I have released four short papers on one idea: making AI-assisted work re-checkable.
- EMET — a byte-level integrity witness whose verdict {MATCH, DRIFT, UNVERIFIABLE} cannot, by construction, express "trusted." Four independent implementations, 44 conformance vectors.
- BuildLang — a compiler that puts ambient capabilities in the function type and seals re-derivable receipts verified by re-execution.
- Witnessed Independence — a mechanism that records whether a verifier graded its own work, and refuses to decide when independence is not witnessed.
- Proof Packets — an envelope for one agent action whose verdict is derived from checks, so a claim can never vouch for itself.
Source and tests for all four are public.
Top comments (1)
Great discussion. The idea of making AI-assisted work re-checkable is one of the most important shifts needed as AI becomes part of real engineering and knowledge workflows.
A lot of current AI systems optimize for generating answers, but production environments require something different:
Can we trace where the output came from?
Can we reproduce the reasoning process?
Can we verify assumptions and intermediate steps?
Can another person or system audit the result later?
In AI engineering, I’ve found that reliability comes less from trying to eliminate model uncertainty and more from designing systems around it:
Capture source context and retrieval paths
Store prompts, model versions, and configuration changes
Keep structured intermediate outputs
Add evaluation and validation steps before final delivery
This is especially important for agentic systems. Once AI moves from answering questions to taking actions, explainability and traceability become operational requirements, not optional features.
The future of AI-assisted work will not just be about generating faster — it will be about creating workflows where humans and AI can collaborate with confidence.
Great perspective. Re-checkability may become one of the defining principles of trustworthy AI systems. 👏