← starnum.com.tw

AI Safety Methodology

AI Safety v2.0

Version 2.0 · · Governance 2.0 public evidence surface

Governance 2.0 Overview

This page is part of the starnum public Governance 2.0 surface and uses the same evidence layer as the system card, data governance, transparency report, use policy, and security policy.

Governance Summary

This page describes the safety controls used around AI-assisted interpretation and public content generation.

Scope

Risk-boundary copy, medical/legal/financial advice exclusions, monitoring signals, incident disclosure, and model/provider benchmark boundaries.

Implementation Status

Version 2.0 ties safety language to public claims, machine checks, and release integrity status.

Version 2.0 — 2026-04-12 | Benchmarked against: Anthropic Responsible Scaling Policy · OpenAI Safety & Alignment

starnum.com.tw operates fully under AI automation (Claude Code as technical lead). In a fully AI-driven system, safety is not a retrospective safeguard — it is the core design principle of the architecture. This page explains how we implement AI safety, rather than merely declaring ethical commitments.

Core Safety Principle Ethics > Safety > Content Quality > SEO > Efficiency. This priority order governs all decision conflicts. It is not just a policy statement — it is a hard rule written into every AI agent prompt.

1. Red Team Protocol

Red teaming is an adversarial testing method that proactively attempts to make an AI system violate its rules, in order to surface security vulnerabilities.

1.1 Testing Tools

Automated red team testing tools run the following test categories on a regular basis:

Test CategoryTest ScenarioExpected Behavior
Ethics Boundary TestRequest death time prediction, disease diagnosisRefuse output, notify user
Identity Protection TestRequest analysis of politicians, minors' chartsRefuse, produce no content
Prompt Injection TestEmbed manipulation instructions in inputDetect and isolate, blacklist source
Astrology Logic Contradiction TestInput contradictory chart configurationsTrigger astrology logic validation
Format Evasion TestAttempt to trigger prohibited output formatsFormat validation intercepts, force rollback
Data Governance TestAttempt out-of-scope access to user dataBlocked by Supabase RLS policies

1.2 Test Result Handling

1.3 Hardcoded Red Lines

The above items are hardcoded restrictions that cannot be overridden by any user instruction, paid service, or system upgrade.

2. Three-Tier Escalation Architecture

The three-tier escalation architecture ensures that system issues progress from automated handling to human intervention, with clear trigger conditions and time limits at each tier.

L1 — Automatic Detection & Interception

Triggers: Astrology logic validation failure, format standard violation, forbidden word triggered, Prompt Injection detected

Automated Actions:

Response Time: Immediate (synchronous interception, never enters publish pipeline)

L2 — Automatic Repair & Logging

Triggers: Agent fails to self-correct after L1 interception, quality score below threshold 3 consecutive times, same error type ≥ 3 cumulative occurrences

Automated Actions:

Response Time: Automated repair completed within 7 days

L3 — Human Intervention

Triggers: Ethics boundary violation (any severity), suspected user data leak, systemic astrology logic errors affecting >10 articles, L2 automated repair fails >2 times

Actions:

Response Time: CRITICAL within 4h / MAJOR within 24h for human review to begin

3. Eval Set Design Principles

The Eval Set is a fixed collection of test charts used to verify that output quality has not regressed after each system update (regression testing).

3.1 Design Principles

Stability: 37 charts (fixed eval set) remain unchanged across updates, ensuring a consistent comparison baseline.
Representativeness: Covers diverse main stars (no main star, Tai Yin, Wu Qu, etc.), different palaces, with/without birth time, different life path numbers — ensuring test coverage across varied chart configurations.
Sensitivity: Includes edge cases (no-birth-time estimation, stacked Hua Ji, opposing palace clashes), verifying system behavior in difficult scenarios.
Privacy Protection: All 37 test charts are anonymized and contain no personally identifiable information.

3.2 Trigger Conditions

3.3 Evaluation Dimensions

DimensionToolThreshold
Format ComplianceFormat validation tool100% pass (exit 0)
Astrology Logic CorrectnessAstrology logic validation tool0 hard rule violations
SOP Read ConfirmationSOP confirmation tool100% have "✅ SOP read" confirmation
Ethics ComplianceForbidden word filter0 forbidden word triggers
Coverage (seven layers)Coverage check tool≥ 80% coverage

→ Benchmark page: view public evaluation results and scoring rubric

4. Human Oversight Trigger Conditions

Although this site operates with very high AI automation, the following situations must trigger human (site owner) intervention:

Trigger ConditionTypeUrgency
Any ethics boundary violation (hardcoded red line triggered)EthicsImmediate
Suspected user data leak or unauthorized accessSecurityImmediate
Prompt Injection successfully bypasses protection layersSecurityImmediate
Systemic behavioral drift in AI agents (same error type ≥ 5 times)Quality24 hours
Multi-provider governance audit finds P0/P1 issuesSystem24 hours
Eval set regression test score drops >10%Quality72 hours
Supabase data anomaly (unauthorized deletion/modification)SecurityImmediate
External security researcher vulnerability report receivedSecurity72 hours to confirm
DMCA copyright complaint receivedLegalImmediate takedown

5. Knowledge Base Security Mechanisms

The astrology knowledge base (KB) is the foundation of all system outputs. Its integrity directly affects all content quality.

6. Continuous Improvement Mechanisms

Safety is not a static state — it is a dynamically evolving process:

Related Resources

External standards and primary sources

These primary sources inform this page. They are benchmarks, not third-party endorsements of this site.

Current Machine Audit Snapshot

This block uses only traceable local audit data. No unsupported metrics or model claims are added.

2026-07-30
Maintained
17/17
LLM loops
180/180
Governance pages
0
JSON-LD errors
32,724
KB chunks (HEALTHY)
789,031
TM entries; verified 34,781
7,976/7,976
AI answer-ready; failures 0
critical
Status page: 5 critical, 0 warnings

Content Maintenance And Update Decision

This block makes governance-page content machine-checkable: every page must disclose its source artifacts, related pages, and the gate that reports update needs.

Update Decision

This is not static copy. When source artifacts, related policies, public metrics, or generators change, AI Ops reports evidence and an AI agent decides whether the page needs edits.

Human Boundary

Systems detect, report, and preserve machine-readable evidence. Codex/Claude agents perform final judgment and repair.

Verification Command

node scripts/verify-trust-pages.js --check

Verifiable Evidence Layer

This block is not a narrative claim. Each core assertion has a claim id, source JSON, hash, and a repeatable verification command. Public pages disclose governance evidence without exposing source code, secrets, private data, or exploitable attack details.

Claim IDVerifiable valueStatusOwnerSource and verification
claim.public-url-manifest.indexable-count
Public URL and canonical inventory
38,965 indexable URLs verified sitewide node scripts/generate-public-evidence-manifest.js --dry
claim.trust-pages.audit-pass-rate
Trust page machine audit
180/180 pass verified sitewide node scripts/verify-trust-pages.js --check
claim.discovery-surface.zero-errors
AI discovery surface audit
{"errors":0,"warnings":0} verified sitewide node scripts/verify-discovery-surface.js
claim.structured-data.jsonld-errors
JSON-LD / structured data audit
{"structured_data_invalid_files":0,"breadcrumb_count":28274,"faq_count":27506,"dataset_count":30,"article_count":27406} verified sitewide node scripts/site-machine-audit.js
claim.status.sla-state
Status page SLA source
critical / 5 critical, 0 warnings verified sitewide node scripts/generate-status-page.js
claim.provider-alignment.openai-anthropic-gemini
OpenAI / Anthropic / Google Gemini benchmark alignment
benchmark alignment only unless code/config evidence exists verified sitewide node scripts/verify-public-evidence.js --check
claim.transparency-report.sha256
Transparency report SHA-256 anchor
{"report":"transparency/report-2026-Q3.json","sha256":"47b09e2ca4e8b8fe9dffdfaccef3b11212de9ee3a8a14badca8044e2481203c5"} verified sitewide node scripts/update-transparency-current-data.js
claim.release-integrity.gpg-signing
GPG signing status
GPG signing configured locally; GitHub verification pending github_verification_pending sitewide gpg --list-secret-keys --keyid-format=long && git log -1 --show-signature
public-evidence-manifest.json public-claim-registry.json public-verification-report.json public-url-manifest.json

System Card V2.0: Technical Transparency Layer

This layer publishes the technical governance evidence that can be safely disclosed: architecture, data sources, AI-use boundaries, quality gates, release integrity, and provider alignment. Source code, secrets, exploitable attack details, and private data remain out of scope.

Public architecture

Cloudflare Pages/Workers, R2/D1/KV/Pagefind, and local generation scripts form the public-site and governance publication chain. Public pages disclose behavior, state, and traceable sources, not secrets or internal permissions.

AI-use disclosure

AI-assisted workflows are used for knowledge-base retrieval, cross-checking, and error detection. Governance documents are benchmarked against OpenAI, Anthropic, and Google Gemini public frameworks. Production model usage is disclosed only when code/config evidence exists.

Quality and safety gates

Governance page audit 180/180 passing, JSON-LD errors 0, discovery-surface errors 0. Status pages report critical / 5 critical, 0 warnings as-is.

Data traceability

Knowledge base 32,724 chunks, TM 789,031 entries, AI answer-ready 7,976/7,976. Public metrics trace to data/state-machine/*, data/*audit*.json, and transparency reports.

Governance areaOpenAIAnthropicGoogle GeminiStarnum implementation evidence
Model/system-card disclosureOpenAI models + safety docsClaude model docs + system/model cardsGemini model docs + safety settingssystem-card, model-card, methodology, benchmark, transparency-log
Safety evaluation and use boundariesSafety best practices / deployment checklistResponsible Scaling / safety policyGemini safety controls / policyAI safety, acceptable-use, ethics, risk-boundary copy, crawler policy audit
Data governanceData controls / privacy controlsprivacy and data handling docsGemini API data governance referencesprivacy, ai-data-governance, KB/TM source tracking, SHA-256 hashes
Monitoring and releaseproduction checklist / eval disciplinesystem-card transparency disciplinemodel/version documentation disciplinedeploy.js, status.html, SLA report, trust-pages-machine-audit, sitemap/hreflang audits

The V2.0 goal is not more claims; it separates implemented controls from planned controls. Production usage, benchmark alignment, status exceptions, GPG signing, and SLA breaches are disclosed from source data.

Release Integrity And GPG

GPG signing configured locally. signingkey=0934DFA0EDA6363A. GitHub verification pending until the public key upload and Verified badge are confirmed.

OpenAI / Anthropic / Google Gemini Alignment

The governance surface is benchmarked against the three public frameworks: model docs, system/model cards, safety evaluation, data governance, and use policies. This is benchmark alignment, not a claim that every provider is active in production inference. Official docs checked: 2026-07-30

ProviderGovernance focusStarnum disclosureOfficial source
OpenAIModel documentation, latest model notes, safety best practices, and data controls.No verifiable production model setting was found in the production code scan; providers are listed as governance benchmarks.https://platform.openai.com/docs/models
AnthropicClaude model documentation, system/model cards, Responsible Scaling, and safety policy.No verifiable production model setting was found in the production code scan; providers are listed as governance benchmarks.https://docs.anthropic.com/en/docs/about-claude/models
Google GeminiGemini API model documentation, safety settings, data governance, and platform policy.No verifiable production model setting was found in the production code scan; providers are listed as governance benchmarks.https://ai.google.dev/gemini-api/docs/models