Agents assert things.
We make them provable.
Agentic systems fail in three predictable ways: they assert things that are wrong, they decide without structure, and they run on operations nobody can govern. Dabslabs builds a tool for each failure — and all three run in production today.
Counts and figures read from the live production cluster and recorded bake-offs. The routing figure is the typed-decision layer against the incumbent on a 232-state rerun; every number on this page traces to a measurement in the company's evidence ledger.
The problem
Every team deploying agents hits the same three walls: outputs nobody can verify, decisions nobody can trust, and operations nobody can prove. The industry's answer has been to add more guards to read and more logs to ignore. Dabslabs' answer is that trust is a layer — the same way TLS is a layer — and it is built, not bolted on.
What it does
One thesis, three tools: agentic systems earn trust the way people do, by showing their work. The provenance engine makes assertions checkable. The typed-decision layer makes decisions machine-readable and bounded. The governed estate makes the operations provable. Each is deployed and measured; together they are the trust layer an agentic system stands on.
The three pillars
The Provenance Engine
Agent claims are gathered and validated against ground-truth data with natural-language inference. Every claim carries its source; what enters the graph is what survived checking, with a confidence score that means something.
The Typed-Decision Layer
A GPU-served joint-head model scores every allowed option in one forward pass and returns the answer with a probability attached to each candidate. Out-of-range answers are impossible by construction.
The Governed Estate
The company's own cluster is delivered by GitOps, governed by recorded due-diligence on every consequential change, and operated by its own agents — the discipline an earlier system, Corvus, proved: signed events that scripts verified before acting, authorization checked against the action itself rather than asserted in prose.
FLAGSHIP DEPLOYMENT
Invisible Strings
A music-lore platform where fans of any artist explore the influences, histories, and relationships woven through songs — every connection carrying its source and a confidence score. Audio is transcribed by speech recognition on an NVIDIA Jetson edge device; claims are verified by natural-language inference on RTX-class GPUs. Taylor Swift's discography is the founding corpus and the muse; the platform is artist-agnostic by construction, built for every artist after her.
Who it is for
- Teams deploying agentic systems who need outputs their reviewers can actually verify.
- Operator estates that want every consequential change carrying a recorded receipt.
- Communities and researchers who need the story behind the data to come with its sources — the flagship deployment exists to prove the depth.
Where this actually stands
This is an early-stage software company: pre-revenue, one operator, no employees, and no claims of traction. The tools run in production on the operator's own cluster — sourcing, GPU transcription, claim verification, decision serving, graph database, and web frontend, all deployed and monitored. The discovery layer is designed and is the next build. The platform is not yet publicly launched. That is the honest position, stated here rather than smoothed over.
Boundaries
Provenance
The engineering method behind this company is documented in public, in the founder's own voice: measurements, failures, and the reasoning that produced them.