-
Notifications
You must be signed in to change notification settings - Fork 3
Master‑Runners‑Codex‑V50
The Master Runners Codex is a companion to GCP V50. It provides a library of deterministic runner definitions that a large language model can execute without additional context. Each runner encapsulates domain‑specific expertise and enforces a consistent contract.
After the Init GCP prompt, the protocol executes an early sequence of setup tasks: Data Rights, CartridgesParsePack, Planner and Memory. The planner generates plan/sequence.json, which lists the ordered runner calls. Each runner call is wrapped in a JSON envelope specifying the runner name, inputs, settings (mode, risk), run ID and phase. The runner returns outputs, a gate status (pass, needs‑human, fail) and telemetry.
All runners adhere to a universal contract:
| Field | Description |
|---|---|
| Inputs | File names or JSON values required to start the runner. |
| Steps | Ordered list of actions for the LLM (e.g. tools to call, prompts to use, sub‑runners to invoke). |
| Artifacts | Files produced by the runner. Names and formats follow the GCP file layout. |
| Gate | Name of the gate and criteria for pass/needs‑human/fail. |
| Prompts/Templates | Pre‑defined prompts saved in prompts/ to guide complex tasks. |
| Telemetry | OpenTelemetry spans and metrics (token counts, durations, novelty scores, etc.). |
| Observability | Instructions for capturing exceptions, retries and rate‑limit events. |
| Dependencies | Other runners that must execute before or after this one. |
Mandatory artifacts include updates to Gate_Signals.json, Evidence_Index.jsonl and ClaimGraph.json. Some runners compute additional metrics, such as Hallucination Confidence Bound (HCB) or Model Risk Metric (MRM), and write them to HCB.yaml and MRM_CARD.md. Runners that process personal data must produce a Data_Rights_Card.md.
The codex defines runners across all phases of GCP V50. Examples include:
- Creative Discovery – generates and evaluates creative sparks.
- Research & RAG – retrieves and summarises research data and computes relevancy and reliability metrics.
- Classification – categorises ideas or data points using structured taxonomies.
- Simulation & Evaluation – runs numerical models or benchmarks to assess designs.
- Ethics & Compliance – evaluates data rights, export control, AI Act compliance and risk tiers.
- Documentation & Packaging – creates reports, diagrams and packages for the Exit Wizard.
Each runner defines micro‑gates and risk controls tailored to its domain. For example, simulation runners must bound runtime and memory usage; research runners must filter out low‑reliability sources; privacy runners must enforce differential privacy budgets.
Refer to the full codex (GCP V50 Master Runners Codex.md) for detailed definitions and ready‑to‑run prompts.