Zero-Hallucination Evidence
Every RDTII mapping carries byte- and bbox-level citations to the source PDF. Agent C runs a red-team reverse-grounding pass before anything reaches a human.
Atman-RDTII is a self-architected, multi-agent legal AI audit system. It reads scanned government gazettes, maps every clause to the UN RDTII framework with byte-level evidence, and lets experts co-pilot the verdict in real time.
Built from scratch on Actor Model + Pub/Sub. Property-based tests pin down every promise — invariants P-1 through P-15 are enforced both in code and at the database layer.
Every RDTII mapping carries byte- and bbox-level citations to the source PDF. Agent C runs a red-team reverse-grounding pass before anything reaches a human.
No LangChain, no AutoGen. A 33-step queryLoop drives Pub/Sub agents across Cloudflare Queues + Durable Objects + D1 + Vectorize.
queryGuard + commandQueue mirror the Claude-Code mid-layer loop. Slash commands and natural-language instructions arbitrate cleanly.
Every expert re-classification produces a <Prompt, Chosen, Rejected> tuple. Vectorize fewshot-live updates in <1 s — no fine-tuning required.
Cloudflare Pages, Workers, D1, KV, R2, Queues, Vectorize, Durable Objects — Singapore POP serves ASEAN+ regulators sub-100 ms.
15 invariants P-1..P-15 verified by fast-check across 51 test files / 223 tests. Every claim above is mechanically checked.
Macro layer: a Pub/Sub event bus drives Agent A (ingest + OCR), Agent B (RDTII mapping via ReAct + Few-Shot), Agent C (red-team reverse-grounding), Agent D (Co-Pilot synthesis). Mid layer: queryGuard + commandQueue arbitrate human input. Micro layer: a 33-step queryLoop powers each agent with streaming tool execution, autocompact, and unkillable error recovery.
Read the full design →Pub/Sub Event Bus · Agents A·B·C·D
Co-Pilot REPL · queryGuard · commandQueue
33-Step queryLoop · streaming · autocompact