The proof layer LLMs are missing
LLMs solved the language-translation half of knowledge interop. The verification half is largely unbuilt: provenance, claim attribution, signed assertions, grounding evals, shared verification standards.
The proof layer LLMs are missing
Engelbart in 2004: knowledge bases across organisations should not have to differ, and you still have to communicate and prove things. Two requirements, often conflated.
LLMs solve half of this. The half that’s still unbuilt is where the value actually compounds.
The communication half (mostly solved)
A model can read company A’s knowledge base and answer questions in company B’s vocabulary without a shared schema. Statistical translation dissolved much of the ontology problem that ate research for thirty years. Knowledge bases no longer need to share format to interoperate. The model bridges.
The proof half (largely unsolved)
To “prove” a claim from a knowledge base you need:
- Claim-level attribution. Which source supports this sentence, not which document.
- Source authority. Who wrote it, when, with what credentials, under what review.
- Audit trails. What did the model see, what did it select, what did it generate?
- Reproducibility. Can the same claim be reconstructed by another reader using the same sources?
- Signed assertions. Cryptographic provenance so a claim can travel across orgs without losing its credentials.
- Grounding evals. Tests that measure how often a model’s output is supported by the cited material.
- Shared verification standards. Cross-organisational protocols so a claim verified in company A holds in company B.
Current RAG + citation stacks gesture at some of these. None delivers all of them reliably.
What Engelbart already had
NLS in 1968: every piece of content was author-signed, date-stamped, and structurally linked to its provenance. The Journal was an addressable archive of every interaction, every memo, every decision, with full attribution. Not a research prototype. The actual working environment.
That infrastructure is largely missing from today’s LLM stack. We have the inverse problem of 1968: more powerful language machinery, weaker provenance machinery.
Why this matters
Without a proof layer:
- Cross-organisational knowledge sharing can’t be trusted.
- LLM output can’t be safely used in regulated work (legal, medical, financial, scientific).
- Cultural memory mediated by LLMs accumulates errors silently.
- C-activity at scale becomes impossible because no one can audit what the improvement process actually did.
What to build
Provenance graphs under the model. Claim attribution as a first-class output. Signed assertions as a transport format. Grounding evals as continuous tests. Verification standards as inter-org protocol.
LLMs handle the language. The provenance fabric handles the trust. Both are required. The second one is the one almost nobody is building.