Hybrid Decision Analysis

Combine expert rubric-based review with deterministic scoring to produce reliable decisions across domains.

When to Use This Skill

• You must choose between two or more options and need an auditable answer.
• The decision spans skills, tools, architecture, workflow design, or build-vs-buy.
• You need both current-platform fit and cross-platform viability.
• The user asks for objective ranking, repeatable scoring, or empirical bakeoff evidence.

Inputs

• Decision statement (one sentence).
• Current platform.
• Primary user/workflow context.
• Hard constraints and non-goals.
• Time horizon (immediate, near-term, long-term).
• Initial alternatives (or discovery scope).

Workflow

1. Define decision and constraints.

2. Build option set with at least three alternatives when possible.

3. Choose evaluation mode:

   • static: document and architecture evidence only.
   • bakeoff: scenario-based output testing.
   • hybrid (default): static + bakeoff evidence.

4. Select a rubric pack from references/rubric-packs.md.

5. Confirm criteria, weights, platform set, and mode before scoring.

6. Evaluate each alternative against the same rubric.

7. Apply hard-constraint gate:

   • Mark alternatives as feasible or infeasible.
   • Do not recommend an infeasible option unless all are infeasible.

8. Score deterministically with:

   python3 skills/comparative-analysis/scripts/score_alternatives.py \
     --input <analysis-input.json> \
     --output <analysis-report.md> \
     --json-output <analysis-result.json>

9. Run reliability checks from references/scenario-bakeoff-protocol.md:

   • Coverage and evidence-gap check.
   • Recommendation-consistency check (rank, constraints, action alignment).
   • Optional weight-sensitivity check for close scores.

10. Produce recommendation and decision record using assets/hybrid-decision-record-template.md.

Evidence and Scoring Rules

• Use one rubric for all options in a run.
• Each criterion must include:
  • id
  • name
  • weight
  • metric
  • data_source
  • scoring_rule
• Use 0-100 criterion scoring for deterministic aggregation.
• If raw rubric is 1-5, convert with score_100 = score_5 * 20.
• Record missing evidence as gaps, not assumptions.

Recommendation Actions

• select: top option is strong and clearly ahead.
• compose: top two are both strong and close.
• improve: best option is viable but below direct-select threshold.
• extend: best option is strong overall but weak on current platform.
• build-new: no viable option meets minimum requirements.

Always justify action with concrete score deltas and feasibility status.

Output Requirements

• Ranked list of all options.
• Feasibility status (feasible or infeasible) per option.
• Major-platform average and current-platform score.
• Final action and chosen option(s).
• Key risks, evidence gaps, and follow-up actions.

Quality Gates

• Do not score before criteria confirmation unless user explicitly asks for one-shot output.
• If one-shot output is required, state assumptions explicitly.
• Do not rank options using different rubrics.
• Do not hide hard-constraint violations inside averaged scores.
• Keep rationale short and evidence-linked.
• Apply an agentic-eval loop (Generate → Evaluate → Critique → Refine), max 3 iterations.
• Use structured JSON for critique output and stop if no improvement between iterations.

References

• Rubric packs: references/rubric-packs.md
• Bakeoff protocol: references/scenario-bakeoff-protocol.md
• Reusable bakeoff fixture: assets/bakeoff-fixture.v1.json
• Reusable bakeoff results template: assets/bakeoff-results-template.v1.json
• Deterministic input schema: ../comparative-analysis/references/input-schema.md
• Decision record template: assets/hybrid-decision-record-template.md