jurisdictions, each with different agencies, deadlines, and evidence requirements. This creates legal liability — a single missed NCMEC report is a criminal violation under 18 U.S.C. § 2258A, regardless of the analyst's intentions or workload.
SentinelCore response: System auto-reroutes cases by content type to the correct queue. Displays the relevant reporting agency based on the case's jurisdiction — NCMEC for US CSAM, IWF for UK, INHOPE for EU. Shows a live SLA countdown timer so the analyst always knows time remaining before a statutory deadline is breached. Mandatory reporting checkboxes are force-checked (non-optional) for zero-tolerance categories. → Full solution: Urgency-Based Queue Segmentation (V1)
3.2 Human Challenges — The Analyst
Research into Trust & Safety analyst working conditions revealed systemic issues that directly impact enforcement quality and workforce retention.
Eye Fatigue & Visual Overload
Analysts reviewing graphic content for 8-hour shifts report eye strain, desensitization, and cumulative trauma. Research from Thorn and the Technology Coalition recommends minimum-necessary-viewing principles: never show raw content when metadata and classifier output can inform the decision.
SentinelCore response: B&W grayscale toggle, default-blur for all high-severity media, permanent blur-lock for CSAM. → Full solution: Media Queue (V2)
Psychological Trauma & Burnout
CSAM reviewers have significantly higher PTSD rates than the general population. Industry best practice (Meta, YouTube, TikTok) mandates mandatory wellness breaks and limits daily media exposure. Without a dedicated media queue with exposure controls, every analyst is exposed to traumatic content at the same frequency.
SentinelCore response: Dedicated Media Queue with exposure controls per shift, hash-first workflow, forensic indicators instead of raw content. → Full solution: Media Queue & Analyst Wellness (V2)
Cognitive Load & Decision Fatigue
Research on T&S operations shows that analysts making 200+ enforcement decisions per shift experience significant accuracy decline in the final 2 hours. Multi-column interfaces with 40+ simultaneous data points exacerbate this — the brain processes information sequentially, not in parallel.
SentinelCore response: Single-column content-first layout, tiered information hierarchy, auto-close eligibility for high-confidence cases. → Full solution: Intelligence-Led Investigation (V5)
3.3 Technical Challenges — The Attacker
Research into threat actor behavior revealed specific technical gaps that attackers exploit to bypass detection systems. Each finding directly informed a SentinelCore feature.
Language & Dialect Gaps in Machine Translation
Machine translation systems (Google MT, Microsoft Translator) routinely fail on regional dialects, dual-meaning phrases, and coded extremist language. Arabic content using Egyptian colloquial phrases — where 'الصفوف' (al-sufuf) literally means "rows" but in extremist context means "fighting ranks" — is systematically misclassified as benign religious discourse.
SentinelCore response: Native analyst override workflow, RTL display, cultural context flags, defensible audit trail. → Full solution: Global Linguistic Intelligence (V4)
QR Code Abuse to Bypass URL Detection
Attackers increasingly embed phishing URLs inside QR codes to bypass text-based link scanners. A QR code is an image — URL reputation databases never see the destination until the QR is decoded. OpenCV research shows that QR-based phishing grew 240% in 2024-2025 as link scanners became more effective against raw URLs.
SentinelCore response: QR decoding at intake, Google Safe Browsing + phishing DB, URL shortener expansion, domain age check. → Full solution: Media Queue with QR detection (V2)
Cross-Platform Proliferation & Synthetic Media
Threat actors use AI generation platforms (ElevenLabs, Runway, Midjourney) to create synthetic abuse content, then distribute across multiple consumer platforms. Isolated moderation removes individual copies — the origin generator account and the synthetic asset remain active. C2PA provenance standards exist but adoption is inconsistent; most synthetic media has stripped metadata.
SentinelCore response: C2PA acoustic scanning, origin API takedown, perceptual hash sharing with Lantern DB & GIFCT, Lumen Database reporting. → Full solution: SentinelCore Federation (V3)
Offender Attribution & Ban Evasion
Attackers create new accounts within minutes of a ban by rotating IPs (VPN), emails (name+1, name+2 pattern), and clearing cookies. Isolated per-case review cannot detect this — the analyst sees a "first offense" from a "new user" when it is actually a serial offender. Google's Project Strobe documented that 94% of ban-evading accounts reuse at least one device attribute (canvas hash, WebGL fingerprint, font list). NYU research demonstrates 87% cross-account identification via stylometric analysis alone. Without cross-account attribution, platforms inadvertently grant clean slates to repeat offenders with every ban — undermining every enforcement action they take.
SentinelCore response: Offender Graph cross-references device fingerprints, IP subnets, email naming patterns, payment instruments, geo-coordinates, and behavioral signatures against a banned-account database. Linked accounts surface in the investigation sidebar with match confidence scores. Currently simulated via deterministic hash-based linking; full implementation would integrate with device attestation APIs (WebAuthn) and behavioral ML models.
2. Architecture Index
Core Foundations
- Policy Playbook — 26-category taxonomy with decision trees
- Decision Tree Engine — Branching a/b enforcement logic per violation
- Enforcement Ladder System — Policy-determined actions, 7 tier categories
- SentinelCore Dashboard — Live investigation environment
Operational Infrastructure
- 4 Queue Form-Factor Segmentation — Regulatory, Standard, Media, Appeals
- Unified Investigation Layout — 5-component HUD, 3-step resolution
- OSINT & Threat Enrichment — IP geolocation, VPN detection, behavioral velocity
- Agency Reporting Engine — 26 policies × 4+ jurisdictions mapped
- Ops Health Panel — Live queue metrics (backlog, SLA%), intake charts, team status, system health monitoring
Advanced Features
- Cross-Platform Federation — Signal sharing with ElevenLabs, Lantern DB, Lumen, Google Safe Browsing
- Linguistic Intelligence — Dialect-aware LLM analysis, native analyst override workflow
- Auto-Triage & Confidence Routing — ML-driven prioritization with 3-tier color-coding
- Second-Reviewer Governance — Mandatory approval gate for Critical-tier cases
5. The Evolution Ledger
Each entry documents the strategic challenge that drove a phase, what was built, and how it was resolved.
V0 — Policy Foundations & Unified Queue
Implemented
Core Implementation: Policy Playbook — 16-category taxonomy (later expanded to 26). A single unified investigation queue handling all violation types — hate speech, CSAM, DMCA, fraud — in one interface. Appeals queue for user disputes. Policies were a single undifferentiated pool — no queue-based organization.
Strategic Challenge: Inconsistent manual moderation. Analysts applied different standards to the same violation type. No single source of truth for enforcement decisions. One queue for everything meant CSAM sat next to spam — no prioritization, no specialization. All 16 policies existed in one flat list — analysts had to memorize which required LE reporting.
Resolution: Built a strict policy-to-decision tree mapping system. Every violation type links to a playbook entry with numbered decision steps. Established the Appeals queue as a formal secondary review pathway. But the flat policy structure remained — this would be addressed in V1 when queues demanded policy alignment.
V1 — Urgency-Based Queue Segmentation
Implemented
Core Implementation: Split the single unified queue into Regulatory Queue and Standard Queue. Simultaneously reorganized the policy taxonomy: 7 policies assigned to Regulatory (violent_extremism, self_harm, election, CBRN, DMCA, data_privacy, public_threat), 6 to Standard (hate, fraud, cyber, harassment, regulated_goods, sexual_content_text). Each policy now had a designated queue based on its operational protocol — legal mandates route to Regulatory, volume violations route to Standard.
Strategic Challenge: Urgency mismatch AND policy-protocol mismatch. A CSAM report requiring 15-minute NCMEC reporting sat next to a 72-hour spam report. But also: a DMCA takedown follows a civil legal process (notice → takedown → counter-notice), while hate speech follows a graduated enforcement ladder (warn → remove → suspend). These belong in different operational environments with different workflows, different SLA pressures, and different analyst training requirements.
Resolution: Segmented by operational protocol, not just urgency. Each policy was assigned to the queue that matched its enforcement pattern — civil processes to Regulatory, graduated enforcement to Standard. Agencies are jurisdiction-mapped per policy — NCMEC for US CSAM, IWF for UK, INHOPE for EU. The queue determines the analyst's workflow: Regulatory analysts see deadline pressure and mandatory reporting checklists; Standard analysts see enforcement ladders and volume metrics.
V2 — Media Queue & Analyst Wellness
Implemented
Core Implementation: Added a dedicated Media Queue, requiring a significant policy realignment. CSAM, nudity, and graphic_violence were moved from Regulatory/Standard into Media. New policies were created to complete the queue: synthetic_media, qr_code_violations, and animal_abuse — each with full playbook dossiers. The Media queue now held 6 policies with distinct forensic protocols. Implemented analyst safety: default-blur for high-severity, permanent blur-lock for CSAM (no unblur toggle), B&W grayscale for graphic violence, hash-first workflow where automated matching occurs before human viewing.
Strategic Challenge: Dual challenge — analyst wellness AND policy-protocol clarity. CSAM was sitting in the Regulatory queue because of its mandatory reporting requirement, but its investigation protocol is visual (hash matching, PhotoDNA, perceptual fingerprinting) — not legal (document review, jurisdiction checks, statutory deadlines). The policy-to-queue mapping violated operational logic: the legal mandate (NCMEC reporting) overrode the actual investigation workflow (visual forensic analysis).
Resolution: Reassigned policies based on investigation protocol, not just legal classification. CSAM moved to Media because its primary workflow is hash-first visual analysis — the mandatory NCMEC reporting is preserved as a forced requirement within the Media queue's resolution step. This established the principle: a policy's queue assignment reflects HOW you investigate it, not just WHO you report it to. Media queue enables load distribution: not every analyst handles media every shift — exposure is limited per analyst wellness protocols.
V3 — SentinelCore Federation
Implemented
Core Implementation: Cross-platform federated threat intelligence — C2PA provenance scanning, acoustic fingerprint matching, and signal federation with industry clearinghouses (ElevenLabs API, Lantern DB, Lumen Database, Google Safe Browsing). Demonstrated via Case MED-999 (ElevenLabs voice clone → C2PA scan → origin API takedown → Lantern hash sharing) and Case MED-022 (YouTube music video piracy → DMCA takedown → Lumen Database transparency report → YouTube T&S notification).
Strategic Challenge: Cross-platform abuse proliferation. A bad actor generates synthetic audio on ElevenLabs, uploads to our platform, and distributes links via Telegram. Isolated moderation only removes our copy — the origin remains active.
Resolution: Built API-driven signal sharing protocol. On synthetic media detection: (1) validate provenance via C2PA/acoustic scan, (2) flag origin platform, (3) generate perceptual hash, (4) transmit to industry clearinghouses. All while maintaining GDPR/SCA compliance — no PII shared between consumer platforms.
V4 — Global Linguistic Intelligence
Implemented
Core Implementation: Multi-lingual translation tooling with native analyst override. Egyptian Arabic dialect detection, machine translation comparison, and cultural context flagging for coded extremist recruitment phrases. Demonstrated via Case REG-030 — an ISIL recruitment post where machine translation misclassified 'الصفوف' as "the path" instead of the correct "fighting ranks."
Strategic Challenge: Global dialect/slang gaps in machine translation. Arabic content containing 'الصفوف' (al-sufuf) was machine-translated as 'the path' — its literal meaning. A native Arabic analyst identified it as a known ISIL recruitment phrase meaning 'the fighting ranks.'
Resolution: Integrated contextual LLM analysis for non-English content. The system displays original text (RTL), machine translation (flagging known failure modes), and native analyst override — with credentials, linguistic rationale, and cultural context documentation. This creates a defensible audit trail for regulatory reporting.
V5 — Intelligence-Led Investigation
Implemented
Core Implementation: ML-driven auto-triage, confidence-based queue color-coding, auto-close eligibility for high-confidence hash-matched cases, second-reviewer governance gate for Critical-tier cases, and classifier feedback loop integration.
Strategic Challenge: High analyst cognitive load. Reviewers were processing every case manually — including clear hash-confirmed CSAM matches that require zero human judgment. This creates operational bottlenecks for cases that actually need human expertise.
Resolution: Implemented confidence thresholds and auto-routing. Cases with ≥90% AI confidence AND confirmed hash matches are flagged auto-close eligible. Critical cases require second-reviewer approval. Every closed case feeds back into classifier training. Color-coded queue cards give instant visual triage priority.
V6 — Emergency Auto-File & Retraction Queue (Planned)
Future Phase
Core Implementation: SLA-gated emergency auto-submit protocol. When a mandatory-reporting case (CSAM, human trafficking, CBRN) approaches its statutory deadline without analyst resolution, the system auto-files to NCMEC/FBI/INTERPOL with available metadata + "Supplemental to Follow" flag. Dedicated Retraction & Correction Queue surfaces all auto-submitted cases within 24 hours for human review: supplement with full evidence, or formally retract false positives. Gates on hash-match confirmation only — AI-only detections never auto-submit.
Strategic Challenge: The legal reality: wrong submission in good faith is safer than late or no submission. 18 USC §2258A creates criminal liability for missed CSAM reports. GenAI can theoretically produce content that hash-matches known CSAM files — a false positive that would trigger an erroneous NCMEC report. The system must distinguish between auto-file (hash-confirmed — always file) and human-reviewed (AI-detected only — never auto-file).
Proposed Resolution: Tier 1 auto-submit: hash-match + SLA < 2min → emergency file with metadata. Tier 2 human augmentation: within 24h, analyst reviews content. If correct → supplement. If GenAI false positive → retract via CyberTipline correction. If uncertain → flag for legal review. The Retraction Queue is not a failure mode — it's a safety net that enables aggressive auto-filing without fear of irreversible error.
V7 — Role-Based Access Control (Planned)
Future Phase
Core Implementation: A comprehensive RBAC system that gates every dimension of the investigation workflow. Analyst profiles define: language proficiency (auto-route cases by dialect), queue clearance (Media requires wellness certification), policy authority (CSAM = certified reviewers only, CBRN = SME-only), SME status (enables cross-queue reclassification and appeals), wellness caps (monthly psych evaluation score → auto-adjusted daily media limits), evidence viewing permissions (CSAM unblur requires CSAM certification + second-reviewer approval), escalation authority tiered L1→SME→Legal→Executive, reporting authority (auto-submit = system, human file = SME+, retraction = Legal only), second-reviewer eligibility (must be SME, different shift, not case owner), jurisdiction routing (GDPR→EU-trained, UK DPA→UK-trained), audit access (self-only default, team lead = team, Legal = all), and certification expiry auto-revoke (CSAM cert lapses quarterly → auto-removed from Media queue).
Strategic Challenge: Without RBAC, every analyst has identical access — CSAM reviewers and spam reviewers see the same queues, the same content, with the same workload. This creates three failure modes: (1) uncertified analysts viewing CSAM, (2) burned-out analysts receiving more media cases than their wellness score permits, and (3) language-mismatched analysts processing Arabic content through machine translation only — the very gap REG-030 exposed. A single analyst profile must become a multi-dimensional clearance matrix that dynamically gates routing, viewing, enforcement, reporting, and workload.
Proposed Resolution: Build an Analyst Profile System with per-dimension clearance levels. Cases are auto-routed by matching analyst language + jurisdiction + queue clearance. A Wellness Gating engine reads monthly evaluation scores and auto-adjusts media case caps. A Certification Tracker monitors expiry and auto-revokes clearance. A Workload Balancer distributes cases across eligible analysts based on current caps. A rigid Escalation Chain (L1→SME→Legal→Executive) prevents unauthorized escalation. The net effect: an analyst only sees cases they are qualified, certified, and well enough to handle — reducing liability, improving accuracy, and protecting analyst health.
6. Known Policy Gaps
The 26-policy taxonomy covers the Trust & Safety landscape comprehensively — but gaps remain. These are identified categories that major platforms enforce but SentinelCore does not yet address. Each represents a genuine harm vector requiring dedicated policy definition, enforcement protocol, and investigation workflow.
Gaps Closed — Implemented
Impersonation & Identity Fraud
Originally missing: No dedicated impersonation policy. Early versions had only fraud for financial scams — no coverage for parody-without-disclosure, brand impersonation, private-individual impersonation, or verified badge circumvention.
Closed: Added as standalone Standard queue policy with 10 context rules covering public figures, private individuals, brands, parody disclosure requirements, profile cloning, and platform staff impersonation. V1 phase.
Status: Implemented. Full playbook dossier with 3-step decision tree, 4 examples, and 10 context rules.
Human Exploitation & Trafficking
Originally missing: No trafficking policy. FOSTA-SESTA and EU Anti-Trafficking Directive require mandatory LE reporting — early versions had zero coverage for sex trafficking, labor trafficking, forced marriage, or human smuggling.
Closed: Added as Regulatory queue policy with 9 context rules covering sex trafficking, labor trafficking, forced marriage, human smuggling, organ trafficking, survivor content exemptions, and mandatory NCMEC/FBI/National Hotline reporting within 15 minutes. V1 phase.
Status: Implemented. Full playbook dossier. Mandatory LE reporting forced-checked in resolution step.
Spam & Inauthentic Behavior
Originally missing: No spam/inauthentic policy. Early versions had fraud for scams but nothing for bulk unsolicited messaging, fake accounts, bot networks, engagement manipulation, or coordinated inauthentic behavior (CIB).
Closed: Added as Standard queue policy with 10 context rules covering bulk spam, sockpuppet accounts, bot networks, engagement manipulation, CIB, review bombing, follower inflation, scraping accounts, and parody/satire disclosure. V1 phase.
Status: Implemented. Full playbook dossier. EU DSA Articles 34-35 systemic risk assessment covered.
Gaps Identified — In Progress
Online Grooming & Enticement of Minors (Text-Based)
Current coverage: csam covers visual child sexual abuse material (Media queue). human_trafficking covers trafficking/exploitation (Regulatory queue). sexual_content_text covers adult solicitation (Standard queue). No policy addresses text-based grooming — an adult building a relationship with a minor through chat, messages, or comments for the purpose of sexual exploitation.
Why it matters: NCMEC received 1.4 million online enticement reports in 2025 (+158% YoY). Grooming is the pre-trafficking stage — the predator builds trust, isolates the child, and moves toward exploitation. The evidence is chat logs and message patterns, not images. Major platforms (Meta, TikTok, Discord) enforce this as a standalone policy category because the investigation workflow is fundamentally different from both CSAM (visual hash matching) and human trafficking (adult exploitation, forced labor).
Recommended placement: Standard queue — text-based linguistic analysis and behavioral pattern detection. Mandatory NCMEC reporting is preserved as a forced requirement in the resolution step. This follows the same precedent as CSAM: the queue reflects how you investigate (text forensics vs. visual hash matching), not who you report to (NCMEC in both cases). Would integrate with NCMEC CyberTipline for online enticement reports (separate category from CSAM).
Status: Identified — in progress. Appears in Research Foundation NCMEC 2025 statistics. No dedicated policy dossier, enforcement ladder, or investigation workflow yet.
7. Design Decisions & Discarded Paths
Not every architectural decision survives contact with reality. These paths were explored, prototyped, and ultimately discarded.
Case Review HUD — 4-Step Wizard Mode
Built: A dedicated 4-step wizard with 3 hardcoded mock cases, SLA countdowns, and auto-advance between cases.
Discarded: The wizard pattern created visual fragmentation. Each step replaced the previous one — the content was not always visible.
Lesson: Multi-step wizards work for forms, not for evidence review. Content must always be visible.
3-Column Permanent Panel Layout → Single-Column Content-First
Built: A 3-column grid (Signals 240px | Investigation | OSINT 260px) with all panels permanently visible. 5 always-open panels showing 40+ data points simultaneously.
Discarded: This was the most fundamental architectural decision of the project. We built it, tested it with real case data, and concluded it was wrong. The layout created cognitive overload — analysts processed signals, OSINT, and enforcement simultaneously rather than sequentially. Real T&S investigation tools (Thorn, Hive, Zendesk) follow a content-first hierarchy where the reported evidence dominates 60%+ of the viewport.
Lesson: Information hierarchy is non-negotiable. Tier 1 (content + detection) must always be visible. Tier 2 (policy guidance, enforcement) surfaces when relevant. Tier 3 (OSINT, user history, linked accounts) is on-demand only. The pivot from 3-column to single-column with collapsible accordion sidebar was the single highest-impact UX decision in the project.
Generic 5-Tier Enforcement Slider
Built: One-size-fits-all Ignore → Warn → Mute → Suspend → Ban slider for every policy.
Discarded: CSAM has exactly one option. DMCA is a civil process, not a punishment ladder. Generic sliders create liability risk.
Lesson: Enforcement must be policy-determined. The system should only surface legally and operationally valid actions.
CSAM Queued to Regulatory — Wrong Protocol
Built: CSAM was assigned to the Regulatory queue because of its NCMEC mandatory reporting requirement.
Refactored: CSAM's investigation workflow is visual (hash matching, PhotoDNA, perceptual fingerprinting) — not legal. The legal mandate (NCMEC reporting) had overridden the actual investigation protocol. Moved CSAM to Media queue, preserving forced NCMEC reporting within the Media resolution step.
Principle established: A policy's queue assignment reflects HOW you investigate it, not just WHO you report it to.
Reclassification Locked to Current Queue
Built: Early reclassification dropdowns only showed policies from the case's current queue — a Standard case could only be reclassified to other Standard policies.
Refactored: Analyst feedback: "What if this Standard case is actually Regulatory?" The restriction was artificial — a misclassified case should be reclassified to ANY applicable policy. Reclassify now shows all 26 policies organized by 3 queue tabs.
Principle established: Classification is the analyst's judgment call. The system should enable correction across queues, not enforce the initial routing.
Mandatory Audit Notes Block Resolution
Built: Resolution buttons were disabled until the analyst entered audit notes — marked with a red asterisk as required.
Refactored: Discovered this creates dangerous friction for time-critical cases. An analyst handling a CSAM report with a 15-minute NCMEC deadline should not be blocked from resolving while they compose notes. Notes were made optional — the resolve buttons are always active.
Principle established: Safety-critical actions (reporting, blocking, escalating) cannot be gated by documentation fields. Documentation can follow resolution.
Landing Page as Queue Gate
Built: The dashboard opened on a landing page with executive summary. Analysts clicked "Launch Live Queue" to access cases.
Refactored: This added an unnecessary click for returning users. The landing page is reference material — the live queue is the operational tool. Dashboard now defaults directly to the live queue. Landing page accessible via header link for first-time visitors.
Principle established: Operational tools should open directly to operations. Reference material is secondary.
CSAM Blur-Lock: One-Size-Fits-All → Hash-Gated
Built: All CSAM cases were permanently blurred with no unblur toggle — regardless of detection method. Hash-confirmed CSAM sat alongside AI-flagged CSAM under the same restriction.
Refactored: Hash-confirmed CSAM (PhotoDNA/DNAHash match) stays permanently locked — mathematically verified, no analyst discretion needed. AI-suspected and user-reported CSAM gets a controlled reveal path: default-blur with an amber "CSAM Suspected — Awaiting SAG Review" notice, metadata-first forensic review, and SAG-gated unblur access for certified reviewers only.
Principle: Hash-match is mathematical certainty — no visual confirmation needed. AI classifier outputs are probabilistic — analyst judgment is required. The blur gate must reflect what is known, not just what is suspected.
8. Future Roadmap
Advanced Real-Time Synthetic Deepfake Detection
Integrate real-time C2PA verification at upload time. Expand acoustic fingerprinting to video provenance (Runway, Sora). Build cross-referencing against known synthetic media clearinghouse databases.
Automated Risk-Scoring for User Reputation
Build a User Trust Score factoring account age, violation history, device fingerprint, behavioral velocity, and cross-platform activity. Auto-escalate high-risk accounts before content is posted.
Expanded Regulatory Jurisdiction Mapping
Add real-time legal requirement lookup per jurisdiction. Auto-populate mandatory reporting agencies, statutory deadlines, and documentation templates per country code.
Real-Time Collaborative Investigation
Enable multi-analyst case collaboration for complex cross-queue investigations. Shared case notes, real-time status updates, and escalation handoff between queues.
SLA-Gated Emergency Auto-File Protocol
When a mandatory-reporting case (CSAM, human trafficking, CBRN) approaches its statutory deadline without analyst resolution, the system auto-files an emergency submission to the relevant agency (NCMEC, FBI, INTERPOL) containing all available metadata, hash matches, and an "Emergency Auto-Submit — Supplemental Report to Follow" flag. This prevents criminal liability for late or missed reporting. A dedicatedRetraction & Correction Queuesurfaces all auto-submitted cases for human review within 24 hours. Analysts can supplement with full evidence, or — in the case of a false positive (e.g., GenAI content triggering a false hash match) — formally retract via the agency's correction mechanism. Auto-submit gates on hash-match confirmation only — AI-classifier-only detections never auto-submit. The hierarchy is explicit: wrong submission in good faith is legally safer than late or no submission.
OSINT-Based Offender Attribution & Ban Evasion Detection
Build a real-time cross-account attribution engine that queries banned-account fingerprints at case creation time. Integrate device attestation (canvas hash, WebGL fingerprint, font list, screen resolution — 94% unique per Google Project Strobe), IP subnet overlap detection, email naming pattern regex (name+1, name+2 detection), payment instrument hashing, geo-coordinate clustering, and behavioral stylometric analysis (87% author identification accuracy per NYU research). When a match is found, the case is automatically flagged as a repeat offender with full linked-account history — preventing analysts from unknowingly treating a serial violator as a first offense. This transforms the current simulated OffenderGraph into a live attribution pipeline.
SentinelCore · System Architecture & Evolution · Internal Document · June 2026