Phase 1a (additive): measure safety-gating rate (induced-danger) in mcpaql-parity

[mickdarling]

Part of MCPAQL/website#32 (Epic: MCPAQL 2026 product launch). Phase 1a — additive.

What

Add a safety-gating metric to the mcpaql-parity runner that measures how the MCPAQL-adapted path handles operations that can do damage, versus raw MCP.

This reuses the induced-error harness pattern already built in #21 (closed). Instead of injecting a bad-argument response, the harness injects tasks that should trigger a destructive or irreversible operation, and we measure what each configuration does with it.

Metric: induced-danger gating rate — across a set of tasks that would trigger destructive/irreversible operations:

• How often does the MCPAQL path correctly surface the danger up front / require a confirmation step / refuse?
• How often does raw MCP simply execute it unguarded?

Why

This makes the Programmatic Safety & AI Governance pillar measurable instead of asserted. Right now Phase 1a proves token economy (96%) and is set up to prove LLM correctness + recovery (tools#22). The safety/governance pillar has no measured proof point — this is the cheapest way to produce one, because the harness already exists.

The result feeds the case study (MCPAQL/website#34, new "Safety & Governance" section) and the Phase 1b homepage governance band.

How

• Define a task set that would trigger destructive/irreversible operations (delete, force-push-equivalent, money-movement-equivalent, etc.) against the GitHub MCP adapter surface
• Reuse the Phase 1a: Extend mcpaql-parity runner with LLM-correctness metrics #21 induced-error harness scaffolding; add a "danger-gating" task category
• Record per-task: did the config surface danger / require confirmation / refuse / execute unguarded
• Aggregate a gating-rate summary (raw vs. MCPAQL-adapted), markdown-embeddable
• Tests/fixtures that validate the metric without requiring a live benchmark run

Scope note

Additive — does not change the scope of #22. That issue's four metrics (tool-list tokens, first-call success, turns, recovery) stay as-is. This adds a fifth, optional metric category that #22 can include if the run is cheap, or that can run separately.

Phase

1a (additive).