TRANSPARENCY / BUILD LOG

What is actually implemented

This page is the architecture disclosure. “Implemented” means deployed code and persistent runtime state—not a roadmap label.

Official-source ingestion

Implemented

OpenAI RSS, Anthropic Research, DeepMind News, Hugging Face Blog, and arXiv cs.AI/cs.CL.

RAG

Implemented

Overlapping chunks, 256D deterministic embeddings, D1 vector persistence, hybrid vector/keyword retrieval, filters, reranking, citations, and abstention.

Multi-agent research

Implemented

Versioned orchestrator registry with Healthcare, Financial Services, Retail, Manufacturing, and Legal agents.

Verification gate

Implemented

Exact evidence-span and citation binding checks; failed claims cannot reach verified status.

Human approval

Implemented

Owner-authenticated review console; only verified drafts can be published.

Daily execution

Implemented

A daily 6:00 AM America/New_York automation triggers ingestion; a date-aware first-request fallback prevents missed daily runs.

External LLM

Not enabled

No Anthropic/OpenAI generation API is claimed. Current agent outputs are deterministic and source-bound.

Supabase / pgvector

Not used

The deployed platform uses Cloudflare D1 and its own stored vectors; this avoids an unnecessary second database.

Publication state machine

Ingested evidence → draft digest → deterministic verification → owner review → published. There is no route from draft directly to public.

Security model

Public routes are read-only. Administrative APIs require the authenticated owner identity. Source documents, vectors, pipeline runs, verification logs, and review actions are stored in D1.