Agents act now. They read codebases, run commands, move money, call APIs. When something goes wrong, nobody can answer the basic question: what did the agent actually do, and who told it to?
JUWEL OS is an AI operating system in the browser built on one rule: every consequential agent action leaves a signed, verifiable receipt. Each receipt is a cryptographic artifact (Ed25519 signature, chained to its predecessor) that records what happened, what authorized it, and what the inputs and outputs were. Anyone can verify a receipt independently — no trusted server, no vendor, no blockchain needed. JUWEL OS 1.0 launched September 4, 2026. A receipts-as-a-service API (issue/verify, API-key auth, per-key metering) is built and running.
This proposal funds turning that mechanism into public verification infrastructure: an open spec, open tooling, a public benchmark, and independent red-team evaluation — so agent accountability stops being a marketing claim and starts being checkable by anyone.
The problem: agentic AI is already executing in production (coding agents, ops agents, customer-facing agents) with no standard way to prove what an agent did. Incident response, audits, and liability all assume a trustworthy record exists. It usually doesn't — logs can be edited, edited logs can be lost, and vendor dashboards require trusting the vendor. Verification must be for everyone, not a badge a lab hands out.
The solution: signed action receipts as an open, checkable standard. When an agent takes a consequential action, the system produces a receipt containing the action description, inputs, outputs, authorization context, a timestamp, and a hash-chain link to the previous receipt, then signs the whole thing with Ed25519. A third party can verify the signature, confirm the chain is unbroken, and confirm nothing was altered after the fact — without asking us for anything.
Milestones:
1. Benchmark MVP (Tranche 1): publish a public benchmark for agent action receipts — an open-source verifier library, a test corpus of receipts (valid, tampered, replayed, chain-broken), and a public report documenting which tampering classes the format resists and which it doesn't. Results published as a technical report on Zenodo.
2. Standard deployment (Tranche 2): stable public spec (v1), a hosted verification endpoint anyone can query, integration examples for common agent frameworks, and an independent red-team evaluation by outside reviewers, published in full — passes and failures alike.
3. Full-time (Tranche 3): ship the receipt system as maintained public infrastructure — conformance test suite, a second independent security review, a reference audit-trail integration for real agent deployments. Twelve months of full-time work, reported publicly every quarter.
Tranched so one regrantor can make the first move. Each tranche unlocks at its own funding bar and pays for a defined, public deliverable.
Tranche 1 — $10,000: Benchmark MVP (~6 weeks). Open-source verifier library + test corpus published under a permissive license. Public demo page where anyone can paste a receipt and see verification succeed or fail, with tampering examples. Technical report on Zenodo. Done = repo public, demo live, report published, linkable.
Tranche 2 — $30,000: Standard (cumulative; ~4 months total). Everything in Tranche 1, plus: stable public spec v1 with versioning and a conformance test suite; hosted verification endpoint (free, rate-limited) anyone can query without an account; integration examples for common agent frameworks; independent red-team evaluation, published in full — passes and failures. Done = spec v1 frozen, endpoint live, red-team report public.
Tranche 3 — $75,000: Full-time (cumulative; 12 months). Everything in Tranche 2, plus: founder full-time for 12 months on receipts infrastructure — maintenance, spec evolution, second independent security review, reference audit-trail integration for real agent deployments; quarterly public progress reports; all code and docs open. Done = 12 months of maintained public infra, second audit published, reference integration running in a real deployment.
Budget note: solo founder, pre-revenue company. Money goes to founder living costs (so she can work on this full-time at Tranche 3), cloud hosting for the verification endpoint and demo, contractor hours for the independent red-team review (Tranche 2+) and second security review (Tranche 3), and publication costs. No blockchain, no tokens, no equity offered.
I'm Annalea Layton, solo founder of Vext Labs Inc (founded March 2026, vextlabs.ai), a Maryland-based independent AI research lab. I work alone — no team, no contractors on core engineering — and I ship.
- Built JUWEL OS end to end: an AI operating system in the browser where every consequential agent action leaves a signed, verifiable receipt. Version 1.0 launched September 4, 2026.
- Built and shipped a receipts-as-a-service API: issue and verify receipts, API-key auth, per-key metering, Ed25519-signed. Running now.
- Research agenda "Human Intent" (machines understanding what people mean): six papers deposited on Zenodo, written as scoped, verifiable claims with published evidence gates — results that pass stay up, failures stay public.
- Nobody adopts the standard. Mitigation: the benchmark and verifier are useful on day one even for a single deployment; I dogfood it on my own agent fleet and publish the results either way.
- A cryptographic or design flaw the red team finds that can't be fixed cheaply. Mitigation: the failure gets documented, not hidden, and the spec says exactly what it does and doesn't guarantee.
- Scope creep into building an agent platform instead of verification infrastructure. Mitigation: tranches are scoped to receipts and verification only; the spec is the product.
- Solo-founder bandwidth. Mitigation: Tranche 1 is deliberately small (6 weeks, one deliverable set); later tranches fund contractor review hours so I'm not the only eyes on the cryptography.
None awarded to date. Pre-revenue; all work so far self-funded by the founder. Applications submitted elsewhere and pending (no awards confirmed): Amber Grant ($10K), Her Agenda ($25), Emergent Ventures, TEDCO BRIDGE assessment. A $99 founding-member offer for JUWEL OS Pro exists with a live payment link but zero paid checkouts — it is not being promoted until fulfillment ships, and it is not counted as funding. SAM.gov UEI Q826JUJWLM77 assigned Sep 15, 2026 (needed for federal grants, not for Manifund).