Think and Find

This repository implements one generic research task type using iterative Episodes: discovering and assessing examples of pragmatic causal pluralism in natural phenomena. Every Episode owns a recurring decision within an assigned question. It chooses work, receives actual outcomes and revises its next choice. The user's original scientific instructions define the requested result.

The implementation foundation is the existing research composition in case_reports/research/ and the shared method_loop runtime. The discarded parallel causal-discovery implementation is deleted. The launcher now runs this causal-inference research application. The scientific parent and investigations populate and calculate the global sheet; acquisition uses the existing Search Strategy/Search/Page/reading hierarchy.

Episode design

• Episode ownership and recurring decisions
• Reasoning, task instructions and steering
• Task authority
• Search and reading responsibilities
• Durable state and reconstruction
• Credit and numerical continuation
• Executable causal scoring sheet draft

The episode hierarchy narrows each child's decision problem. The parent defines its assignment and requested information; the child iterates and returns that information. The parent then chooses useful work from the updated state. A primitive search, extraction, calculation or reasoning operation is an available action within this structure.

Source layout

Location	Responsibility
method_loop/	Shared iteration, identity, contracts, trace and reconstruction
case_reports/research/	Existing scientific decisions, research state and acquisition episodes
credit_assignment/	Numerical projections and continuation mechanisms
task_executors/	Provider adapters and deterministic tools
configured_providers.py	Provider composition and model routing
checkpoint.py, environment.py, cli.py	Persistence, configuration and launcher

Python imports retain the package name episodic_workflow. Source runs locally; external repositories are design references, not runtime dependencies. Provider configuration is described in .env.example. Credentials and research artifacts belong outside tracked documentation.

Run

python3 __main__.py --progress

Use --root-yield-threshold 0 for a standing investigation. The estimator yields only when nonnegative recent gain is strictly below its threshold, so zero makes that cutoff unreachable. Child thresholds retain their normal values. Provider or input failures remain operational failures. Checkpoints retain the threshold.

The default objective is defined in case_reports/research/task.py. Optional request text or --request-file can supply a scientific scope. Multiple candidate inferences can emerge, each with its own theory applications and scoring subsheet. There is no task selector or disease/scenario argument.

Literature assigned to the coding agent informs the implemented formalism and rubric. It is not a runtime reading list or a prescribed scientific example. Argo is the default model provider through the local shim. Codex OAuth requires an explicit provider selection. Provider credentials and model routing are read from this checkout's .env and shell environment. --checkpoint-path selects a new durable run; --resume continues that run under its recorded request and compatible implementation. A numerical yield is an execution outcome, not a claim of scientific completeness.