AI for Jigsaw Sorting Experiments
As of 2026-10-19 · the AZET team. We write this educational guide as AZET's makers, for fun and reference; every example and result below is hypothetical.
Sorting a jigsaw can become a second puzzle. A piece showing sky, a roof, and a dark patch has three plausible destinations. Asking an assistant to name more categories does not resolve the practical question: which arrangement makes the next piece easier to find?
A useful AI jigsaw sorting experiment starts with a limited comparison. Give a text assistant your category rules and observations, then ask it to identify ambiguities and summarize the tradeoffs. Keep physical placement and the final choice with the person handling the pieces.
Choose the next task before choosing trays
Define what you will do after sorting. A method suited to finding lettering may be awkward for finding connector shapes.
For an illustrative session, suppose the next task is finding six pieces that match six supplied visual descriptions. The descriptions are written by a helper before sorting and remain unchanged. Each search ends when the helper checks a selected piece against its description. This prevents a quick but incorrect selection from counting as a success.
Set a space limit, such as four trays plus an undecided area. Count sorting time separately from search time. Also record how many pieces you move between trays and how many remain undecided. Keep these measures separate to preserve the tradeoffs.
Write rules that handle mixed pieces
Choose two methods to compare. In this hypothetical exercise, method A uses dominant visible color. Method B uses recognizable image details: lettering, straight architectural lines, foliage, and background texture. Both permit an undecided area.
Define a tie rule before starting. For A, a piece with no clearly dominant color goes to undecided. For B, lettering takes precedence over architectural lines, which take precedence over foliage. Background texture receives pieces that match none of those details. This ordering belongs to this exercise, not to jigsaw sorting in general.
Ask the assistant to examine written piece descriptions against the rules. An illustrative description, “white letter crossing a green leaf,” should land in lettering under B. If the answer depends on an unstated assumption, revise the rule or leave the piece undecided. Do not ask the model to infer unseen artwork.
A worked comparison with visible limits
Imagine using the same set of 60 loose pieces twice. The following invented record shows how a summary might look:
| Observation | Method A | Method B |
|---|---|---|
| Initial sorting time | 4 minutes | 7 minutes |
| Six completed searches | 9 minutes | 5 minutes |
| Pieces left undecided | 8 | 3 |
| Pieces moved during sorting | 6 | 2 |
The recorded sorting and search totals are 13 minutes for A and 12 for B. B takes longer to prepare but leaves fewer undecided pieces in this fictional session. That is a modest, task-specific difference, not evidence that detail sorting is universally faster.
There is also a serious comparison limit: B was attempted second. The sorter had already seen the pieces and search descriptions. Familiarity could affect the later result. An assistant should flag that limitation before suggesting a winner.
For another session, you could reverse the order. Alternatively, use two different batches and record that their difficulty may differ. Neither choice removes every source of variation. The goal is a useful personal note, not a scientific ranking of puzzle methods.
A prompt that preserves the comparison
Paste your actual observations after this prompt, keeping invented practice data clearly labelled:
Help me review a jigsaw sorting exercise. Use only my supplied tray rules, piece descriptions, timings, search checks, and attempt order. Calculate sorting plus search time separately for each method. Identify ambiguous rules and comparison limits. Do not infer missing measurements or recommend a universal best method. Suggest one small rule revision for my stated next task. Return observations, interpretation, and a human check as separate paragraphs.
Read the response for category drift. If the assistant replaces “recognizable lettering” with “interesting details,” it has changed the experiment. Restore the original terms before comparing records.
Check the trays, then keep a modest conclusion
A person should sample pieces from every tray and apply the written rule again. Record disagreements rather than quietly relocating pieces. A recurring disagreement can reveal that a category label needs a clearer boundary.
Check the arithmetic against the original notes, including whether pauses were counted. Then choose a conclusion with a defined scope: in the illustrative session, B suited the six specified searches, while A required less setup. You might keep A for a short evening and B for a longer search session.
FAQ
Must I time the whole puzzle?
No. Time the task you defined. Whole-puzzle completion includes many activities beyond sorting, making it harder to connect a result to tray rules. State plainly that your comparison covers only preparation and the selected searches.
What if almost everything goes into one tray?
Check whether that tray is still useful for the next task. If searching it feels cumbersome, divide it using an observable distinction and give the revised method a new label. Do not rewrite the old record as if the new rule had always applied.
Can an assistant sort from photographs?
This exercise assumes human-supplied descriptions. Any image-based suggestion would need separate checking against the actual pieces, especially where artwork or boundaries are unclear. A plausible category label does not establish that a piece belongs there.
The product boundary
Our E1 maker introduction provides a historical product baseline dated October 3, 2026: azet was described as an AI assistant with an early-access waitlist. It does not verify today's live status. This workflow is a general illustrative exercise, not a documented AZET capability.
For the AZET waitlist, visit azet.io.