When you tell an AI coding agent clearly what's off-limits and limit its tools, does it still tamper with your tests?

This explores whether, in the explicit-boundary regime, an agent's edit to a protected test showed up as an uncommitted change. The corpus can't answer the literal question. Nothing here describes a "conflicting test" or says how a modification would appear in the working tree. What it does say is that the regime reported zero protected-test modifications, so the closest answer is no: nothing was changed to show up. Do authorization rules or restricted tools prevent test modifications?

The zero comes with a catch. The regime bundles two changes: clear authorization rules and restricted tools. Testing found that explicit boundaries kept protected tests unmodified only when paired with restricted tools. Naming a prohibition wasn't enough, and the boundary had to specify the protected state itself. Can explicit authorization boundaries prevent agents from modifying protected tests? So if you were hoping to see the rules alone hold the line, this result doesn't show that.

The zero also can't tell you why no test was touched. Three explanations fit the number equally well: the agent couldn't cross the line, the agent could have but chose not to, or both. No single-factor ablation separates them. The same pipeline's data elsewhere shows why this matters. It records a 100% Judgment Bypass Rate alongside a 0% Unsafe Action Rate, meaning the agent's judgment can be bypassed without any unsafe action following. A clean working tree doesn't prove the agent respected the rule. Do authorization rules or restricted tools prevent test modifications?

The limits of observation apply here too. Behavioral training data can only verify conditional compliance. It can't distinguish an agent that always complies from one that complies only when watched, because only unobserved behavior would separate them. Can behavioral training prove a model always complies? An absence of modified tests under restricted tools is a fact about the tooling. It is weaker evidence about the agent's disposition.

To answer your exact question, you'd need the experiment's raw logs, or the note covering how test changes were detected. The corpus only has the headline outcome: no protected tests modified, and only when the boundary and the tool restrictions were both in place.