multi-perspective-analysis

name	multi-perspective-analysis
description	Composition layer for multi-perspective deliberation primitives: calibrated confidence annotation, lens-independence discipline, heterogeneous model orchestration (Haiku-proposer / Opus-aggregator), and two-tier disconfirmation (Tier-A always-on for architectural ADRs; Tier-B cross-model adversarial challenge gated to high-stakes contested decisions). Triggers: high-stakes design decisions with genuine uncertainty, adversarial stress-testing of architectural choices, calibrating confidence on research claims, parallel lens fan-out that must remain independent, pre-mortem failure-imagination at planning→implementation boundary.
allowed-tools	["Read","Glob","Grep"]
compatibility	Claude Code
metadata	{"principle":"CLAUDE.md§Context Engineering"}

Multi-Perspective Analysis

Composition layer for multi-perspective deliberation. This skill owns the shared schemas; it does not restate lens methodology owned elsewhere. Every section is a pointer or an activation gate — not a knowledge layer.

No new knowledge — point, don't restate. If this file's body is growing beyond pointers and gate definitions, content has drifted from a reference file into the composition layer. Move it back. Sentinel T06 enforces this.

Satellite files (loaded on-demand):

• references/calibrated-confidence.md — verbal+numeric confidence anchor table (high >80% / med 40–80% / low <40%), GRADE downgrade factors, annotation format, authoring guidance
• references/lens-independence.md — isolation/reconciliation/gate discipline for parallel lens sweeps; correlation-collapse rationale; fragment-file enforcement
• references/heterogeneous-orchestration.md — Haiku-proposer / Opus-aggregator recipe, prompt-caching guidance, cost-model references (MoA, PoLL, MasRouter)
• references/disconfirmation-tiers.md — DI vs. DA distinction; Tier-A (always-on for category: architectural): Falsifier / Steelmanned runner-up / Reversal trigger; Tier-B: cross-model adversarial challenge, gating conditions, protocol

Activation Gate

This skill activates under the honest-uncertainty gate: the invoking agent must be able to name ≥2 plausible paths before any multi-perspective mechanism is invoked. If the invoking agent cannot name ≥2 plausible paths, proceeding would manufacture strawmen — a behavioral-contract violation (Register Objection).

activate_multi_perspective(task) =
    honest_uncertainty == "multiple_plausible_paths"   # always required
  AND (
        stakes ∈ {security, one-way-door, user-visible-breaking}   # OR
     OR tier ∈ {Standard, Full} AND blast_radius ≥ 5_files         # OR
     OR contested_with_evidence                                     # genuine rival supported by citations
     )

HARD gate. Mechanisms behind the gate are invoked only when the gate fires. Cheap, non-multi-agent mechanisms (Tier-A Disconfirmation block, per-claim confidence annotation, pre-mortem gate) are always-on at their respective pipeline tiers — they are not behind this gate.

Mechanism Map

Mechanism	Cost tier	Gate	Owning artifact
Per-claim confidence annotation	Free (inline text)	Always-on for research claims	references/calibrated-confidence.md
Verifier confidence field	Free (inline text)	Always-on	references/calibrated-confidence.md
Tier-A Disconfirmation (ADR body)	Free (prose section)	Always-on: category: architectural	references/disconfirmation-tiers.md, rules/swe/adr-conventions.md
Pre-mortem gate	Free (checkpoint)	Always-on: Standard/Full, planning→implementation	skills/software-planning/references/coordination-details.md
Lens-independence discipline	Free (procedure)	Always-on when fan-out is invoked	references/lens-independence.md
Parallel lens fan-out	Moderate (N agent calls)	Honest-uncertainty gate	references/lens-independence.md
Heterogeneous model assignment	Moderate (mixed tiers)	Honest-uncertainty gate	references/heterogeneous-orchestration.md
Tier-B cross-model challenge	High (extra frontier call)	Honest-uncertainty AND stakes ∈ {security, one-way-door, user-visible-breaking}	references/disconfirmation-tiers.md

Pipeline Wiring

The mechanics of multi-perspective analysis are wired directly into the pipeline artifacts — this skill is the navigation surface, not the execution surface:

• Calibrated confidence → agents/researcher.md (per-claim annotation), agents/verifier.md (structured verdict confidence field)
• Lens-independence discipline → agents/systems-architect.md (Phase 7 DI sub-step), skills/software-planning/references/agent-pipeline-details.md § Multi-Perspective Analysis
• Disconfirmation Tier-A → rules/swe/adr-conventions.md (## Disconfirmation body section, dissent: frontmatter field)
• Disconfirmation Tier-B → agents/systems-architect.md (Phase 7 Tier-B note)
• Pre-mortem gate → rules/swe/swe-agent-coordination-protocol.md § Conversation Checkpoints (named variant), skills/software-planning/references/coordination-details.md § The pre-mortem gate
• DI sub-step → agents/systems-architect.md Phase 7 + skills/software-planning/references/design-synthesis.md § S3
• Heterogeneous orchestration → agents/systems-architect.md (Tier-B cross-model note); references/heterogeneous-orchestration.md for recipe details
• Contradiction mapping → agents/roadmap-cartographer.md (Phase 3.5 cross-lens contradiction map); the cartographer applies lens-independence discipline and per-claim confidence annotation when surfacing contested prioritization evidence

Related Skills

• skills/software-planning — three-document model, step decomposition; this skill's mechanisms are activation-gated extensions of the software-planning pipeline.
• skills/software-planning/references/design-synthesis.md — the lens catalog and convergence signals for pre-implementation synthesis; multi-perspective-analysis is the deliberation layer that complements synthesis.
• skills/spec-driven-development — REQ-ID stability as a convergence signal (cross-referenced in design-synthesis.md).