End-to-End Risk Analytics & Operational Governance Case Study
A complete risk analytics project focused on enforcement scale, automation exposure, internal appeals and quality-governance design using harmonised transparency reports published under the European Union Digital Services Act.
The project combines Python, SQL, Excel and operational risk analysis to transform five platform disclosures into a reproducible cross-platform benchmark, an executive case study and a practical human-review governance framework.
View the full case study · Download the analytical workbook
Large digital platforms publish substantial volumes of enforcement and appeal data, but harmonised reporting does not automatically produce decision-ready insight.
This case study audits the H2 2025 transparency reports of TikTok, Instagram, YouTube, Snapchat and LinkedIn across enforcement volume, automation, internal complaints, appeal outcomes, restriction severity and reporting quality.
The analysis identifies where operational review should be prioritised when four signals converge:
- high enforcement intensity;
- high end-to-end automation;
- elevated reversal outcomes;
- high-impact account, service or monetary restrictions.
Rather than treating reversals as proof of error, the project translates public disclosures into an investigation and quality-assurance framework with explicit methodological guardrails.
The project addresses five principal operational questions:
- How concentrated is enforcement volume after normalisation by active-recipient metrics?
- Where is end-to-end automated enforcement most exposed?
- How do resolved complaint outcomes vary across platforms?
- Which categories and restriction types create the highest potential review impact?
- Where should quality assurance and human review be prioritised first?
| KPI | Value |
|---|---|
| Platforms Benchmarked | 5 |
| Core Harmonised Tables Audited | 55 |
| Reported Enforcement Measures | 359.85M |
| Weighted Automation Exposure | 85.24% |
| Internal Complaints | 24.77M |
| Reversed Decisions | 10.80M |
| Reversal Share Among Resolved Complaints | 43.66% |
| Highest Category Automation | 99.69% |
| Highest Review Priority Score | 65.1 |
A reversal is treated as an investigation signal, not proof that the original decision was erroneous.
The executive view consolidates the main scale, automation, appeal and control-priority metrics used throughout the case study.
It includes:
- reported enforcement measures;
- weighted automation exposure;
- internal complaints and reversed decisions;
- resolved reversal share;
- category-level automation signals;
- a decision statement for risk-tiered human review.
Harmonised DSA Reports
↓
Source and Schema Audit
↓
Data Cleaning and Harmonisation
↓
Hierarchy and Reconciliation Controls
↓
Platform / Category / Decision Master Tables
↓
SQL and Python Validation
↓
Scale, Automation and Appeal Analysis
↓
Enforcement Review Priority Score
↓
Operational Recommendations
↓
Monitoring Framework and 90-Day Roadmap
- Python
- SQL
- Microsoft Excel
- Git and GitHub
- Matplotlib
- Google Sheets
- Google Docs
- Data Cleaning
- Schema Validation
- Cross-Platform Harmonisation
- Exploratory Data Analysis
- KPI Design
- Data Reconciliation
- Category Hierarchy Control
- Normalised Benchmarking
- Appeal Outcome Analysis
- Risk Scoring
- Data Visualisation
- Executive Reporting
- Operational Recommendation Development
TikTok represents the largest concentration of reported own-initiative enforcement measures in the five-platform sample.
- Reported measures: 258.69M
- Share of selected sample: 71.89%
- Measures per 1M active-recipient proxy: 1,450,878
The normalised metric measures enforcement intensity, not unique-user incidence. A single account, recipient or content item can generate multiple measures.
Aggregate automation is high across all five platforms, but category-level deployment varies materially.
| Platform / Category | Measures | Automated Exposure |
|---|---|---|
| YouTube - Protection of minors | 11.16M | 99.69% |
| Instagram - Civic discourse / elections | 5.96M | 2.45% |
| Instagram - Privacy violations | 1.97M | 3.32% |
| LinkedIn - Cyber violence | 23.5K | 0.98% |
| Snapchat - Cyber violence | 618.8K | 0.15% |
This range supports category-specific review thresholds rather than a single platform-wide automation policy.
The five platforms reported 24.77M internal complaints and 10.80M reversed decisions.
| Platform | Complaints | Reversed | Reversal Share Among Resolved |
|---|---|---|---|
| 151,062 | 104,174 | 70.65% | |
| TikTok | 15,411,338 | 7,114,588 | 46.16% |
| 8,668,843 | 3,454,739 | 39.85% | |
| YouTube | 408,939 | 110,187 | 29.15% |
| Snapchat | 130,767 | 15,404 | 11.78% |
These rates are not platform-wide error estimates. Complaint populations are self-selected, may include timing effects and cannot be linked to individual enforcement actions in the public reports.
LinkedIn reports a particularly high account, service and monetary restriction mix, while TikTok's much larger volume is primarily visibility-related.
- LinkedIn severe-restriction mix: 92.03%
- Instagram severe-restriction mix: 53.41%
- YouTube severe-restriction mix: 9.48%
- Snapchat severe-restriction mix: 3.53%
- TikTok severe-restriction mix: 1.69%
This distinction is central to human-review design: higher-impact decisions require stronger precision, escalation and audit controls.
The project creates an Enforcement Review Priority Score using:
- 30% normalised enforcement intensity;
- 25% automation exposure;
- 30% reversal share among resolved complaints;
- 15% severe-restriction mix.
| Platform | Review Priority Score | Operational Interpretation |
|---|---|---|
| TikTok | 65.1 | Highest scale-driven review priority |
| 58.2 | High reversal and account-impact priority | |
| 43.5 | Targeted priority with material data caveats | |
| YouTube | 34.5 | Category-level QA priority despite high automation |
| Snapchat | 24.6 | Lower aggregate priority; targeted controls remain relevant |
The score ranks review need, not platform quality, fairness or regulatory compliance.
| Recommendation | Operational Design | Priority |
|---|---|---|
| Build an appeal-linked quality loop | Map appeals to policy, model, action type and reviewer cohort | High |
| Introduce risk-tiered human review | Raise review thresholds for severe account, service and monetary actions | High |
| Use category-specific automation thresholds | Tune sampling and confidence requirements by category and harm severity | High |
| Implement reporting reconciliation controls | Add schema, hierarchy, period and total-to-subtotal validation | High |
| Establish a governance dashboard and review cadence | Run weekly operational reviews and monthly risk-governance reviews | Medium |
The project translates transparency reporting into a three-layer operating model:
| Layer | Purpose |
|---|---|
| Input Layer | DSA reports, appeals, automation, AMAR and taxonomy |
| Control Layer | Schema, hierarchy, reconciliation, versioning and exceptions |
| Review Layer | Weekly operations, monthly risk and quarterly governance |
| Timing | Phase | Deliverable |
|---|---|---|
| Days 0-30 | Instrument | Publish definitions, map appeals, define severity weights and implement data-quality tests |
| Days 31-60 | Pilot | Launch risk-based samples, category thresholds and root-cause analysis for reversed severe actions |
| Days 61-90 | Evaluate and Scale | Measure quality lift, queue cost and SLA impact; scale only positive controls |
The operating principle is to escalate based on evidence, impact and repeatability - not raw volume alone.
platform-enforcement-reliability-audit/
├── README.md
├── Case_Study_02_Platform_Enforcement_Reliability_Audit.pdf
├── MANIFEST.json
├── dashboard/
│ ├── README.md
│ └── dashboard_preview.png
├── data/
│ └── processed/
│ ├── platform_summary.csv
│ ├── category_master.csv
│ ├── decision_type_master.csv
│ ├── data_quality_checks.csv
│ ├── data_quality_log.csv
│ ├── primary_source_verification.csv
│ └── source_registry.csv
├── docs/
│ ├── DATA_SOURCES.md
│ ├── phase1_data_audit.md
│ ├── phase2_master_dataset.md
│ ├── phase3_analysis.md
│ └── phase4_recommendations.md
├── scripts/
│ ├── build_master_dataset.py
│ └── create_charts.py
├── sql/
│ ├── README.md
│ ├── 01_platform_benchmark.sql
│ ├── 02_category_concentration.sql
│ ├── 03_appeal_reconciliation.sql
│ └── 04_automation_exposure.sql
└── workbook/
└── Platform_Enforcement_Audit_Data.xlsx
- Platform-level benchmark
- Category-level master table
- Decision-type master table
- Data-quality checks
- Primary-source verification
- Source registry
- Data source registry and attribution
- Phase 1 - Data audit
- Phase 2 - Master dataset design
- Phase 3 - Analysis findings
- Phase 4 - Recommendations and monitoring framework
- Master dataset build script
- Chart generation script
- Platform benchmark SQL
- Category concentration SQL
- Appeal reconciliation SQL
- Automation exposure SQL
The published repository includes processed analytical tables, SQL validation queries, chart-generation code and a full primary-source verification file.
The original provider ZIP, XLSX and ODS reports are not redistributed. Their official URLs, reporting periods and treatment notes are retained in:
data/processed/source_registry.csv
docs/DATA_SOURCES.md
Charts can be regenerated from the processed data using:
python scripts/create_charts.py \
--processed-dir data/processed \
--output-dir chartsThe master dataset build script is designed for the official H2 2025 provider reports organised by platform:
python scripts/build_master_dataset.py \
--input-root path/to/extracted_provider_reports \
--output-dir data/processedThe analytical controls validate:
- reporting-period alignment;
- source schema and expected tables;
- numeric normalisation;
- category hierarchy;
- appeal reconciliation;
- provider-total reconciliation;
- primary KPI reproduction;
- provider-specific caveats.
- A reversal is not treated as proof of an original error.
- Measures are actions, not unique affected users.
- Top-level categories are separated from keyword subcategories to prevent double counting.
- YouTube signed-in and signed-out recipient measures are not added together.
- Provider-defined taxonomies are not assumed to be fully identical.
- Instagram demotion and combined-service caveats are preserved.
- The same formulas and weights are applied to every platform.
- No internal TikTok data or non-public employer knowledge is used.
- Business Operations
- Risk Analytics
- Trust and Safety Operations
- SQL
- Python
- Microsoft Excel
- Data Cleaning
- Data Validation
- Cross-Platform Harmonisation
- Exploratory Data Analysis
- KPI Design
- Data Quality Management
- Automation Governance
- Appeal Outcome Analysis
- Operational Risk Prioritisation
- Dashboard Design
- Data Visualisation
- Executive Reporting
- Data Storytelling
- Business Recommendations
- Stakeholder Communication
- Reproducible Analytics
- GitHub Project Documentation
Rayan Braccio
Business Operations, Data Analytics and Risk professional with experience supporting high-volume platform operations.
Following several years in Trust & Safety and Monetization Integrity Operations, this portfolio demonstrates a transition toward Business Operations, Risk Analytics and Data Analytics by combining operational expertise with structured analysis, data validation and decision-oriented reporting.
This repository is part of a growing Business Operations and Data Analytics portfolio.
Related work:
If you found this project relevant, feel free to connect with me on LinkedIn.