Ranking fourteen polymers from what you can actually observe

Plastic Fantastic asks what you can observe about a piece of plastic and tells you which polymer it probably is. The interesting decision was not which questions to ask. It was refusing to build a decision tree.

Why decision trees are wrong for this

Every plastic identification chart you will find is a flowchart. Does it float? Yes, go left. Does it burn with a yellow flame? Yes, go left again. Follow the branches, arrive at an answer.

Flowcharts are lovely on paper and brittle in a workshop, for three reasons.

One wrong answer is unrecoverable. A branch is a commitment. If you misjudge whether a sample sank — and you will, because a trapped air bubble or a bit of trapped label will float HDPE and a dense filler will sink PP — every subsequent question is being asked inside the wrong subtree. The chart will still produce an answer. It will be delivered with exactly the same confidence as a correct one.

It forces answers you do not have. A tree needs a response at every node to continue. In practice you often cannot answer: no flame to hand, a sample too small to drop, an ambiguous smell. The chart has no representation for "I don't know", so you guess, and a guess in a decision tree is indistinguishable from an observation.

It cannot express partial evidence. Real observations are not binary. A sample that sits neutrally buoyant is evidence, and it is evidence against both the clear floaters and the clear sinkers. A tree can only route you down one branch or the other.

A decision tree turns a wrong answer into a wrong destination. Scoring turns a wrong answer into a slightly wrong ranking.

Scoring instead

So the app holds all fourteen candidates the whole way through. Each answer multiplies each candidate's likelihood by how well that observation fits it, and the list re-ranks. Nothing is ever eliminated — it only becomes improbable.

This has properties the tree cannot have:

• One bad answer costs probability, not correctness. Misjudge the float test and the true polymer drops a few places. Two more observations and it usually climbs back. Errors are diluted rather than amplified.
• "Don't know" is a real answer. Skip a question and it contributes nothing. You are not penalised for the tests you cannot run, which — outside a laboratory — is most of them.
• The ranking is the honest output. When the top two candidates are close, that is not a failure to answer, it is the accurate state of the evidence. A tree would have picked one and shown you nothing about how nearly it picked the other. Displaying the runners-up tells you whether to trust the result or run another test.
• Evidence can be weighted by how diagnostic it is. A resin code stamped in the material is strong evidence. A vague chemical smell is weak. A tree treats every node as a hard fact; scoring lets the weak signals be weak and still count.

Cheap questions first

The question order is deliberate, and the ordering principle is cost to you, not information content.

It asks for the resin code first, because it is free and often decisive. Then observations needing nothing but your hands and eyes: does it float in water, what does it sound like dropped on a hard surface, does it whiten when you bend it, is it rigid or waxy. Only after all of that does it consider anything involving heat.

An information-theoretic ordering would sometimes put a burn test early, because flame colour, smoke and smell are extremely diagnostic. That would be optimal and wrong. A test that requires a flame, ventilation and a willingness to set your sample on fire is not comparable in cost to squeezing it, and most sessions should resolve before you ever reach one.

Naming PVC without lighting it

The safety constraint drove real design, not just a warning screen.

Burning PVC releases hydrogen chloride. Burning PTFE above roughly 350 °C releases fluorinated decomposition products, and the resulting polymer-fume fever is well-documented and thoroughly unpleasant. The classic identification charts cheerfully instruct you to apply a flame and note the result. That is a bad instruction to give someone at a sorting bench.

So both are identifiable from cold observations alone. PVC is dense — it sinks in water, and unlike most sinkers it whitens sharply at a bend. PTFE is denser still, distinctly waxy to the touch, and unusually slippery. Neither needs heat.

The app will reach both conclusions and never suggest burning either. Where a burn test is genuinely the discriminator between two remaining candidates, it asks — but only after the free tests have run, and never for the two polymers where the answer is worth less than the risk of getting it.

What it is not

It is a ranking of likelihood from consumer-grade observations. It is not FTIR, and it will not tell you about additives, fillers, plasticisers, flame retardants or regrind content — all of which shift exactly the properties it measures. For sorting, repair, recycling triage and deciding whether something is safe to machine, that is enough. For anything with a compliance requirement attached, it is not, and it says so.

It runs entirely on the device with no network permission at all, in English, Swedish and Polish. Details and download here.