Handheld sensing unit / separate device
A battery-powered unit reads 19 optical channels and answers with every label it can support, each one carrying its own score and its own threshold. Its catalog lives on a memory card. A surface it has never seen can be taught in the field in seconds, with no host computer and no retraining. Everything it already knew reads the same afterwards.
wrong names in 1,219 gate reads
0
Run mlc_v13_s21b, the gate that had to pass before the catalog was pushed to the card. 4,312 label assertions were made across those reads and 346 labels were considered and left unresolved. A run ships only at zero wrong names.
resident on the device
126,264
bytes, sized by stat on the shipped file. The same size to the byte at every deployed catalog from 37 labels to 50 at the largest, and a field write never modifies it.
labels answered per reading
3
on the surface read below, held on all 181 reads of a 20 second census.
Reading 1 / one surface, several true answers
The unit was held against a knotty pine interior wall. It considered 8 labels and asserted three of them. The three it declined stay on screen with their numbers. Dropping them would look like they were never considered.
| label | score | threshold | outcome |
|---|---|---|---|
| wood | >=0.5246 | 0.0460 | asserted |
| pine | 0.0982 | 0.0460 | asserted |
| knotty | 0.0982 | 0.0420 | asserted |
| cherry | -0.0982 | 0.0558 | unresolved |
| shellac | -0.0982 | 0.0558 | unresolved |
| smooth | -0.0982 | 0.2125 | unresolved |
Record / the gate the deployed catalog passed
1,219 reads across 36 recorded surfaces. Each read can assert several labels, so the counts below add up to more than the read count. Four labels were considered on every applicable read and asserted on none of them; they are in the chart at zero.
held out and re-read
0.000
Six labels were removed one at a time and their surface was re-read. The absent label was named on none of those reads. 6 labels were tested this way.
evidence spanning surfaces
25 / 49
labels have been recorded on more than one surface. A label recorded on a single surface may be a stand-in for that surface, so the surface count is published beside every label.
where the unresolved mass sits
231 / 346
of the unresolved readings belong to five labels: the dark-polymer family, which sits close together on this front end, and one squash species whose evidence is still thin. The unit declines rather than guessing in both cases.
Structure / one table, read two ways
The catalog stores recordings of surfaces and the labels each recording carries. Nothing else defines a label: its meaning is the set of recordings that carry it, so a label is general when many diverse recordings carry it and granular when few do. The two cards below read the same table from its two ends, and they meet on one label.
one recording, several labels
cherry_shellac
A shellacked cherry desktop. The catalog holds no combined entry by that name: the recording carries 4 labels, and the unit scores each one on its own evidence and its own threshold.
one label, several recordings
shellac
Two different woods under the same finish. The finish is the part the two recordings share, so the finish is what this label holds evidence for. Each wood species is carried by its own label on the same recordings.
This is what keeps the catalog small as it grows. A new shellacked surface adds one recording, the shellac label gains evidence, and no combined name is ever minted for the pair.
the same labels, arranged by a curator
The tree below is a curator's arrangement of the same labels, carried as data beside the catalog. The identification path never reads it, and the labels outside any family prove the arrangement is optional. Every count is the same surface count published in the record above.
outside any family
the garden, recording by recording
Ten recordings of five living plants, against the ten labels they share. A filled cell means that recording carries that label, so a label's evidence is the union of its column: general labels have long columns, granular ones short, and both kinds sit over the same recordings.
| living | leaf | leaf_top | leaf_bottom | squash | tomato | chili_pepper | zucchini | butternut | buttercup | |
|---|---|---|---|---|---|---|---|---|---|---|
| tomato_leaf_top | carries living | carries leaf | carries leaf_top | carries tomato | ||||||
| tomato_leaf_bot | carries living | carries leaf | carries leaf_bottom | carries tomato | ||||||
| chili_pepper_leaf_top | carries living | carries leaf | carries leaf_top | carries chili_pepper | ||||||
| chili_pepper_leaf_bot | carries living | carries leaf | carries leaf_bottom | carries chili_pepper | ||||||
| zucchini_leaf_top | carries living | carries leaf | carries leaf_top | carries squash | carries zucchini | |||||
| zucchini_leaf_bot | carries living | carries leaf | carries leaf_bottom | carries squash | carries zucchini | |||||
| butternut_leaf_top | carries living | carries leaf | carries leaf_top | carries squash | carries butternut | |||||
| butternut_leaf_bot | carries living | carries leaf | carries leaf_bottom | carries squash | carries butternut | |||||
| buttercup_leaf_top | carries living | carries leaf | carries leaf_top | carries squash | carries buttercup | |||||
| buttercup_leaf_bot | carries living | carries leaf | carries leaf_bottom | carries squash | carries buttercup | |||||
| recordings carrying it | 10 | 10 | 5 | 5 | 6 | 2 | 2 | 2 | 2 | 2 |
Three things this matrix states plainly. The living and leaf columns are identical today, which is why living is reported as unproven: nothing yet distinguishes them. The leaf_top and leaf_bottom columns split the same ten recordings in half, and that split is what let the instrument separate the two by measurement. And each species column pairs with squash only where the plant is a squash, which is how one reading can assert squash while leaving the species unresolved.
The field loop / what an operator does
worked
2026-08-11
A new surface taught in the field
A blue shop towel read as unknown on 34 consecutive reads. It was recorded, checked, and added on the device in about two seconds with no host and no retraining. It then read as itself on 7 of 10 reads, at scores of 0.426 to 0.462 against a threshold of 0.412, and every other surface still read as itself.
The predicted score was computed before the write and matched the measured score to two decimals.
declined
2026-08-11
A surface it declined to learn
A wax-coated paper plate was checked before writing and refused: it sat 0.898 from white polystyrene foam, close enough that adding it would have taken the neighbor down with it. It was added by rebuilding the catalog on the host instead, which separated the pair to 0.423, after which both read as themselves.
The check is the reason nothing was corrupted. The rebuild is a host step and it is not free.
limit
2026-08-11
A limit set by the sensor
Matte white paper and scuffed white plastic separate by 0.146 on 19 optical channels. The host predicted 0.144 before the session and the device measured 0.146. Both stay unresolved against each other, so the instrument says nothing.
The limit sits in the front end. More data does not move it. A different illumination band would.
worked
2026-08-18
Adding a label changes nothing that was already right
Across the campaign the catalog was rebuilt and re-pushed repeatedly, growing and once shrinking, and every deployed build from 37 labels through 50 has passed its gate with no wrong name, most recently across 1,219 reads on the 49-label catalog deployed today.
Labels that were thin at one build recovered at the next after more recordings were added. Thin labels heal by data, never by lowering a threshold.
worked
2026-08-13
It says squash when it cannot say which squash
Three squash species from a garden went into the catalog by leaf. On the deployed build the shared squash label asserts on 212 of 216 applicable reads, while the individual species assert less and decline the rest: buttercup 57 of 108, zucchini 63 of 84. The instrument reports the level the evidence supports and leaves the rest unresolved.
The species labels stay thin until more recordings of each plant exist. Their surface counts are published beside them.
worked
2026-08-18
Two labels that were identical until the data split them
Until this week every leaf in the catalog was recorded from the top, so leaf and leaf_top were carried by exactly the same recordings and nothing could tell them apart. A round of underside recordings later, with no change to the device or its software, leaf_top asserts on all 204 top-side reads with zero wrong assertions across all 120 underside reads, and leaf_bottom does the reverse.
The living label still spans exactly the same recordings as leaf, and the page treats it as unproven until a dried leaf is recorded.
Cost / resident memory as the catalog grows
The deployed unit holds 126,264 bytes resident and reads its 49-label catalog from the card. The blue curve is a simulation, 2026-07-13: a conventional classifier was fitted at four catalog sizes and its stored size recorded. It reports how large that method gets. It says nothing about accuracy.
| materials | conventional classifier | this unit, resident |
|---|---|---|
| 30 | 0.211 MB | 0.120 MB |
| 300 | 2.209 MB | 0.120 MB |
| 1,000 | 3.195 MB | 0.120 MB |
| 3,000 | 5.514 MB | 0.120 MB |
what the catalog costs on the card
Projected card sizes for the shipped configuration, computed from the per-entry size. The largest catalog written to a card so far held 50 labels; the one deployed today holds 49. The resident file has been the same size to the byte across every one of those builds, and a field write never modifies it, while the card file moved from 26,446 bytes at 37 labels to 39,946 today.
The same limits are printed on the device where the operator can read them.