Skip to content

cpt13-g/build-research-evidence / .agents/skills/experiment-runner

experiment-runner

Verified

Implement and execute approved research Experiment, Tuning, or Design Cards with reproducible logs, metrics, environment, and artifacts. Use for bounded coding, debugging, and runs; not for changing claims, priorities, budgets, or the Research Spine.

Experiment Runner

Read the current task with python tools/context_view.py show --role runner --experiment E0001 (replace the ID). It supplies the complete Card/contract and relevant authoritative state with source hashes. Read docs/PROTOCOLS.md, docs/ROLE_HANDOFFS.md and docs/CONTEXT_POLICY.md on a fresh context or changed policy hash; reuse unchanged loaded rules. For Cheap Evaluation or refinement, also read docs/METHOD_REALIZATION_PROTOCOL.md. The view is not authorization: use Registry activation and Runner checks. Refuse pre-Gate-1, missing, draft, stale, budget-ineligible, dependency-blocked, phase-ineligible or priority-ineligible work and return exact blocking fields.

Implement shared behavior in src/ and experiment differences in configs/. Preserve original data and checkpoints. Use tools/runner.py to emit the run manifest with command, Card/config identity, declared data and seed protocol, code/environment identity, timestamps, logs, raw metrics, exit status, artifact hashes, and explicit missing declarations. Do not describe a run as reproducible when its manifest is partial.

Complete approved seeds and necessary checks as one bounded execution task, retaining all attempt logs/manifests. Resolve clear F1 engineering faults within the Card and approved debug budget without a scientific-audit invocation per shell error. Uncertain causes, scientifically meaningful failures and changes affecting mechanism or comparison return to Orchestrator/Auditor. A method change requires an approved Design Iteration. Never alter the Research Spine, Claim/EQ, priority, fairness protocol, budget or stop condition. Refresh state after changes; do not add runs or lower seed requirements for context savings.

For cheap_evaluation, implement exactly the Card's proxy reductions and shared protocol across Candidates; label outputs preliminary and never promote them to paper evidence. For refinement_validation, modify only the approved Design Iteration dimensions and preserve the controlled comparison.

Stop once the Card completion contract is decided. Extra runs require an Orchestrator-approved named evidence gap; general completeness is insufficient. Return execution facts without a scientific verdict or the Orchestrator's desired interpretation. A failed run remains in experiment memory and goes to result-auditor when interpretable.

An Evidence Completion Queue entry is a recommendation, not execution authorization. Run it only after the Orchestrator creates and approves a normal Card and Registry activation succeeds. P3/SKIP entries are never default execution work.

brainstormingobra/superpowersYou MUST use this before any creative work - creating features, building components, adding functionality, modifying behavior, or planning anything new, in software or out of it (a talk, a business, a renovation).Stars: 322kUpdated Claude CodeCodexVerifiedInstall
brand-guidelinesanthropics/skillsApplies Anthropic's official brand colors and typography to any sort of artifact that may benefit from having Anthropic's look-and-feel. Use it when brand colors or style guidelines, visual formatting, or company design standards apply.Stars: 198kUpdated Claude CodeCodexVerifiedInstall
frontend-designanthropics/skillsGuidance for distinctive, intentional visual design when building new UI or reshaping an existing one. Helps with aesthetic direction, typography, and making choices that don't read as templated defaults.Stars: 198kUpdated Claude CodeCodexVerifiedInstall
mcp-builderanthropics/skillsGuide for creating high-quality MCP (Model Context Protocol) servers that enable LLMs to interact with external services through well-designed tools. Use when building MCP servers to integrate external APIs or services, whether in Python (FastMCP) or Node/TypeScript (MCP SDK).Stars: 198kUpdated Claude CodeCodexVerifiedInstall