Verification step 6
Confidence scoring: why ambiguity is a feature
The final verdict combines evidence from multiple rules. Some rules support possible-looking structure, some support impossible-looking contradiction, and some support ambiguity because the input is too sparse or noisy.
The best detector is not the loudest detector. It should say "impossible" when strong signals agree, "possible-looking" when the structure is clean and no contradiction appears, and "ambiguous" when the evidence does not justify a stronger claim.
That restraint matters because impossible shapes are playful. The fun is not in forcing every sketch into a dramatic label. The fun is in noticing when your eye feels certain, then discovering whether the drawing has enough visible structure to support that feeling.
Confidence scoring is the rule-comparison step that chooses possible-looking, impossible-looking, or ambiguous after weighing competing geometry evidence.
Start with the drawing
The plain version
Confidence scoring is the rule-comparison step that chooses possible-looking, impossible-looking, or ambiguous after weighing competing geometry evidence.
Why your eye buys it
The detector should only give a strong impossible verdict when contradiction evidence clearly beats possible structure and ambiguity pressure.
Small details that matter
- Possible, impossible, and ambiguous rules can all fire on the same drawing.
- Ambiguity is intentional when the input is noisy, sparse, or internally unclear.
- Confidence is an educational estimate, not statistical proof.
How to think about the result
- If contradiction evidence strongly outweighs possible structure, then impossible-looking is selected.
- When clean wireframe structure appears without a contradiction, possible-looking is selected.
- If the scores are close or the input is noisy, then ambiguous is selected.
What can shift the verdict
| Cue | A useful sign | How it changes the reading |
|---|---|---|
| Structure | At least a few clean visible segments | The drawing has enough shape to discuss |
| Input clarity | Few noisy edges or unclear overlaps | The result can be less hesitant |
| Contradiction | A clear depth, prong, or loop conflict | The illusion has a stronger impossible-looking case |
Try it with your sketch
- Draw with straight, high-contrast visible edges.
- Leave endpoints and junctions easy to see.
- Compare the sketch with a possible cube, tunnel, or box.
- Use the verdict as a conversation starter, not a final authority.
A small scene
A user can test a cube, a trident, and a noisy upload to see how confidence changes with clearer or weaker evidence.
The core idea
A rule returns a direction and a confidence value. The scorer sums evidence by verdict type, compares the totals, and applies conservative thresholds. This keeps a single misleading signal from deciding the whole result.
Picture three voices in the room. One says the drawing looks coherent. One says a contradiction is forming. One says the sketch is too messy to trust. Confidence scoring is the part that listens to all three before choosing words the visitor can use honestly.
| Evidence type | Example rule | How it affects the verdict |
|---|---|---|
| Possible-looking | Clean repeated direction families with enough long segments. | Raises possible confidence when contradiction evidence is low. |
| Impossible-looking | Known impossible preset hints, nearby crossings, or depth cycles. | Wins only when contradiction evidence clearly outweighs possible structure. |
| Ambiguous | Too few segments, too many noisy upload fragments, or weak crossings. | Prevents a confident verdict when the input does not support one. |
Why ambiguous is useful
Ambiguity can feel less exciting than an impossible verdict, but it is the honest answer when a drawing is underspecified. A 2D picture can suggest an impossible object without containing enough clean structure for a browser-side detector to classify it strongly.
In practice, ambiguous often means "your idea is still interesting, but the lines need another pass." That is a better answer than pretending a fuzzy upload, broken cube, or half-finished fork has been solved. It gives you something concrete to improve.
False positive risk
Calling a normal drawing impossible because a few lines crossed or an upload had noisy edges.
False negative risk
Missing a real contradiction because the drawing hides depth cues in shading or context.
Good ambiguity
The detector says it needs cleaner lines, more explicit junctions, or less noise.
Bad certainty
The detector gives a dramatic verdict without enough evidence to justify it.
Mini experiment
Try the possible cube, the triangular tunnel, and the impossible trident. Then draw a messy hybrid. The point is not to get the same verdict every time; the point is to see when evidence gets stronger, weaker, or contradictory.
After each result, say the label out loud in cautious language: possible-looking, impossible-looking, or ambiguous. That small habit helps keep the tool playful and honest, especially if you share a result with a class, a friend, or another artist.
Possible baseline
A clean cube should produce possible-looking structure and low contradiction pressure.
Impossible baseline
The trident preset should show stronger contradiction evidence than possible evidence.
Ambiguous target
A sparse or noisy drawing should push the detector toward caution instead of drama.