PROJECT SUMMARY

I'm Mike Olivares, an independent researcher and technology-operations practitioner in New Haven, Connecticut. I built RESIDUAL after years of working with production systems where a tool saying “success” is not the same thing as evidence that the intended state changed.

RESIDUAL asks a concrete question: can software-agent workflows become more dependable if the AI can propose actions but a host-controlled layer retains the authority to accept them? The host binds acceptance to evidence about the evaluated candidate, checks explicit authorization, limits recovery attempts, and escalates unresolved cases to a human.

The existing implementation is research apparatus, not proof that the hypothesis is true. This project will test the design against simpler controls and publish the results, including negative findings.

GOALS AND METHOD

Over approximately three months I will:

1. Freeze a benign task corpus, independent checks, ablations, and an analysis plan after a small pilot.

2. Compare minimal controls, individual components, and the complete design using matched tasks and inference budgets.

3. Test stale evidence, duplicate results, misleading completion reports, authorization mismatches, and shared-verifier errors only inside isolated toy workflows.

4. Measure independently correct completions, unauthorized accepted transitions, acceptance coverage, human interventions, and total execution and verification cost.

5. Publish a reproducible evaluation package and a public report describing results, uncertainty, and limitations.

Zero observed violations will not be presented as universal safety. Results on bounded present-day tasks may not transfer to more capable systems.

USE OF FUNDING

The minimum funding level is $5,000. That would support a smaller but meaningful pilot using dedicated local evaluation capacity and paid model/API access.

At the full $15,000 level, the budget is:

- $5,000 for dedicated evaluation hardware: an isolated research workstation or test host, accelerator capacity, memory, encrypted storage, power protection, and related components.

- $6,000 for AI model APIs and research/evaluation subscriptions used in matched runs.

- $3,000 for cloud compute, sandboxing, artifact storage, and reproducibility infrastructure.

- $1,000 contingency for additional evaluation capacity or directly relevant research tools.

Intermediate funding would scale the number of matched runs, models, and replications while preserving the core pilot.

COMMERCIAL AND INTELLECTUAL-PROPERTY BOUNDARY

RESIDUAL predates this grant and may later support commercial products or services. This request funds only the scientific evaluation, dedicated research infrastructure, and direct consumption costs described above. It does not fund or transfer ownership of pre-existing or independently developed source code, architecture, trademarks, customer features, deployment tooling, product design, or commercial know-how.

Grant-funded research results, publications, and the defined reproducibility package may be made public and licensed as required by Manifund's agreement. No ownership, equity, exclusive license, platform control, or commercial rights in the broader RESIDUAL system are offered or implied. Future proprietary and customer-facing development will be separately funded and accounted for. I will maintain a dedicated expense ledger and receipts.

TEAM AND TRACK RECORD

I am the sole human investigator. My professional background includes infrastructure, networking, automation, incident response, and production technology operations. I have already built and published the RESIDUAL prototype, technical documentation, project site, and repository. I have invested roughly 120 human research hours over three months and about $300 of my own funds.

Project: https://ninja-ops-guy.github.io/residual-agent-harness/

Repository: https://github.com/ninja-ops-guy/residual-agent-harness

Portfolio: https://ninja-ops-guy.github.io/theemrld-portfolio/

FAILURE MODES AND VALUE IF THE HYPOTHESIS FAILS

The controls may add cost without improving outcomes, shared verifier errors may defeat the separation, the benchmark may prove too narrow, or limited compute may reduce statistical power. These are useful outcomes if measured honestly. The report will distinguish implementation failure, experimental uncertainty, and evidence against the hypothesis. A negative result would help teams avoid unnecessary complexity.

FUNDING RECEIVED AND PENDING REQUESTS

I have received no grant funding for this study in the last 12 months. Pending requests concern the same proposed work: OpenAI research/API support ($25,000), NSF SBIR Project Pitch 00126349 (a request for an invitation, not a full proposal or award), the EA Funds Transformative AI Fund ($15,000), Emergent Ventures ($15,000), and BlueDot Rapid Grants ($15,000).

These are alternative funding routes, not cumulative budgets. If more than one progresses, I will disclose the overlap before accepting funds, allocate only non-overlapping costs where each funder agrees, or reduce/withdraw overlapping requests so the same expense is not funded twice.

AI ASSISTANCE DISCLOSURE

AI tools helped organize and edit this proposal, as they also assist parts of my research workflow. I reviewed every claim, supplied the underlying facts and plan, and take responsibility for the submission.