ASCEND Closed-loop Contract

ASCEND, the Autonomous Scientific Campaign Engine for Novel Discovery, combines an autonomous, independently maintained nanobot Agent with a durable campaign control plane. It advances hypothesis-to-evidence cycles within explicit policy. Model turns may fail or be retried; Campaign, Iteration, approval, and execution records remain deterministic.

Runtime and Model Configuration

Composer and ASCEND are both based on nanobot, with independent embedded runtime copies. Their provider, configuration, model-preset, and LLM runtime sources are aligned directly. Both use the same YAML model fields:

llm:
  provider: deepseek
  model: deepseek-chat
  base_url: https://api.deepseek.com
  api_key: ""
  chat_completions_path: /chat/completions
  timeout_seconds: 30

The modules do not share configuration or runtime state. Composer owns package generation, session sandboxes, and composition tools. ASCEND owns campaign tools and keeps generic shell execution denied. Each has its own key, workspace, sessions, process, port, and generated nanobot config.

Round Lifecycle

chat -> Campaign Spec -> Iteration candidate
  -> POST Composer /api/composer/jobs
  -> package review
  -> PACE run -> LabFlow -> LabBridge
  -> pace-result-bundle/v1
  -> ASCEND Decision Proposal
  -> next_iteration | stop | request_human | blocked

Each Iteration preserves its candidate, hypothesis, composition job, package hash, PACE run, Result Bundle, and decision proposals. A new round creates a new row and never overwrites prior evidence.

Device Results

Device-specific parsing belongs to the controller implementing a LabBridge capability. output_schema is the controller's result contract. Successful results contain compact JSON values and object-store artifact URIs. LabBridge validates the object before accepting the terminal operation update.

Composer selects declared output paths when building result_contract. PACE generically extracts those paths and emits pace-result-bundle/v1. No device parser is added to ASCEND, PACE, or LabFlow.

Decision Safety

The ASCEND Agent reads only committed Result Bundles and process-scoped PACE history. It records a structured proposal with evidence references, constraint checks, candidate parameters, confidence, and a concise reason. Deterministic code enforces parameter bounds and maximum iterations.

Automatic proposal/application requires campaign policy. Package execution still requires an explicit approval reference. Human Task scientific data is submitted through LabBridge, not chat.