Evidence Layer · Versioned Quality
Methodology & Test Evidence
Reproducibility rules, fixed-seed quality gates, crawl tests, browser tests, limitations, and release controls for Atlas of Names.
Product owner and release accountability: Will Hu · Implementation and automated evidence: Manus AI · Updated · Independent human naming-quality reviewer pending
This MVP implements a deterministic TypeScript engine in the browser. The same engine version, ruleset version, seed, semantic cues, and configuration produce the same ordered output. Base generation makes no AI request.
Automated quality contract
| Area | Automated check | Acceptance |
|---|---|---|
| Determinism | Repeat identical configuration and seed | 100% ordered equality |
| Task morphology | 500 deterministic fixtures for each of five task types | Task fields complete; final strings readable |
| Semantic mapping | Counterfactual briefs with identical seed | Mapped output fields and traces change |
| Batch uniqueness | Generate 12 candidates across fixed seeds | No exact or high-similarity pair |
| Readability | Final length, hostile clusters, vowel runs, pronunciation | Zero accepted failures |
| Safety | Prohibited, retained-failure, and recognizable fictional patterns | Filtered from recommendations |
| Inputs | RTL, CJK, emoji, combining marks, controls, markup, conflicting strict rules | Bounded, normalized, non-injecting, explicit error or recovery |
| Board | Legacy migration and malformed imports | Existing state preserved on error |
| Crawl | Approved routes, redirects, unknown URLs, metadata, canonicals, sitemap | 200/301/404 contract with zero duplicate metadata |
| Accessibility | axe critical/serious scan | Zero violations |
| Privacy | Inspect event payloads and generation requests | No brief, result, seed, Board, IP, UA, or session identifier stored in analytics |
Versioned evidence
The fixed-fixture run writes downloadable JSON and CSV with versions, per-task fixture counts, successful samples, retained failure examples, counterfactual evidence, and independent scoring reasons. A passing automated report is never described as human approval.
Release gates
| EMD acquired | true |
|---|---|
| SERP shape evidence | true |
| External liquidity observation | false — post-release evidence, not fabricated |
| Technical/product MVP gate | true |
| English public indexing allowed | true |
| Chinese pages | 17 remain noindex pending independent language review |
Known limits
Similarity is an editorial signal rather than trademark, copyright, linguistic, cultural, or identity clearance. Pronunciation is an English-readable approximation rather than IPA. The site does not claim cultural authenticity, legal availability, uniqueness, ranking performance, or human review that has not occurred.
Report and revision
Use the persistent contact and correction form to submit the route, generated result, seed, engine version, observed problem, and requested correction. It returns a public reference and status URL. Will Hu reviews correction and rights reports within five business days.