# Terminator

> Guidance for the Terminator's qualitative stop check: judge whether the optimization has plateaued in ways the numeric scores don't capture, from the latest round's analysis text, and return stop / continue.

- Skill: `nvidia-nemo/terminator` (Agent Skill)
- Install (CLI): `npx skillmds@latest add nvidia-nemo/terminator`
- Raw SKILL.md: https://api.skillmd.com/api/skills/nvidia-nemo/terminator/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- License: Apache-2.0
- Author: NVIDIA NeMo (https://skillmd.com/u/nvidia-nemo)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/nvidia-nemo/terminator

---

<!-- SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved. -->
<!-- SPDX-License-Identifier: Apache-2.0 -->

# Terminator — qualitative stop check

You are deciding whether the optimization has **qualitatively plateaued** — in ways
the numeric scores don't capture — from the latest round's `analysis` text. This
runs only after the deterministic checks (round budget, Pareto-front convergence)
were inconclusive.

Return **True (stop)** only when you can cite specific evidence in `analysis`, such as:

- the analyzer flagging the same root cause across multiple rounds,
- survivors converging to architecturally identical mutations,
- diagnostics getting shorter / less specific (a sign of diminishing returns),
- every new mutation reshuffling the same component.

Return **False (continue)** when:

- there is no specific textual evidence of qualitative stagnation,
- the `analysis` is short, missing, or describes ongoing improvement,
- you would have to assume any value not present in `analysis`.

When in doubt, return False — stopping early forfeits the remaining round budget.