zwright8 Epistemic Goal Decomposer
Why This Skill Exists
We need this skill because decisions drift when claims are accepted without verification. This specific skill breaks ambiguous missions into executable units.
When To Use
Use this skill when the request explicitly needs "Epistemic Goal Decomposer" outcomes in the Truth-Seeking and Epistemics domain.
Step-by-Step Implementation Guide
- Define the scope and success metrics for
Epistemic Goal Decomposer, including at least three measurable KPIs tied to false certainty and unverified assumptions. - Design and version the input/output contract for claims, evidence, and confidence traces, then add schema validation and failure-mode handling.
- Implement the core capability using hierarchical decomposition, and produce atomic task trees with deterministic scoring.
- Integrate the skill into swarm orchestration: task routing, approval gates, retry strategy, and rollback controls.
- Add unit, integration, and simulation tests that explicitly cover false certainty and unverified assumptions, then run regression baselines.
- Deploy behind a feature flag, monitor telemetry/alerts for two release cycles, and iterate thresholds based on observed outcomes.
Required Deliverables
- Capability contract: input schema, deterministic scoring, output schema, and failure modes.
- Runtime profile: planning-router using hierarchical decomposition to produce atomic task trees.
- Orchestration integration: truth-seeking-and-epistemics:planning-router routing, approval gates, retries, and rollback controls.
- Validation evidence: unit, integration, simulation, regression-baseline suites and rollout telemetry.
Operational Runbook
Preflight
- Confirm the Epistemic Goal Decomposer request scope, source evidence, and measurable success criteria before execution.
- Verify feature flag skill_0007_epistemic-goal-decomposer, approval gates, and rollback owner before autonomous use.
Execution
- Execute hierarchical decomposition with deterministic scoring and reproducible trace capture.
- Produce atomic task trees plus scorecard, assumptions, and unresolved-risk notes.
Recovery
- Fail closed when required signals, evidence, or approval gates are missing.
- Rollback to the last stable baseline when posture is critical or validation fails.
Handoff
- Publish atomic task trees, validation evidence, and telemetry links to downstream owners.
- Queue follow-up tasks for unresolved risks, threshold tuning, or approval review.
Guardrails
- [quality] Require deterministic scoring and validation evidence before promotion.
- [reliability] Preserve retries, rollback controls, and failure-mode evidence for every run.
- [safety] Route critical posture or missing approval gates to human review before autonomous action.