Research Loop turns an existing research project and a researcher's intent into a bounded, auditable experiment campaign.
Research Goal: The researcher's desired outcome, success criteria, preferences, and constraints expressed for one project. Avoid: Domain module, research recipe
Research Profile: The project-specific, generated representation of a Research Goal plus verified project facts. Avoid: User-authored configuration, domain module
Execution Contract: The approved description of how this project can be run, including commands, working directory, resource class, and limits. Avoid: Environment adapter
Evaluation Contract: The approved description of the authoritative metric source, comparison direction, compatibility checks, and confirmation policy. Avoid: Evaluator module
Research Strategy: The approved approach for choosing among eligible hypothesis-driven Experiments for one Campaign. Avoid: Search algorithm, DAG policy
Strategy Contract: The approved representation of the project shape, initial Selector, rationale, and deterministic Selector transitions. Avoid: Loop Policy, agent preference
Selector: The deterministic policy that scores and recommends the next eligible Experiment candidate under a Strategy Contract. Avoid: Strategy, LLM judgment
Hypothesis Evidence: An auditable observation from a recorded Experiment together with the agent's explicit assessment of how it bears on a Hypothesis. Avoid: Performance result, metric status
Campaign: One approved sequence consisting of a baseline and a bounded number of related Experiments. Avoid: Session, run
Experiment: One minimal, hypothesis-driven project change evaluated against a deliberate parent result. Avoid: Trial, arbitrary code change
Research Ledger: The append-only record that connects every attempted Experiment to its branch, command, artifacts, metric, and decision. Avoid: Log file, notes
Proposal: An agent-authored, runner-validated portfolio of hypothesis and candidate items that changes no state until its accepted items are registered. Avoid: Plan, batch of guesses
Portfolio Lint: Deterministic, Selector-aware coverage warnings over pending and proposed candidates; advisory, never authorizing. Avoid: Critic, gate
Idea Source: A hash-pinned reference to external material that explains where an idea came from; never a substitute for recorded local origin evidence. Avoid: Citation, imported solution
Knowledge Pack: The content-addressed, locally verified set of normalized claim records offered as optional hypothesis-generation input, filled by the user or by approved Agent Retrieval. Avoid: Cache, scraped corpus
Agent Retrieval: The approved, agent-side search-and-normalize workflow that fills a Knowledge Pack from external papers, pull requests, and issues; the runner never fetches. Avoid: Runner network access, web crawling
Research Surface: The approved description of which components may change, the invariants that must hold, and the data flows that are forbidden. Avoid: Allowed paths, executable checks