The point of sorting is not simply to produce a machine label or move an object into another bin. Recovery teams need to know what the outgoing stream contains and whether it meets the needs of its next stage. A line can make useful separations and still leave contamination, missed targets, attached residues or inconsistent material that requires further handling.
Define what quality means for the destination
Quality is not one universal measure. The relevant criteria depend on the material, its intended use, receiving conditions and applicable requirements. A stream intended for additional processing may be assessed differently from one supplied directly to a user. Teams should establish what matters with the people responsible for handling and accepting the material rather than assuming that a visually uniform pile satisfies every need.
For a given stream, questions may include the presence of other materials, soil or fine residue; the range of piece sizes; visible coatings or attached components; and whether the material is suitable for the next operation. Some properties cannot be confirmed by appearance alone. An image may support a surface-level inspection, while further measurement, sampling or specialist assessment may be required for a particular specification.
Use checks that fit the process
Visual checks can help operators spot obvious misplaced items and changes in the stream. Sampling can provide a more structured view of composition, provided the method represents the material being produced. Operational records can track exceptions, machine settings, line conditions and corrective actions. No one check suits every fraction, and this guide does not prescribe a universal sampling rate or acceptance threshold.
Consistency matters when comparing observations. Record what was checked, how it was assessed and what happened next. If a team finds contamination, investigate whether it came from the incoming mix, a recognition error, an actuation issue, material movement, or a different process step. Avoid assigning cause based only on a single image or a final output sample. The purpose of monitoring is to improve understanding and support a measured response.
Input and output are not the same evidence
The input image documents what a camera could see before an action. The output inspection documents what ended up in a stream. These observations answer different questions. One can help assess the system's interpretation and intended action; the other can help assess the resulting material. Looking at both makes it easier to understand where the process worked and where a gap remains.
| Stage | Evidence to review | What it can tell you | What it cannot prove alone |
|---|---|---|---|
| Input view | Camera image, visibility, crowding, surface condition | What was exposed to the vision system | Hidden composition or final stream quality |
| Decision and action | Predicted label, exception, timing and equipment response | How the system attempted to direct the sort | That the correct material reached the intended destination |
| Output stream | Inspection, sampling and receiving feedback | Observed composition and potential process issues | Every property not covered by the selected check |
Respond to changes, not just failures
Incoming construction material can change as projects and handling practices change. A stream that previously produced predictable images may become more mixed, wetter, dirtier or more crowded. A shift in output observations can signal the need to inspect feed conditions, camera cleanliness, machinery, category definitions or the process itself. Treat a change as a prompt to investigate rather than assuming the model is the only possible cause.
Feedback from downstream users is valuable because it connects line measurements to actual handling. Ask what they received, what additional processing was needed and what problems occurred. Where requirements are formal, consult the applicable contract, specification and competent technical guidance. A sorting system cannot replace those requirements or establish suitability in isolation.
Close the loop
A practical quality routine links observed issues to decisions people can take. It might include recording nonconforming examples, pausing or adjusting a process under established procedures, reviewing the model's uncertain cases, checking equipment and discussing the stream with its receiver. Define responsibilities in advance so an operator knows who can authorise changes and where observations should be recorded.
Use evidence proportionately. Do not claim a recovery rate or guaranteed purity without a clear measurement method and supporting data. Report what was actually checked and acknowledge limits. For teams considering automation, a modest, transparent account of material variation is more useful than a broad promise that every recognised object will become a recovered product.
The full process is observation, interpretation, action and verification. Revisit the camera view, classification limits and physical separation to see how the stages connect.