Every claim on this site is backed by an artifact: real numbers from real work orders, with reproducible commands.
One work order’s terminal gate, fully navigable: every review round in true order, every refusal on the way to green, the false-success catch, the fix chain, and the witness record that closed it. Reconstructed from real records, with a guided tour that explains every screen.
The entire estate rendered from the system of record: governed work wrapped around the grounding core of code and memory. Interactive, replayable, anonymized.
Validated against the production runtime. The benchmark intentionally exceeded expected production behavior by driving individual retrieval surfaces at full concurrency.
Measured on production workloads, not synthetic benchmarks. 22× stock TensorRT-LLM throughput.
State survives the worker. The resume packet is regenerated from PostgreSQL on every stand-down.
Gates don't pass on assertions. The work graph is the record; every claim is backed by an artifact on disk.
Before any code change ships, the graph tells you the exact blast radius. Sub-second, no LLM.
Ground decisions in code, not model recall. Every hit returns file path, line range, and score.
Same PostgreSQL state regenerates the same packet bytes. Determinism is a property of the bytes.
Cost is a field on every task, not a line on an invoice. Sub-penny precision per call.
Every call routed by required task_type and schema_id. No ungoverned dispatches.
These eleven capabilities are the foundation. More evidence entries land as the system grows. If you want to put your own work through AgentOS, get started below.