Verity

A research agent that refuses to publish until every important claim can be traced, verified, scored, and attested onchain.

Track 2 submission for the Mantle Research Challenge. Verity is a verification-first onchain-finance research agent: it gathers data (Dune) + context (Exa), synthesizes claims that each carry provenance, then hard-gates the output through a deterministic checker before rendering a PDF and anchoring its hash to ERC-8004 on Mantle.

Why it's different

The field is full of "AI research agents" competing on autonomy and speed. Verity competes on trust. Every number must equal its source value (checked by code, never by an LLM), every address must be on a hand-verified allowlist, and nothing un-verified is ever published or attested.

Where trust actually comes from: the public, re-runnable Dune query IDs + the recomputable PDF hash. The on-chain attestation is a timestamp/anchor — not the source of trust.

Pipeline

        User question / claim
                  |
                  v
        Operator (plan + cost estimate)
                  |
                  v
        Scouts  -- onchain-finance-scout (allowlist -> Dune)
                +- web-scout (Exa / Firecrawl)
                  |   (summaries + query IDs + sources)
                  v
        Synthesizer  -> draft claims, each with provenance + confidence
                  |
                  v
   +------ Verification Gate (HARD) ------+
   |  (a) Deterministic provenance checker |  <- the spine
   |  (b) Configurable LLM-as-judge (qual.)|
   +---------------+-----------------------+
            pass    |    fail -> back to synth/scout
                    v
        Report builder -> PDF (charts, sources, confidence, cost)
                  |
                  v
        Attestation -> recompute hash -> ERC-8004 Validation Registry (Mantle)

Run it

1. npm install && npx playwright install chromium
2. Copy .env.example → .env, fill keys (Dune, Exa, Anthropic, Mantle).
3. Offline demo (no keys needed): npx tsx src/cli.ts --fixture → writes examples/mantle-rwa-q2-2026.pdf.
4. Register the agent (once): npx tsx src/attest-8004/register.ts → mints the ERC-8004 agentId on Mantle and prints VERITY_AGENT_ID + VERITY_VALIDATOR_ADDRESS to paste into .env.
5. Live: set VERITY_QUERY_IDS=<comma,separated,dune,ids>, then npx tsx src/cli.ts "your question".

Model providers

The synthesizer and judge are provider-agnostic. Default is Anthropic; to test with OpenAI set VERITY_LLM_PROVIDER=openai, OPENAI_API_KEY, and point VERITY_SYNTH_MODEL / VERITY_JUDGE_MODEL at OpenAI model names. To exercise the real LLM path cheaply over cached data (no Dune/Exa keys): VERITY_FIXTURE_LIVE_LLM=1 npx tsx src/cli.ts --fixture.

How it's built

TypeScript + vitest, fully TDD. Spine = src/verify/provenance-checker.ts (deterministic). See docs/superpowers/ for the full design + plan. Packaged as a Mantle Agent Skill (SKILL.md).

Live example

The checked-in examples/mantle-rwa-q2-2026.pdf answers "Did Mantle's RWA growth actually accelerate in Q2 2026, and are tokenized-equity adoption claims supported onchain?" — with confidence scores, re-runnable Dune sources, a cost panel, and an attestation anchor.

Status / roadmap

v1 uses a hand-verified address allowlist and a Mantle attestation anchor (confirm the ValidationRegistry address on mantlescan.xyz before live use). Roadmap: dynamic verified resolution, per-run + reputation attestation, autonomous loop/goal, multi-model worker fan-out, real x402 pay-to-run.

Submission checklist (manual)

• Push repo public
• Publish X thread tagging @Mantle_Official; follow + like/share the campaign article
• Join the Mantle creators Discord
• Submit the participation form with the correct wallet address