- Carbon black absorbs across the near-infrared, so a black PP tray on a black belt is a zero-signal void to every NIR sorter on the market (TOMRA, Machinex, Pellenc). It falls to residue and landfill. But that same tray is not silent in the mid-wave IR. We read it there. 🧵
- 3 to 15% of the waste stream is black plastic lost this way (WP45 research, 2026). On a 50,000 t/yr MRF at 5% black that is ~2,500 t/yr of recoverable rPP, on the order of a 2.2 to 2.7M euro/yr opportunity. The signal was always there. Nobody was reading it.
- Assay is a side-belt sensing station that reads the black-heavy reject from your primary NIR sort (we do not replace it). Per object: a 154-band MWIR spectrum, 2.7 to 5.3 microns, into a from-scratch numpy 1D-CNN that labels PP/PE/PS/ABS/PVC in a ~5 ms hot path.
- Held-out balanced accuracy is 0.894, measured on a deliberately dirtier-than-train test set (n=1,700). We anchor it near the 83.4% peer-reviewed field benchmark (Resources, Conservation & Recycling, Jan 2026), not the ~99% lab figure. An honest cold-start number.
- A MRF does not sell accuracy. It sells a bale that meets a spec: APR Grade A rPP is at least 97% PP and at most 0.5% PVC. Miss it and you have a rejected bale, a landfill redirect fee, and a damaged downstream-recycler relationship. The spec is the product.
- So we ran the same 0.894 model on the same stream (demo seed 11) and flipped one toggle. Confidence gate ON: the bale lands at 99.2% PP, 0% PVC leak, PASS at APR Grade A. Gate OFF (naive argmax): 94.2% PP, 1.94% PVC, FAIL. The classifier never changed.
- The gate is a calibrated abstain: if max softmax is under 0.86, the object routes to MANUAL QC, never guessed into the bale. PVC gets a hard-reject on a low threshold, because one PVC false-negative in a PP bale is a chlorine bomb. Plain code, sitting outside the model.
- The cost lands on throughput. At those thresholds it routes ~12.6% to manual QC and captures ~99.8% of PVC (1 leak across 60 bales). Across 40 random bales the gated line clears APR Grade A ~85% of the time vs ~22.5% ungated. It abstains rather than contaminate the bale.
- Every passing bale ships a signed sha256 certificate: measured composition, each APR clause checked, model hash, abstain and reject counts, recovered tonnage and revenue. Filable purity evidence an EPR or PPWR auditor can read. That record outlives any future model.
- One question for anyone running a sort line: is a ~12.6% manual-QC rate a fair price for a bale that clears APR Grade A, or does that abstain rate break your throughput math? The tradeoff worth arguing is the gate, not the model. #Recycling #CircularEconomy #EdgeAI #MLOps
- Run Assay yourself: the live MWIR sort, the gate ON pass versus gate OFF fail, and the signed bale certificate, every decision on file. https://veriprajna.com/demos/materials-recovery-ai
Published on X · September 4, 2026
On social media
See this post on its original platform
In our archive