Project summary

ValiChord is open-source infrastructure that makes independent verification structurally tamper-proof. Independent validators — AI agents now, humans later — reproduce a claim blind, seal cryptographic commitments to their verdicts, and reveal simultaneously; a tampered reveal is rejected on-chain. The result is a permanent, publicly queryable Harmony Record that preserves agreement and disagreement, and that no lab, operator, or developer (including me) can edit after the fact.

AI capability and safety claims increasingly drive deployment and policy decisions, yet rest on self-reported numbers. ValiChord makes eval integrity a structural property rather than a matter of trusting whoever ran the test.

• Live demo, full protocol end-to-end in ~90s: https://valichord-demo.onrender.com/demo

• Repo (Apache-2.0): https://github.com/ValiChord/ValiChord

• Governance framework (DOI): https://doi.org/10.5281/zenodo.18878108

What are this project's goals? How will you achieve them?

Goal: turn ValiChord from a working demo into infrastructure the AI-evals world actually uses. Over the funded period I will:

• Land the lm-evaluation-harness integration — code ready to run and submitted upstream — so benchmark results can be sealed into a tamper-proof record.

• Run AI-validator rounds at volume on real published claims, producing 50+ Harmony Records that show independent validators contesting results blind.

• Harden and independently audit the commit-reveal implementation (the component everything else depends on).

• Decentralise operation — bring on the first node operators who aren't me, so the network stops depending on a single party.

How will this funding be used?

Full ask $63,000 / 12 months; minimum viable $10,000 / 3 months (integration + first batch of records).

• ~$53,000 — stipend, sole full-time developer (subsistence, not salary)

• ~$6,200 — infrastructure + AI-validator model API costs

• ~$3,800 — independent security audit of the commit-reveal protocol

There is no organisation and no overhead; funding goes to one person and the machines the network runs on.

Who is on your team? What's your track record on similar projects?

Solo. I'm Ceri John, a UK-based builder — Music teacher and award-winning documentary filmmaker — who taught himself distributed systems and cryptography and, over the past six months, designed and shipped the entire ValiChord stack alone and unfunded:

• 4-DNA Holochain application (Rust) with on-chain commit-reveal verification

• Svelte/TypeScript UI for researcher, validator, and governance roles

• Python attestation library — 537 tests, 97% coverage — with adapters for Inspect AI, Inspect Evals, lm-evaluation-harness, and AILuminate

• 183 integration tests; permanently live five-node deployment; v0.6.0 released

• Published governance framework designed to resist institutional capture

What are the most likely causes and outcomes if this project fails?

• Adoption risk (most likely): the protocol works but no eval org or journal adopts it, so records stay demonstrative rather than load-bearing. Outcome: a proven mechanism looking for a home — still a public good, and re-usable when the "why should we trust this score?" pressure grows.

• Validator-independence risk: AI validators on similar base models may be correlated rather than truly independent; at scale, collusion/sybil resistance needs work. Outcome: informative negative result about the limits of automated validation.

• Funding runout: without support I return to teaching and the project drops to evenings/weekends — the code survives, but the live network stalls.

Even in the failure cases, the honest documentation of what did and didn't work is a contribution.

How much money have you raised in the last 12 months, and from where?

$0. Entirely self-funded to date. One application under assessment (Long-Term Future Fund) and several submitted or in progress (grantmaking.ai, BlueDot, Cosmos Institute, Emergent Ventures). No prior grants, no institutional backing.