← All work 13 / Open source

RDLLM

I built a reference implementation that attaches evidence to AI claims and checks rights before producing allocation instructions.

v1.0.0 · Beta reference

Role: Author & maintainer
Duration: 2026
Team: Independent project
Status: v1.0.0 · Beta reference

Which passage supports this sentence?

A list of sources can tell me what an AI response cites. If I want to check one sentence, I still need to find the passage behind it. I built Royalty Driven LLM to carry that connection with the answer when it’s copied, exported or passed through an API.

The reference implementation can ground a response before generation, audit existing text or produce allocation instructions. It records evidence spans and hashes and uses Ed25519 signatures so others can verify the receipts. I kept those interfaces provider-neutral so the records can travel across model vendors.

Evidence and rights get separate checks

A passage might support a claim while its rights prohibit a proposed use. I made those two separate decisions. Claims with weak support, conflicting evidence or rights restrictions can go to review or escrow. The source includes an entailment channel, licence parsers, and Shapley and Banzhaf allocation options.

The MIT-licensed v1.0.0 beta has a 187-case internal conformance suite and a 17-page white paper. Its output stops at allocation instructions. Payment execution and hidden training-data detection are outside its scope. It remains a reference implementation, with no production certification or PyPI release.

Explore the repository and white paper.

The record

What the evidence says.

Internal conformance cases: 187; Project conformance suite
White paper pages: 17; Published project paper

Royalty Driven LLM / Workflow

What must pass before RDLLM allocates?

Ground or audit
Ground a response before generation or audit existing text.
Link each claim to evidence
Attach the supporting source span and its hash to the claim.
Check evidence and rights
Assess whether the span supports the claim and whether the rights allow allocation.

Two possible evidence and rights outcomes

Choose an outcome to read why the next action changes. This illustrates the logic; it does not run a verifier.

Pass · Generate instructions Produce allocation instructions after both checks pass. Fail · Hold allocation Route unsupported, disputed or rights-blocked claims to review or escrow.

Pass · Generate instructionsBoth evidence support and rights checks pass. The output is an allocation instruction, not a payment.

Fail · Hold allocationEither insufficient support or a rights restriction is enough to hold allocation. A supporting source alone cannot override a rights restriction.

Worked explanationFollow a supported but restricted claim

Input: A source passage supports a claim, but its rights restrict the proposed use.
Decision / rule: Evaluate support and rights separately. Evidence support alone does not clear the allocation.
Outcome: Hold allocation and route the claim to review or escrow rather than issue cleared allocation instructions.

Scope of this exampleAn illustrative beta workflow. Signed receipts verify record integrity; they do not settle disputed rights or execute payments.

Source footers travel with the output. Ed25519-signed receipts make records publicly verifiable. This beta reference implementation produces allocation instructions; it does not execute payments.

What I’d do next.

Validate provider integrations and rights-handling assumptions beyond the internal conformance suite.

Working through a similar decision?

Talk to me about your project

Next project / 14 / Research KAPREKAR I found a smaller exact description of a digit routine, then used it to pose a separate, testable question about model uncertainty.