Transparency Log
Transparency Logdata/transparency-log.json (not published as a public download)starnum.com.tw is committed to publicly and transparently recording system events, quality issues, and improvement measures. This page benchmarks against the transparency standards of the OpenAI Postmortem page, Google Transparency Report, and Anthropic Usage Policy Violation Summaries.
Current System Status
| Service | Status | Last Confirmed |
|---|---|---|
| Content Production System | Operational | 2026-04-12 |
| Chart Analysis Service | Operational | 2026-04-12 |
| Knowledge Base (KB) Index | Operational | 2026-04-12 |
| Quality Review System | Operational | 2026-04-12 |
| multi-provider governance audit | Operational (R15 4/4 CLEAN) | 2026-04-12 |
| Supabase Database | Operational | 2026-04-12 |
| Cloudflare CDN | Operational | 2026-04-12 |
| Neo4j Knowledge Graph | Operational (343 nodes / 681 relations) | 2026-04-12 |
| Qdrant Vector Search | Operational (12,378 chunks) | 2026-04-12 |
Incident Record (2026)
1 recorded incident | 0 ongoing | All major incidents publicly disclosed within 72 hours
INC-2026-001: Plain-Language Output Format Violation
Severity: Low (quality issue, not data loss or security problem)
Impact: An AI agent used non-standard formatting in chart plain-language analysis (【title】 format + overly short description tags), causing some output to fail our format specification.
Root Cause: The agent did not read the required analysis-output-spec.md skill pack, relying on an outdated template from memory. This resulted in bh-tags with 3–5 character short words (spec requires 12–22 character descriptive sentences) and use of the prohibited 【title】 format at block beginnings.
Mitigations:
- Added "do not write formats from memory" warning to
chart-analysis.md - Added explicit source annotation at the top of
templates/chart-analysis-prompt.mdpointing toanalysis-output-spec.md - Created
prompts/common-agent-base.mdmandatory SOP read-confirmation mechanism - Created
scripts/validate-output-format.jsautomated format validation (exit 1 forces rejection)
Resolution Time: 2026-04-10 (resolved same day)
Follow-up: The validation script has been integrated into the CI/CD pipeline for every article, preventing the same issue from recurring.
No other major incidents recorded.
No records does not mean no mechanisms — it means the system is operating normally.
This log is manually updated whenever there is a major incident; minor issues are summarized in monthly digests.
Transparency Standards Comparison
| Benchmark Item | Our Implementation | Status |
|---|---|---|
| Public incident record page | This page (transparency-log.html) | ✅ Complete |
| Machine-readable incident data | /data/transparency-log.json | ✅ Complete |
| Root cause analysis (postmortem) | Each incident includes full root cause + mitigation | ✅ Complete |
| Severity classification | CRITICAL / MAJOR / MINOR / INFO | ✅ Complete |
| Resolution time commitments | CRITICAL 4h / MAJOR 24h / MINOR 7d | ✅ Complete |
| AI-system-specific disclosures | AI agent behavior issues included in scope | ✅ Complete |
| JSON-LD Schema.org Report | Structured data injected into this page | ✅ Complete |
Incident Classification and Response Times
| Severity | Definition | Disclosure Deadline | Resolution Deadline |
|---|---|---|---|
| CRITICAL | User data breach, complete service outage, security vulnerability | Within 4 hours | Within 24 hours |
| MAJOR | Core function failure, chart analysis error rate >5%, major data consistency issue | Within 24 hours | Within 72 hours |
| MINOR | Below-standard quality, format issues, single function failure | Within 7 days | Within 30 days |
| INFO | System optimization, preventive maintenance, process improvements | Monthly digest | Per schedule |
AI-System-Specific Disclosures
This site uses a fully AI-automated architecture (Claude Code as technical lead). Therefore, the transparency log's scope includes special categories not applicable to conventional websites:
- AI agent behavior issues: Agents not following SOPs, output format violations, reasoning errors
- Model version impacts: The effect of LLM version upgrades/downgrades on output quality
- Joint audit findings: Systemic issues found by the multi-provider governance audit (Anthropic / OpenAI / Google Gemini governance benchmarks)
- Knowledge base contradictions: School conflicts or data quality issues found in the astrology KB
- Quality feedback loops: Content quality degradation signals from GA4 / GSC data
Related Pages
- Ethics Statement — Ethical framework and commitments for AI use
- AI Safety Methodology — Red team testing protocol and safety architecture
- AI Data Governance Policy — How user data is handled
- Security Disclosure Policy — How to report security vulnerabilities
- Acceptable Use Policy — Platform usage guidelines
- Content Production Methodology — Quality standards and technical architecture
- System Status — Real-time service status