README scanner section: the "third mutation" claim is never demonstrated, and glm-5.2:cloud confuses viewers

[higagan]

From the UX review. Two related issues in the Red-Team Scanner section.

1. The marquee "adaptive" claim is never shown succeeding

Line 123 sells the differentiator: "evolving each attack until a seed breaks through … a model that resists the obvious attack may still fall to its third mutation."

But the glm-5.2:cloud transcript (lines ~162–189) shows, for every seed: gen 1 → [✅ SAFE] → [🧬 MUTATING] → <variant printed>, then 3 attack attempts across 3 seeds and 0/3 broke through. The mutations are generated but never actually probed — the run stops at generation 1. So the headline feature ("falls to its third mutation") is never demonstrated firing. A critic can note in 60 seconds that the marquee feature didn't run in the project's own demo.

Note: that transcript also predates the 0.3.6 rewrite (separate attacker call). With current code, an aligned target's mutation lineage ends when the attacker declines — so a real gen-2+ breakthrough is only reachable against a target willing to author attacks (see #35, --attacker-model).

Fix: either capture a run where a mutation is actually probed on a later generation (attempt count > seed count, e.g. SAFE → MUTATING → 🚨 VULNERABLE at gen 3), or soften line 123 to something like "generates a mutated variant for the next generation, budget permitting" and drop "evolving each attack until a seed breaks through."

2. glm-5.2:cloud reads as unfamiliar to an outside audience

To be clear: this model is real — it's a local Ollama model on the maintainer's machine, and the transcript is genuine output (verified). But to a demo/HN viewer who doesn't know that, "glm-5.2:cloud at http://localhost:11434" is momentarily confusing (a "cloud" model at localhost?) and the name isn't recognizable next to gpt-4o-mini / qwen2.5:1.5b.

Fix: use a model a viewer will recognize for the "resistant model" example, or add a one-line note that :cloud is an Ollama-proxied hosted model. This connects to the mutation point — a hosted --attacker-model (#35) would let you show a real gen-2 breakthrough against a nameable resistant model.

Done when

• The scanner section either demonstrates a real mutation breakthrough, or the claim is softened to match what the transcript shows
• The resistant-model example uses a recognizable model, or explains :cloud