INGRID acts, and proves it worked

Until now INGRID could explain your machine. This is the point where she started operating it — narrowly, and with the result checked rather than assumed.

• She finds what is actually installed. Application discovery reads the shortcuts, registry entries and Microsoft Store packages already on the machine, so the list reflects your computer rather than a catalogue somebody maintained by hand. It is cached between runs so the first request is not slow.
• Open, focus, close — one path for all three. Those started as separate mechanisms and have been converged onto the same discovery layer, so an application that can be launched can also be focused and closed, instead of the three commands disagreeing about what exists.
• A sentence becomes a checked plan. Rather than mapping phrases directly to actions, an ordinary request is turned into a plan, carried out, and then evaluated against what was actually asked for. The interesting half is the last step: deciding whether the person got what they wanted is a separate judgement from whether a command executed without error.
• Authority stays local. A model may propose a plan. What is allowed to run is decided here, against a fixed set of permitted actions, and the decision is not something the model can talk its way past.
• Failure modes that only appear in reality. A launch that triggers an elevation prompt no longer hangs waiting for it. A process that exits during verification no longer crashes the check. Both were found by running it, not by reasoning about it.

Named commands still take precedence over open-ended interpretation, so a phrase INGRID already understands exactly is never handed to something that has to guess.