Wever Labs Commerce Proof

Inspiration

Agent work often ends inside one conversation or API response. The next person or system cannot easily tell what was requested, what the agent did, which evidence supports the result, or whether the work is ready to continue. We built Wever Labs Commerce Proof so human intent, agent execution, reviewed evidence, and a verifiable result travel together.

What it does

Wever Labs Commerce Proof turns a human-directed agent run into a result the next person, agent, or application can use.

The customer defines a BUIDL diligence job, including the agent, tools, time, budget, and expected result. The Atlas Diligence Agent returns a structured Tokenization Atlas research package covering the asset's legal claim, backing, custody, investor eligibility, redemption, and transfer path.

GPT-5.6 maps each claim to its sources, identifies incomplete or conflicting support, and summarizes the evidence in structured output. Wever Labs then scores completion, evidence, scope, authorization, and handoff readiness. The complete request, execution record, GPT review, scores, and next step are packaged into a signed Agent Commerce Passport with an independently verifiable receipt.

The hosted demo also includes scenarios for missing evidence, conflicting evidence, and an unavailable AI review. Each produces a clear handoff that tells a person what needs attention.

How we built it

The browser experience runs on Netlify with a server-side JavaScript workflow. Tokenization Atlas supplies the BUIDL research package. GPT-5.6 is called through the OpenAI Responses API with strict structured output to map claims to evidence, find missing or conflicting support, and produce a reusable summary.

Deterministic Wever Labs code handles short-lived job authorization, the five quality scores, the next-step decision, Ed25519 signing, and receipt verification. The hosted judge path requires no login, API key, payment, or production write.

Codex helped audit the existing Wever Labs modules and build the guided interface, isolated server workflow, GPT-5.6 evidence-review boundary, adversarial tests, submission documentation, and desktop and mobile verification.

Challenges we ran into

The main challenge was making an AI-reviewed result useful without making the model the authority. We separated evidence interpretation from authorization, scoring, signing, and final workflow state. We also designed a realistic end-to-end demo that judges can run without credentials or setup.

Accomplishments that we're proud of

One browser run carries the customer's request through agent execution, GPT-5.6 evidence review, five consistent quality scores, a clear next step, a signed Agent Commerce Passport, and independent verification.

The same authorization cannot start a second job. Alternate scenarios turn incomplete evidence, contradictions, and model unavailability into understandable handoffs instead of unexplained failures.

What we learned

Reusable agent work needs more than a good answer. The request, execution context, evidence, review, quality state, and verification have to remain connected. The product becomes much easier to trust when people can understand both a successful result and what needs attention.

What's next

Next we will add configurable research templates, live source refresh, portable verification libraries, and cross-agent passport exchange so more products and agents can carry completed work into the next workflow.

Built With

• api
• codex
• cryptography
• gpt-5.6
• javascript
• mcp
• netlify
• node.js
• openai