When documents disagree, AHS catches it — and proves it.
Paste two documents below. The Skeptic engine scans for direct contradictions,
numeric disagreements, and version drift; every step is recorded in a
hash-chained audit trail. No API key required — everything runs locally.
Click any situation to load a real example and run it instantly.
Who is AHS-Core?
AHS-Core · Agentic Hyper-graph Synapse
The conflict-resolution & forensic-traceability layer
for agentic RAG. When documents disagree, AHS catches the contradiction, surfaces both
versions with citations, flags it for human review, and writes a tamper-evident
audit record. It's a library you bolt on —
not another framework to migrate to.
MIT · Python 3.10+ · Offline-first
Speculative Retrieval
Parallel fan-out of sub-queries with bounded concurrency and backpressure.
Skeptic Engine
Flags direct contradictions, numeric disagreements, and version drift.
Forensic Log
Hash-chained JSONL — tamper with one line and verify() catches it.
Grounded Answers
LLM is forced to cite evidence IDs; conflicts are surfaced, not hidden.
Pluggable Stack
Swap embedders, retrievers, and LLMs — OpenAI, Anthropic, or local.
Offline-First
Deterministic hash embeddings — demo, test, and CI without API keys.
⚙ Pipeline architecture
Questionuser input
→
Retrieverparallel hops
→
Skepticconflict scan
→
LLMcited answer
↓ every step sealed to ↓
Forensic Loghash-chained JSONL
0
Tests
all green
0
Python versions
3.10 · 11 · 12
0
Capability modules
retrieval → audit
0
API keys required
fully offline
What it deliberately does NOT do in v0.1 — ship clinical/legal
agents (no domain validation), claim "100% traceability" or "O(1) latency" without
numbers, or replace your orchestrator. It's a layer you add, not a migration.