How isb333/crucible works
Crucible is a verifier-grounded search engine: a user supplies a task with editable regions, and independent LLM workers iteratively edit candidate artifacts until a verifier accepts one or the run budget is exhausted. The orchestrator manages workers and rechecks a prospective winner in a fresh sandbox; integrity checks guard against altered frozen regions and gaming. Results, episodes, artifact versions, and model reasoning are recorded in SQLite for later inspection. The README also documents scored optimization and code, proof, and chemistry verifier capabilities. Wiring for unsampled modules is shown only where documented or supported by the tree.
Main entry points: orchestrator.py Worker orchestration · worker.py Worker loop · verifiers Verifier implementations · cli.py CLI
Components 19
Entry and Tasks
- CLI
crucible/cli.py - SDK entry
crucible/__init__.py - Task definition
crucible/task.py - Editable artifact
crucible/artifact.py
Search Engine
- Worker orchestration
crucible/orchestrator.py - Worker loop
crucible/worker.py - Region edits
crucible/edits.py - Search budgets
crucible/budgets.py
Verification and Integrity
- Verdict contract
crucible/verify.py - Verifier implementations
crucible/verifiers - Integrity gate
crucible/integrity.py - Execution sandbox
crucible/sandbox.py
Models and Advice
- LLM session API
crucible/llm.py - Model providers
crucible/providers.py - Advisor
crucible/advisor.py
Run Records
- Run store
crucible/store.py - Reasoning inspection
crucible/cli.py
Other components
- Task author
- LLM services
Connections 26
- Task author → CLI: starts run
- CLI → SDK entry: invokes
- SDK entry → Worker orchestration: starts search
- Worker orchestration → Task definition: loads task
- Worker orchestration → Editable artifact: builds initial
- Worker orchestration → Worker loop: dispatches workers
- Worker orchestration → Search budgets: configures limits
- Worker orchestration → Integrity gate: creates checks
- Worker loop → LLM session API: requests turns
- Worker loop → Region edits: applies edits
- Worker loop → Verifier implementations: checks candidate
- Worker loop → Integrity gate: checks edits
- Worker loop → Run store: records episodes
- Worker orchestration → Verifier implementations: rechecks winner
- Verifier implementations → Verdict contract: returns verdict
- Verifier implementations → Execution sandbox: runs checks
- Integrity gate → Execution sandbox: fresh reverify
- Worker orchestration → Execution sandbox: creates workspaces
- Worker orchestration → Model providers: creates sessions
- Model providers → LLM session API: implements sessions
- Model providers → LLM services: sends prompts
- Worker loop → Advisor: consults optionally
- Advisor → LLM services: requests advice
- CLI → Reasoning inspection: displays history
- Reasoning inspection → Run store: reads history
- Worker orchestration → Run store: records run events