You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
keep AI-generated outputs non-authoritative: provider responses, critic reports, retrieval results, risk scores,
previews, and action proposals can inform a human reviewer, but cannot themselves authorize execution or writes.
The project develops and validates a practical safety architecture built around explicit human approval, hash-bound
action proposals, fail-closed gates, and append-only audit evidence. Funding will support focused engineering,
adversarial testing, reproducible offline validation, and public documentation of the system’s authority
boundaries.
The intended public benefit is reusable safety infrastructure for developers and researchers building AI-assisted
tools: a concrete reference implementation showing how to separate model suggestions from consequential authority,
rather than treating model output, scoring, or guardrails as permission. AIOA does not provide autonomous
execution, provider-driven control, or automatic approval.
The project’s goal is to create and validate an open-source safety architecture for AI-assisted systems in which
model outputs never become authority by default. It will demonstrate how AI suggestions, risk classifications,
critic feedback, retrieval results, and proposed actions can remain useful to a human operator while consequential
actions—such as writes or controlled execution—remain behind explicit, verifiable human approval.
We will achieve this by developing a local-first reference implementation with four core mechanisms: inert action
proposals and previews; explicit, hash-bound human approval for each consequential action; fail-closed policy gates
that reject missing or mismatched authorization; and append-only audit records that preserve evidence of what was
proposed, reviewed, approved, or denied. The work will include focused engineering, adversarial and regression
tests, reproducible offline validation, and clear public documentation of the system’s capabilities and limits.
The outcome will be a reusable, inspectable reference for developers and researchers who want to build AI-assisted
tools without confusing model output, automated scoring, or safety guardrails with permission to act
Funding will be used to move the project from development on limited netbook hardware to a reliable, reproducible
research and engineering environment.
The primary costs are:
- Workstations and essential hardware for secure local development, testing, offline validation, and reproducible
builds.
- API and compute credits for controlled, human-supervised evaluation of different AI models and safety-relevant
behaviours. Model access will be used for testing and comparison; it will not create autonomous authority or
automatic execution.
- Secure storage, backups, and development infrastructure needed to preserve audit evidence, test artifacts, and
reproducible releases.
- Focused engineering time for implementation, adversarial testing, documentation, and open-source maintenance.
The funding will directly strengthen the project’s ability to test safety boundaries under realistic conditions,
publish reproducible results, and make the resulting reference implementation useful to other developers and
researchers.
AIOA–NiFe spArk HAT is my original, founder-led project. I currently lead its architecture, implementation, safety
design, testing, and documentation.
I am also working to build a small international team with a collaborator based in Ireland. The team is still
forming, so I remain accountable for the project’s technical direction and delivery. As the project develops, I
plan to bring in complementary expertise for engineering, evaluation, research review, and community or partnership
work.
My track record is primarily hands-on: I have independently designed and developed AIOA’s local-first safety
architecture, including explicit human approval boundaries, hash-bound action proposals, fail-closed gates, audit
evidence, and deterministic offline validation. I have also developed multiple related technology projects, which
are collected at These projects reflect an ongoing focus on responsible AI, practical
safety controls, local-first systems, and transparent technical documentation.
The most likely failure modes are limited time, funding, and hardware capacity; difficulty recruiting the right
collaborators; and the possibility that the architecture proves too complex or insufficiently useful for external
developers to adopt.
A second risk is that testing may reveal gaps in the proposed safety boundaries. This would be treated as an
expected research outcome, not hidden or worked around: the affected capability would remain disabled or be removed
until it can be supported by clear evidence and tests.
If the project does not reach its intended scale, the outcome would be a smaller but still useful open-source
reference implementation, test suite, and documentation rather than a deployed autonomous system. aioa is
deliberately designed so that failure does not create uncontrolled execution risk: provider output, scores, and
proposals remain non-authoritative, while consequential actions stay behind explicit human approval and fail-closed
gates.
We have raised €0 in grant or external funding during the last 12 months.
The project has so far been developed independently, without institutional backing or external investment. We have
submitted AIOA NiFe spArk HAT to NLnet; the project has been accepted into their review process, and we are
currently awaiting a final funding decision. No funds have been awarded or received from NLnet at this stage.
There are no bids on this project.