Research & Methodology

What This Page Actually Answers

Starnum's research page is not an academic journal or a marketing page. It is a public workbench explaining how we connect traditional astrology knowledge, chart calculation, AI-assisted writing, multilingual translation, AWR/TM scanning, and deployment gates. Readers should be able to tell here which judgments come from the astrology system, which are just engineering quality checks, and which still need further review by a human or an AI agent.

Every research claim is split into three layers: the first is cultural and astrological methodology, the second is data and tooling evidence, and the third is post-release verification results. Systems are only responsible for detecting, reporting, and preserving machine-readable evidence; whether to revise copy, retranslate, or adjust rules is decided by AI agents such as Codex/Claude.

Current Research Tracks

Research Evidence Specification

• Astrology methodology: Every astrological inference must be traceable to a school baseline, palace/star/Sihua conditions, and its uncertainty boundary.
• Translation quality: Multilingual pages must be checked against the zh-TW source text and classified as "needs full retranslation," "needs a partial patch," "an AWR false positive," or "TM terminology inconsistency."
• Public governance: Every public claim must link to data/public-claim-registry.json, a public evidence manifest, or a re-runnable verification command.
• Release verification: After a production deploy, the page must at minimum pass an HTTP smoke test, return 200 on key pages, have R2/Worker deployment records, cache purge, and a search-indexing submission.

Research Content Publication Rules

Data Sources and Responsibility Boundaries

The research page uses only traceable internal data and public artifacts, and never presents an off-the-cuff judgment as fact. Each type of data has a clear responsibility: astrology data owns methodology boundaries, machine audits own state evidence, AWR/TM own multilingual-quality signals, and AI agents alone own final judgment and repair.

Astrology and knowledge-base sources

docs/kb/zwds/, docs/kb/numerology/, and research articles define stars, palaces, Sihua, chart patterns, and numerology rules. This data can only support "method adoption" and "interpretation boundaries" — it must never be written as a scientific guarantee.

AI Ops and public evidence

data/public-claim-registry.json, data/public-evidence-manifest.json, data/trust-pages-machine-audit.json, and deployment records prove what the site currently discloses publicly, which gates it passes, and which states remain a warning or degraded.

AWR / TM multilingual governance

AWR scans articles, pages, CSS, and untranslated residue; TM stores translation memory and terminology consistency. Both only produce routing and evidence — they never rewrite content directly. When the scanner and translation memory conflict, an AI agent decides based on the zh-TW source text, target-language fluency, and evidence severity.

Front-end and release surfaces

Front-end true solar time, region selection, font loading, footer consistency, schema.org, llms.txt, OKF, and sitemap are all release surfaces that gates can verify. Any copy change must avoid breaking indexing, accessibility, or cross-language hreflang.

Research publishing tools

data/research-topics.json is the current queryable research topic pool; the legacy publisher lives at scripts/archive/publish-research.js and is an archived candidate tool. Before it can be re-enabled it must be rewritten for the current routes (/research/zh-TW and /research/posts/...), the current sitemap/llms/OKF pipeline, and AI Ops result reporting — it must not be treated as an active schedule as-is.

AWR / TM Decision Matrix

True Solar Time Research Notes

Chart-time research currently displays "original birth time" and "true solar time" in separate fields: the original time preserves the user's input, and true solar time shows only the region and the corrected time, e.g. "Taichung 12:29." The page does not additionally display the correction in minutes, to avoid users misreading an engineering correction value as a chart conclusion.

Region selection derives its default range from the site language: Traditional Chinese shows Taiwan-region options, Japanese shows Japan-region options, Korean shows Korea-region options, Thai shows Thailand-region options, Indonesian shows Indonesia-region options, and Malay shows Malaysia, Singapore, and related regions; English and Spanish are primarily country selection, using the capital or a major city as the true-solar-time reference.

The research page does not claim that "every region is always correct." Accuracy depends on the front-end options, latitude/longitude data, time-zone data, true-solar-time calculation tests, and post-deploy smoke tests all being maintained together; where regions such as Malaysia, Taiwan, Luoyang, Beijing, or others have time-zone history or longitude discrepancies, this must be documented with test cases.

Claims We Do Not Currently Make

• We do not claim that astrological analysis has medical, legal, financial, or scientific predictive validity; content on this site is for self-exploration and learning reference only.
• We do not claim that AWR or TM can directly judge whether an article is good or bad; they can only flag suspicious spots and produce machine-readable evidence.
• We do not claim that OpenAI, Anthropic, or Google Gemini are all used for production inference; without code or configuration evidence, this can only be written as governance benchmarking.
• We do not claim that a multilingual translation is automatically natural just because it matches the zh-TW source word-for-word; the target language must still pass tone, terminology, cultural-fluency, and schema quality checks.

Research Queue and Ongoing Maintenance

The research article queue is managed by data/research-topics.json. The monthly publishing tool is currently still an archived candidate and must not be treated as an active automated schedule; before it is re-enabled, research articles should be drafted by an AI agent from the queue, reviewed by a human or an agent, passed through the gates, and then published. Priority topics fall into three categories: true-solar-time and region data, Zi Wei Dou Shu hard rules and school differences, and multilingual article-quality governance. After every publication, status must be written back, and AWR/TM must rescan the corresponding pages.

• Short term: Fill in true-solar-time region test cases, especially for Taiwan, Malaysia, Japan, Korea, Indonesia, Thailand, and Singapore.
• Medium term: Convert the AWR/TM multilingual decision matrix into machine-readable issue-routing rules, so full retranslation and partial patches don't get mixed together.
• Publishing tool: To restore publish-research, it must first be rebuilt as a current-generation script: it must not write the old research.html, must not append directly to sitemap/llms, and must go through the current public-discovery pipeline and structured reporting.
• Long term: Bring the research page, system card, data governance, transparency report, FAQ, and article-publishing gate into a single AI Ops state machine.