Retrieval · evidence policy · safe failure
- Role
- Independent prototype author
Public applied-AI prototype · 2026
Evidence decides
whether the system answers.
I built a bilingual retrieval system that cites, checks, repairs, or escalates instead of polishing a guess.
- Route
- Retrieve
- Govern
- Answer
- Escalate
- Evaluate
Governed source trail
Swipe through the evidence03 / 06
- Prototype routing
Sensitive, financial, vague, or malformed requests stop before an answer is drafted.
NORMALIZELanguage + termsGUARDSSensitive / financial / vagueREVIEWReason retained - Hosted embedding dependency
Vector and lexical search merge without hiding where a source ranked.
VECTORSemantic rankBM25Lexical rankTRACELane + rank metadata - Bounded conflict families
Authority, dates, and bounded contradictions decide whether retrieved evidence may be used.
POLICYPreferred + authoritativeDATESNewer source can governCONFLICTExplicit HR receipt - Google model APIs
The system validates a cited draft, restores the requested language, then checks citations again.
DRAFT[SOURCE: path] tokensFR / ENLanguage restoredAUDITRepair or review - Prototype review state
A failed draft can be corrected twice; the third failed validation exits to review.
REJECTGuard or audit failureRETRYFresh evidence · boundedHUMANReason + sources - Curated prototype set
Three curated suites exercise routing, citations, escalation, and a faithfulness proxy.
15Base15Paraphrased06Reasoning
System viewSource policy decides whether evidence may be usedPrecedence is not conflict: authority and date rules resolve some differences; bounded contradictions escalate.