June 10, 2026 · 5 min read

Judging evidence with LLM oracles on GenLayer

Proven — the dispute-resolution protocol my team started at the Aleph & GenLayer Hackathon — rests on an unusual primitive: smart contracts that ask LLMs to evaluate evidence. Here's what building on non-deterministic oracles actually feels like.

Determinism was the whole point. Now what?

Classic smart contracts are valuable because they're deterministic: same input, same output, every node agrees. Subjective-logic contracts on GenLayer break that on purpose — validators run LLM inference over the evidence each party submits, and the network reaches consensus over judgments, not hashes.

That trade unlocks a category of problems blockchains couldn't touch (was this delivery late? is this photo consistent with the claim?), but it imports every operational reality of LLMs into your protocol design:

• Latency is consensus-bound. Your slowest validator's inference time is your floor. UI patterns that work for chat (streaming, optimistic rendering) don't apply when the answer is a verdict.
• Cost scales with evidence size. Every page of submitted evidence is tokens each validator pays to process. We learned to treat evidence like a legal brief: structured, bounded, with the burden on the submitter to be concise.
• Disagreement is a feature. Validators won't always agree — that's the subjective part. The protocol has to define what happens at the margins, exactly like appellate review exists because trial courts get things wrong.

Where the law background pays off

Designing the evidence flow, I kept reaching for procedural law concepts: admissibility (what evidence the contract accepts), burden of proof (who must submit what), and standard of review (how much validators can weigh context). These aren't metaphors — they're requirements. A dispute protocol without them produces verdicts nobody trusts.

The interesting discovery of the hackathon was that writing the prompt for a validator looks a lot like drafting jury instructions: you're specifying, in plain language, how a reasoner should weigh evidence within bounds.

What's next

We're still building Proven past the hackathon — the current focus is making verdicts reproducible enough to audit without making them trivially gameable. If you're working on subjective consensus or computational dispute resolution, get in touch.