AIOA spArkHAT — Personal AI on Nebius

What we are building

AIOA spArkHAT is a human-bound Personal AI runtime for long-running agents that need to remember, reason, recover from interruption, and act safely across real workflows.

The core design principle is simple:

model capability is not execution authority.

The model may reason, retrieve evidence, critique, remember verified corrections, and propose actions. Consequential effects remain behind explicit human approval, deterministic authority checks, guarded execution, independent verification, durable receipts, and replay protection.

For the Nebius x NVIDIA Global AI Hackathon we are evolving the existing AIOA spArkHAT codebase into a Nebius-native Personal AI system powered by NVIDIA Nemotron through Nebius Token Factory.

Why this matters

Long-running agents fail in ways that ordinary chat systems do not.

After a timeout, crash, network failure, or restart, an agent may not know whether an external action already happened. A naive retry can duplicate a consequential effect. A model can also repeat the same factual or procedural mistake across sessions unless the correction is independently verified and made reusable.

AIOA addresses both problems:

durable state instead of hidden conversational state; explicit human authority boundaries; independent effect verification; replay-safe receipts; persistent verified memory; bounded multi-model critique; verified epistemic deltas and ZERO_WRITE duplicate prevention; long-running recovery without silently replaying effects.

Existing foundation before the Nebius submission period

AIOA existed before August 26, 2026. Earlier work established several reusable components that are disclosed as prior work rather than claimed as new Nebius-hackathon code:

the AIOA control architecture; human approval and Non-Zero authority concepts; persistent Memory Patch / Knowledge HAT work using CockroachDB; provenance-aware retrieval and verified correction concepts; bounded agent execution and recovery primitives.

Significant updates since August 26, 2026

During the submission period the project was substantially rebuilt and converged into one integrated runtime.

Major post-cutoff engineering includes:

Critical Prompt Loop 1+3+1

A bounded review loop consisting of one primary draft, three ordered critic roles, and one final revision. CPL is advisory-only: critic agreement is not treated as independent proof and cannot authorize execution.

Core-native Non-Zero convergence

Previously separate human-bound execution semantics were integrated into AIOA's Core while preserving explicit approval, durable state, fail-closed transitions, and no silent success.

CockroachDB Memory integration

The existing Memory Patch capability was integrated into the unified runtime with scoped persistent memory, HAT selection, provenance, owner isolation, restart durability, and explicit backend truth.

Verified Epistemic Delta

AIOA can represent a correction as a small verified difference rather than storing an entire chat transcript. A correction is eligible for reuse only after independent evidence checks. Repeating an equivalent verified correction produces ZERO_WRITE instead of duplicating knowledge.

Personal Delta

Verified corrections can be retained as private, owner-scoped model experience subject to consent, HAT scope, freshness, provenance, conflict rules, and storage limits.

Pheromone / DVM research layer

AIOA added bounded tau+/tau- memory dynamics and DEEP/HOT/WORKING/ARCHIVE reference tiers. These signals help rank eligible verified experience, but they never become truth or action authority. Competition behavior remains SHADOW.

Restart, replay, receipts and Service Guard

The runtime now separates proposal, approval, effect dispatch, durable receipt, independent verification, restart reconstruction, and duplicate-effect prevention.

NVIDIA trajectory and reviewer evidence

The runtime was integrated with NVIDIA open model workflows and hardened with deterministic reviewer paths, explicit LIVE versus TEST_FIXTURE separation, reproducibility evidence, and long-running endurance validation.

Nebius-native work

A dedicated branch, nebius-personal-ai, is being developed independently from the frozen competition baseline.

Already implemented on that branch:

dedicated Nebius Token Factory provider; official Token Factory endpoint validation; NEBIUS_API_KEY, NEBIUS_BASE_URL, and NEBIUS_MODEL configuration; NVIDIA Nemotron competition profile; fail-closed Nebius routing with no silent provider fallback; Token Factory /v1/models discovery and Nemotron filtering; exact provider/model identity checks; CPL 1+3+1 exact routing through Nebius; request ID, token usage, and latency metadata; regression tests proving all five CPL generation calls can be bound to Nebius transport.

The next implementation gates are:

first bounded LIVE Token Factory validation; CPL → Verified Delta → reusable Model Experience bridge; Tavily runtime evidence acquisition; Hermes advisory integration; OpenShell / NemoClaw execution isolation; Nebius Serverless packaging and public demo.

Intended Personal AI flow

User task → Knowledge HAT + persistent CockroachDB memory → currently eligible verified model experience → evidence acquisition → Nebius Token Factory / NVIDIA Nemotron → CPL 1+3+1 → independent verification → Verified Delta / ZERO_WRITE → explicit human approval → Non-Zero + Service Guard → bounded sandboxed execution → independent effect verification → durable receipt → restart-safe continuation.

What makes the project different

AIOA does not try to make one model the controller of everything.

Instead it separates:

cognition from authority; memory from canonical truth; critique from proof; scoring from permission; execution from verification.

The goal is an always-on Personal AI that can become more useful over time while keeping the user's data, consent, and decision authority explicit.

Current status

The Nebius provider and Nebius-bound CPL path are implemented and covered by deterministic tests. LIVE Token Factory evidence will be added once the project API key/credits are configured. We will keep LIVE claims separate from fixture evidence throughout the submission.

Team

This is a team submission led by Łukasz Żuchowski with collaborator Natalia Kumoch. Team membership on Devpost is being managed through the project invite flow.

Repository

Public development repository: https://github.com/luciferprosun/AIOA-spArkHAT

Project site: https://sparkhat.eu

Public contact

LinkedIn — Łukasz Żuchowski: https://www.linkedin.com/in/łukasz-żuchowski-807160316/