About

How Atman-RDTII works.

A from-scratch, multi-agent, zero-hallucination legal-AI audit system. Every claim below is checked by property-based tests.

Three nested loops

Macro · Pub/Sub bus. Four agents subscribe to AgentEvent kinds. Agent A discovers + parses gazettes, Agent B maps to RDTII via ReAct + Few-Shot, Agent C runs a red-team reverse-grounding pass, Agent D synthesizes the report and arbitrates expert instructions.
Mid · Co-Pilot REPL. Mirrors the Claude-Code call chain: PromptInput → handlePromptSubmit → executeUserInput → onQuery → query() → queryLoop(). queryGuard ensures no two engines ever run concurrently (P-3).
Micro · 33-step queryLoop. Streaming tool execution, autocompact / microcompact / contextCollapse compaction chain, unkillable error recovery with bounded retries (P-14), strict tool_use ↔ tool_result pairing (P-4), guaranteed termination (P-5).

Front-end: Next.js 14 · TypeScript 5 strict · Tailwind · Framer Motion · pdf.js · D3 force
Edge runtime: Cloudflare Pages · Workers · Pages Functions WebSocket
Storage: D1 (CHECK-enforced P-1/P-8/P-13) · KV · R2
Async: Cloudflare Queues + DLQ · Durable Objects
Vectors: Vectorize × 3 (taxonomy · clauses · fewshot-live)
LLM: DeepSeek Chat (streaming SSE) via self-architected gateway
Tests: vitest + fast-check · 51 files / 223 tests · 15/15 properties

Each row is enforced by at least one fast-check property test. CI fails the PR if any invariant is violated.