ledgerkeep

Ledgerkeep: a guardrailed operations agent as a decentralized AI service

Applicant: doom2quake (builder collective) Programme: ASI Deep Funding (SingularityNET, Artificial Superintelligence Alliance) Programme page: https://deepfunding.ai/ Requested: milestone-based grant, non-dilutive, IP retained by the team Project repo: github.com/doom2quake/ledgerkeep (new repo, purpose-built for this entry) Status of this document: draft grant proposal, testnet-only scope, no mainnet deployment


0. Grant verification (read this first)

We checked the programme against ASI / SingularityNET’s own pages before writing.

VERIFIED on official ASI / SingularityNET pages:

CONFLICT / GATE we are flagging honestly: the official application site (deepfunding.ai) returned HTTP 403 to our fetch, so we could not read the live round state directly. Before submitting, an operator must open deepfunding.ai in a browser and confirm there is an open round or RFP, its deadline, its pool, and its current milestone-payment terms. Treat the per-project amount, the rolling-versus-round cadence, and any token or on-chain disbursement mechanics as operator-verify items, not settled facts. This is the single gating check before we invest applicant time.

INFERRED (our mapping, not a verbatim label): that a guardrailed operations agent published as a callable service maps onto Deep Funding’s decentralized-AI- service thesis and its MCP composition story. The mapping is ours.


1. The problem

A payments configuration is pushed at 14:03. It routes one region’s card traffic to a gateway that declines most authorisations. Order volume still looks normal, because customers keep trying to buy. Settled revenue quietly falls, and nobody notices until the next morning. The team that suffers this is the on-call operator: a person paged at 03:00 who has to reconstruct, from a dashboard that only shows the symptom, which of forty deploys that day moved a number, how much money has already been lost, and what one line to change. That reconstruction is slow, it happens under pressure, and it is where the cost lives. A single mis-routed region can cost six figures per day while it goes unattributed.

The reason this stays unsolved is not detection. Anomaly detection is a solved commodity. The unsolved part is the step after the alert: an agent that can be trusted to investigate, attribute the cause to a specific change, quantify the loss, and propose the exact fix, without being trusted to execute anything a human has not approved. The market is full of demo agents that will happily run generated SQL against production or act on model text that carries an injection. The gap between a demo agent and one an operator would let near their systems is almost entirely guardrails, durable memory, correct routing, and an honest report of what the agent actually did. That is the layer nobody wants to build twice, and it is the layer we have already built and tested.

2. Why the ASI ecosystem, why now

Deep Funding is explicit about funding decentralized AI services: builders who publish agents and tools that others can discover, call, and pay for, while retaining ownership of their work. An operations agent that a team must trust is exactly the kind of component that benefits from being a standalone, independently callable service with a published, auditable contract, rather than a black box buried inside one company’s stack.

Ledgerkeep maps onto that thesis directly. It is packaged as a marketplace service: a stable, typed request and response, a guarantee about what it will and will not do, and an audit object returned on every call so the caller can verify the chain from a number moved to this proposed fix. The funder’s technology is load-bearing behind an adapter seam: when a seed phrase and the uagents SDK are present, the service is a real uAgents endpoint on Agentverse (testnet only); otherwise it answers the identical contract over an in-process transport, keyless. The Model Context Protocol surface we already implement is the interoperability seam: the same agent that answers a marketplace call can be consumed as a tool by another agent in the ASI ecosystem, which is the composition story the Alliance is built around. The timing is that the tooling to publish and compose agent services matured in 2026 to the point where an operator can actually call one, and the guardrail question has moved from academic to urgent as autonomous agents start touching real systems.

3. Evidence we ship

Most Deep Funding applicants arrive with a plan and no artifact. We arrive with working, tested code. Milestone 1 is already built and green in this repo.

Measured facts, reproducible on any machine with no network and no credentials:

The reusable control plane is the vendored agent_core package (guardrails, a domain-aware router, durable run state with recurrence detection, and an MCP bridge). It carries its own passing tests here, so the safety layer is exercised in this repo rather than merely referenced. Ledgerkeep is built on this package, which is why the safety layer arrives on day one rather than being promised for a later milestone.

An independent code-review pass over the changed surface is part of our normal process and will be run and its findings published for each milestone below.

4. Milestone roadmap

Funding is requested against milestones. Each states a concrete deliverable, how a reviewer verifies it independently, and what it unlocks. All Web3 or on-chain interaction is testnet only for the duration of the grant.

Milestone 1: the re-themed service, its contract, and the audit object (weeks 0 to 4). BUILT. Deliverable: the doom2quake/ledgerkeep repository with a documented marketplace service contract (typed request and response) and the audit object returned on every call, with the offline loop and guardrail suite green. Verify: clone the repo, run the suite, confirm 82 tests pass offline with no network; call the service contract locally and inspect the returned audit object. Status: complete and green (see Section 3). Unlocks: a callable, self-contained service that later milestones publish and compose.

Milestone 2: MCP interoperability and composition (weeks 4 to 9). Deliverable: Ledgerkeep exposed over the Model Context Protocol so another agent can call it as a tool, with a worked example of a second agent invoking it and consuming its audit object, plus a conformance test for the MCP surface. Verify: run the provided MCP client example against the running server and observe the tool call, the response, and the recorded guardrail decisions; run the conformance test. Unlocks: the composition story the ASI ecosystem is built around, and the seam other builders integrate against.

Milestone 3: testnet-anchored audit provenance (weeks 9 to 15). Deliverable: an optional, env-gated path that anchors the hash of each run’s audit object to a public testnet, so a caller can verify after the fact that a returned audit trail was not edited. This is a tamper-evidence feature, not a custody feature; it degrades to a labelled no-op when unconfigured. Verify: run a cycle with anchoring on against the testnet, then independently recompute the audit hash and confirm it matches the anchored value; run the same cycle with anchoring off and confirm a labelled no-op and no external call. Unlocks: verifiable provenance for a service other parties are trusting.

Milestone 4: Agentverse / marketplace publication and documentation (weeks 15 to 20). Deliverable: Ledgerkeep published as a discoverable service on Agentverse / the ASI marketplace (per the platform’s current publication process), with a public quickstart, the service contract reference, and a limitations document that separates what has run from what has not by evidence class. Verify: locate the published service listing, follow the quickstart to make a call end to end, and confirm the limitations document is present and specific. Unlocks: real discovery and third-party calls, the point at which usage and feedback can begin to accumulate.

After the grant. The repository and the agent_core control plane stay open-source under doom2quake ownership. The service remains published and callable. The immediate post-grant work is hardening the marketplace path from testnet to whatever mainnet publication the operator decides on, and widening the domain library beyond the payments and infrastructure scenarios shipped during the grant. None of that requires further grant money to keep the delivered artifact alive.

5. Ecosystem impact

The durable, reusable output is not one agent, it is the safety layer under it. The agent_core package (guardrails, router, durable state, MCP bridge) is MIT-licensed and separable, and every improvement made for Ledgerkeep lands there, so other ASI builders can vendor the exact tested control plane rather than re-implementing guardrails from a blog post. The MCP surface delivered in Milestone 2 is documented as an integration seam, so a Ledgerkeep call is a worked reference for how to publish and compose a guardrailed agent service on the platform. The testnet audit-anchoring path from Milestone 3 is a general pattern for verifiable agent provenance, liftable by any builder who needs a caller to trust a returned audit trail. All of it is documented, tested, and open.

6. Sustainability and honest limits

What keeps it alive after the money ends: the delivered artifact does not depend on the grant to run. The offline loop and the test suite are reproducible on any machine at zero cost; the published service is small and scale-to-zero; the control plane it depends on is a library we maintain for our own other builds regardless. There is no server we must keep paying for to stop the work from evaporating.

What is NOT built, deployed, or measured (state plainly):


Cite:

@software{sarkar_ledgerkeep_2026,
  author  = {Dipankar Sarkar},
  title   = {Ledgerkeep: A Guardrailed Autonomous Operations Agent as a Decentralized AI Service},
  year    = {2026},
  url     = {https://github.com/doom2quake/ledgerkeep},
  license = {MIT}
}

License: MIT, held by doom2quake. Testnet only; no mainnet, no real funds.