ORPHEUS Ω
ORPHEUS Ω is an autonomous evidence-to-pilot engine built for the Taskmaster category. A human supplies a measurable mission; ORPHEUS reconstructs why a technical idea failed, designs modern candidates, hardens the implementation plan, executes applicable deterministic tools, attempts to falsify its own conclusions, and returns an evidence-controlled decision.
It is not a standard chatbot and it is not one model role-playing as a team.
Real Google ADK architecture — v0.9
The Gemini execution surface now contains 18 real LlmAgent specialists, 4 real ParallelAgent squads, and one deterministic SequentialAgent root. Every specialist writes to a unique ADK output_key, so downstream state is explicit and testable.
The workflow is:
- ORION creates the measurable mission contract.
- VIGÍA, NYX-7, and VEGA work concurrently on evidence, provenance, failure modes, and verification thresholds.
- ATLAS-9 creates candidate architectures and rejection rules.
- FORJA Ω works concurrently through three real engineering agents:
- FORJA CORE defines typed modules, state, tools, retries, timeouts, failure states, and security contracts.
- FORJA TEST defines acceptance tests, regression gates, failure injection, and conditions that block false closure.
- FORJA UX defines a light chat-first proof interface backed only by real runtime events.
- SPARK calls applicable deterministic tools and reports raw results.
- RECURSOR-Ω and NÉMESIS-Ω run concurrently:
- RECURSOR detects plan weaknesses, architecture and programming defects, repeated failures, unsupported closure, technical debt, and defects in the work method itself.
- NÉMESIS ruthlessly attempts to disprove the preferred route and find a stronger lawful alternative without bypassing safety, privacy, authorization, intellectual property, platform, or hackathon rules.
- HELIX-8 runs after both audits and scores only demonstrated evidence from 1 to 5 for operational utility, architectural discipline, and demo/production readiness.
- AUREUS-7, BASTION, ECHO, RIFT, and VANTA-0 work concurrently on sustainability, approval gates, provenance, blockers, and alternate routes.
- KIRA Ω resolves disagreements and produces the final evidence-controlled decision.
KIRA cannot declare completion when RECURSOR returns FAIL or HELIX reports that mandatory submission viability is FAIL.
Hypermodern mission chat
The /adk interface was rebuilt as a light ChatGPT/Codex-style mission cockpit with three evidence-oriented zones:
- a Constellation rail loaded from the actual
/architecture/agentstopology; - a live mission conversation showing user direction, agent output, tool calls, tool results, state changes, artifacts, failures, and KIRA's final answer;
- an evidence inspector counting real events, tool calls, errors, final-response observation, backend, model, dependency state, and session identity.
Agent statuses are driven by actual ADK event authors and session-state deltas. The interface does not generate fake activity.
Honest live execution bridge
FastAPI connects to Runner.run_async() through:
GET /adk/readinessPOST /adk/runPOST /adk/stream
The bridge refuses to imitate a cloud-agent run when the backend is mock or credentials are missing. The public trace includes actions and evidence but suppresses private chain-of-thought, raw binary data, secrets, and private file URIs.
Bounded scientific provenance
The passive-cooling catalog includes five mechanisms supported by at least ten primary or peer-reviewed engineering sources. Every entry contains a DOI or stable proceedings URL, evidence type, known limitations, and a bounded statement of exactly what the source supports.
ORPHEUS separates source verification from application validation. A paper may verify that a mechanism exists without validating the current ORPHEUS geometry, target climate, food safety, manufacturability, patent position, demand, or economics.
Current deterministic reference mission
Design an affordable, locally manufacturable, grid-free food-preservation concept by combining passive-cooling ideas.
The repository includes a deterministic climate-aware proxy and independent verifier. It can reject candidates that fail climate, cost, safety, or performance constraints. The model is preliminary: it is not CFD, field validation, food-safety certification, patent clearance, measured customer demand, or confirmed revenue.
Reproducibility and proof
The public repository now includes:
- real sequential and parallel Google ADK orchestration;
- 18 specialist agents with unique state outputs;
- FORJA engineering contracts before execution;
- RECURSOR evolutionary auditing;
- lawful NÉMESIS falsification;
- HELIX evidence-only scoring;
- deterministic execution tools under SPARK;
- live browser streaming from the real ADK Runner;
- a light responsive mission chat;
- provenance integrity checks;
- explicit approval gates;
- machine-readable topology;
- automated tests and GitHub Actions verification;
- reproducible local, Gemini API, Vertex AI, Docker, and Cloud Run instructions.
GitHub Actions verified the v0.9 graph, unit tests, reproducible deterministic mission, Python compilation, JavaScript syntax, and UI asset contracts before the implementation was merged to main.
Truth boundary and remaining proof
The official criterion maximum is 5/5; permitted bonuses can raise the final hackathon score to 6/6. “10/5” is an internal aspiration for impact, not a fabricated rubric score.
ORPHEUS does not yet claim:
- a preserved real Gemini or Vertex execution;
- completed Cloud Run deployment proof;
- shared multi-instance session persistence and crash-safe resume;
- a completed external beneficiary action;
- final architecture PNG/PDF;
- final public four-minute demo video;
- field testing, CFD validation, food-safety approval, patent freedom-to-operate, confirmed customers, grants, pricing, or revenue.
Those items remain the critical path before final submission. ORPHEUS Ω was created during the All Things Agentic submission period, and pre-existing libraries, public datasets, papers, templates, and open-source components will be disclosed and attributed according to their licenses.
Log in or sign up for Devpost to join the conversation.