What RPCTION cannot do.

The selection is sometimes wrong

RPCTION has no object classifier. It does not know what a person or a car is. It works structurally, whether a region is bounded more strongly on the outside than it is divided on the inside, whether its outline is supported the whole way round, whether something visually continues it.

That is enough for most things and not enough for all of them. Scenes that are measurably hard:

• An object on a strongly patterned surface, carpet, woodgrain, marble. The selection tends to spill into the surface. This is the one you are most likely to meet, because desks and floors look like this.
• Two objects of the same colour that touch. There is no contour between them to find. Not a weak one, none.
• Something partly behind something else can merge with it.
• Glass, and reflections. A reflection has a real outline and can read as a real object.
• Heavily compressed or downscaled images, where the evidence has been thrown away before RPCTION ever sees it.

The selection is always drawn before you act on it, and ↑ ↓ walks the hierarchy to correct it. That is the intended answer to all of the above, not an afterthought, every failure listed here is one you can see and fix before it reaches your clipboard.

Confidence does not know when it is wrong

This one is worth stating plainly because it would be easy to hide. We measured whether a confidence threshold could turn wrong selections into refusals, across the whole adversarial test corpus. It cannot: the score a wrong answer arrives with is statistically indistinguishable from the score a right one arrives with. Raising the threshold would silence correct selections at the same rate as incorrect ones.

So RPCTION does not pretend to a certainty it has not got. What protects you is that you see the selection first, not a number.

SOURCE cannot reach real source files

SOURCE reads what is on your screen. It does not attach to other applications, read a browser's internals, use a debugging port, install an extension, or fetch anything over the network.

So private, server-side or unloaded code is out of reach, and RPCTION says so rather than approximating it. Everything SOURCE produces today is labelled RECONSTRUCTED or PARTIAL, nothing in this build can report AVAILABLE, which is reserved for a real source file obtained legitimately.

Reconstruction is imperfect in the ordinary way: a character misread is misread, and a block mixing code with other content is clipped to the code with the amount dropped reported.

Text recognition is not perfect

Small text, low contrast, unusual fonts and heavy compression all cost accuracy. Recognition is English only in this release.

Motion loses targets

Tracking survives an object passing behind something else and reports OCCLUDED while it does. It gives up on long disappearances, extreme motion blur, and targets that leave the screen, and reports LOST rather than following the wrong thing.

Two identical objects that cross at the same point with mirrored paths are genuinely indistinguishable at the moment they meet. There is no measurement that determines which continued in which direction, so RPCTION reports the moment as ambiguous instead of guessing. That is the behaviour working, not failing.

AUDIO captures the whole mix

Windows combines system audio before applications can see it, so AUDIO records everything you are hearing rather than one application on its own. Per-application isolation is not offered because it cannot be delivered.

There is no transcription in this release.

Platform and packaging

• Windows 10 / 11, 64-bit only. macOS and Linux are not supported and are not being claimed.
• Not code-signed. Windows SmartScreen will warn on install. The SHA-256 of every published file lets you verify it instead.
• No automatic updates, which is also why it never contacts the network.
• Around 330 MB installed and ~450 MB of memory while running, most of it the runtime and the bundled recognition model.

What this is not

RPCTION is not a cloud service, an assistant, or an agent that does things on its own. ACT is user-directed: you select the target, you choose the action, and only then does anything happen.