Impact Engine — prove growth

Run impact snapshots, scores, stress, and reports to prove the agent is getting smarter.

What you will learn

• impact_run
• impact_score/reports
• impact_enforce

Prerequisites

• Efficiency Engine — token budgets and loops

Surface

What this feature does

Impact Engine audits growth across dependency maps, builders, semantic constraints, and guardrails. Plateaus can be maturity (optimal_plateau), not failure. Adversarial stress proves learning is real.

How to call it

Call these MCP tools through your AGS / tower-gov host — impact_get, impact_validate, impact_enforce, impact_plan, impact_run, impact_compare, impact_score, impact_reports, impact_stress, impact_feed_efficiency.. Prefer scan/get before validate/diff, and enforce only when the change set is ready.

1. impact_plan
2. impact_run
3. impact_stress
4. impact_score → impact_enforce
5. impact_feed_efficiency when coupling budgets

{
  /* Example MCP call — fill args from schema_get / docs */
  "tool": "impact_get",
  "arguments": {}
}

What success looks like

ImpactScore produced

impact_enforce ok

No silent no-op verdict

Common failure modes

Treating plateau as hard fail

Skipping impact_stress

Entitlement lacks impact_full for deep reports

Related reference

• Architecture
• Licensing