AIOA spArkHAT — Personal AI on Nebius

What we built

AIOA spArkHAT is an existing personal AI system extended into one local Linux application: chat, persistent candidate notes, planning, model calls, bounded tools, review and durable receipts. The central rule is that a model proposal is not execution authority.

The demonstrated integration uses NVIDIA nvidia/nemotron-3-super-120b-a12b through Nebius Token Factory. Exact provider/model identity is checked; there is no silent provider fallback. The application has one WorkspaceStore and reuses AIOA Core admission, warrants, ServiceGuard, intent journaling and provenance. Local operator admission uses Linux AF_UNIX/SO_PEERCRED. This delegated UID1000 profile does not prove human presence or certify a production owner.

Why we built it

After a crash or lost acknowledgment, an agent can be unsure whether an effect happened. Repeating the action can cause a duplicate. AIOA records irreversible intent before dispatch, verifies the result independently and retains uncertain outcomes as UNKNOWN. It does not treat a plausible model answer or a second click as evidence that an effect is safe to repeat.

What actually worked

• Two recorded end-to-end READ missions used real Nemotron calls: tool proposal, exact local Core approval, bounded LICENSE read, independent hash verification and a subsequent actual model turn receiving the tool result.
• A real model proposal produced a small JSONL receipt reporter. The corrected program passed seven independent sandbox cases. It counts PASS/FAIL/UNKNOWN; CSV export was not implemented.
• The existing guarded Git integration created a real public draft PR #22, with separate approvals for commit, push and draft PR. No merge is claimed.
• Three recorded post-hoc advisory reviews used three roles of the same Nemotron model. These are not independent models, a full live 1+3+1 review, or execution authority.
• Candidate notes and completed receipts survived application restart. Verified completed actions were inspected without redispatch. A separate local Git/SQLite fault test covers lost acknowledgment and restart recovery; it is not evidence of a live cloud crash in every transition.

The local ledger now contains 18 real model attempts: 16 COMPLETED and 2 UNKNOWN. The two UNKNOWN outcomes and their budget holds remain intact. One coding final response is UNKNOWN, so we do not claim that the model delivered a verified final coding answer. Acknowledgment and filesystem changes are not one atomic transaction.

The current interface

The existing frontend-v2 now presents Chat/Agent, Missions, CPL, Models, GitHub, Linux Tasks, Knowledge/Academy and Non-Zero in the same application. Model selection comes from the backend catalog. Linux Tasks offers approved-contract inspection and logs, not an arbitrary host shell. Muse and browser automation are disabled.

The local owner has now explicitly enabled LIVE Nebius inference for recording under the existing USD 1 allowance. A new bounded model request completed successfully through the actual AIOA application. Previously verified model executions are labelled HISTORY/LIVE RECORDED. A fresh bounded public LICENSE preview can run locally without inference; that preview is explicitly not a new native Core START or a new Nemotron turn.

Validation and cost boundaries

The preceding exact local candidate passed 2,457 Python tests plus 1,627 subtests, with five documented optional-dependency skips. The final frontend change has an additional targeted regression result of 175 Python tests plus 152 subtests and 55 Node UI tests. These are local Python 3.12 results, not a remote full CI matrix or production certification. Automated browser testing was not run in this final phase.

The operator selected a total local inference allowance of USD 1. The ledger estimates USD 0.00991440 settled and retains USD 0.00722220 in holds, leaving a conservative USD 0.98286340 available. These figures are a local conservative ledger, not a provider invoice or account-balance statement. The initial frontend phase used no paid inference. The later explicit LIVE activation was verified with one real model request; it did not execute a native tool. Previous verified native READ missions remain recorded evidence.

Prior work and significant updates

AIOA existed before August 26, 2026. Earlier Core, Memory Patch/HAT, authority, Non-Zero and recovery work is prior foundation. The current integration adds exact Nebius/Nemotron routing, binding of real tool turns to existing Core authority and receipts, bounded native READ and reporter execution, guarded Git publication, persistent same-application integration, negative/replay tests and the consolidated Linux recording interface. Historical CockroachDB/M12 research remains separate; the current local demonstrator uses SQLite WorkspaceStore and does not claim production M12 or CockroachDB certification.

Limits and judging access

The public reporter PR and curated offline inspector build cover their stated functions. They are not the complete authenticated native agent release. Publication of a safe full native source/build and funded, appropriately scoped LIVE judging access remain unresolved. The laptop and owner control transport are not exposed to the Internet. A local offline build or a recorded video does not itself close those gates.

The final public demo video and submission confirmation must be supplied by the project owner. This description does not assert that the project has been submitted.

Feedback

Nebius Token Factory made it practical to test actual Nemotron tool calls with exact model identity and usage metadata. Token limits, interrupted responses and uncertain billing outcomes required explicit handling. The most useful improvements for this use case would be stronger request reconciliation/idempotency, clearer reasoning-token bounds, and an easy way to provide safe, funded judging access. We did not run a controlled cross-model quality benchmark and do not claim a numerical superiority result.

Built With

• cockroachdb
• nebius-token-factory
• nvidia-nemotron
• openai-compatible-api
• python