-
-
Approved Thumbnail A and core promise.
-
Persistent goal, purpose, voice and boundaries.
-
Local preparation completed; consequential action awaits approval.
-
Existing readable receipt crop: validated answer, tokens and cost.
-
Receipt reuse after restart, without another request.
Inspiration
Creative projects need continuity across sessions. A real creator promotion workflow exposed the need to remember the project's purpose, preserve voice, ground claims in sources and avoid repeated approvals for ordinary preparation. CreatorAgent translates those lessons into a persistent local worker and a consolidated review process. The public demonstration uses original fictional material only.
What it does
CreatorAgent retains goal, purpose, priorities, identity and voice alongside tasks and evidence. Local preparation continues until a consequential boundary. Explicit simulated approval resumes a fake-destination workflow. Reviewed source metadata separates source statements, quotation, personal experience, interpretation and hypotheses. The prototype does not infer sales demand from writing.
A separately approved competition-runtime task makes a real bounded NVIDIA Nemotron call through Nebius Token Factory, validates it strictly and persists a provenance receipt before local review. On October 10, the actual worker's synthetic count audit returned twelve with exact-match acceptance. Restart with transport disabled reused the receipt without another request.
How we built it
Python and SQLite provide durable tasks, context, permissions, dependency scheduling, artifacts and recovery. The runtime path is Worker → exact consent and fresh preflight → atomic budget reservation/intent → dedicated Nebius adapter → strict validation → atomic usage/receipt → local review. Source identity, speaker, subjective status and permissions are application-owned. Public execution remains a fake adapter; live model transport is disabled in default reproduction commands.
Challenges we ran into
We preserved V1–V8 exactly rather than retroactively improving scores. V8 passed fact, quotation, unsupported-demand and publication-boundary gates, then failed personal-experience machine classification despite correct prose. The voice case was not called. That finding moved provenance ownership into trusted metadata instead of model wording. Neither the full audit nor a general production NVIDIA workflow is claimed successful.
Accomplishments that we're proud of
The previous clean public candidate passed 82 local tests and offline walkthroughs with socket creation blocked. Persistence, scoped approval, bounded accounting, strict validation and crash/restart behavior are covered. The actual worker subsequently completed one consent-gated NVIDIA request: 632 input / 383 completion-reasoning tokens, $0.00012984 API-derived credits under a $0.001 cap, with zero out-of-pocket. The persisted receipt survives database reopen with one request total. New runtime official billing remains pending.
What we learned
Autonomy needs durable purpose, not only a task list. A model's plausible explanation can disagree with protected fields. Source-grounding, conservative budgets, explicit consent and visible failure are part of the product, not afterthoughts. Historical failures are useful evidence when their limits remain visible.
What's next for CreatorAgent
Broader creator-launch generation, general multi-source annotation review, production integrations and unattended operation need additional validation and consent. Immediate release work is recording review, public repository/test-build access and final submission review. No new model calls are needed to prepare this package.
Log in or sign up for Devpost to join the conversation.