---
title: "A Single Class Choice Cannot Express Co-Occurring AI Tags"
description: "Leviathan Systems — Technical content Published 2026-10-21; updated 2026-10-21."
canonical: "https://www.leviathansystems.co/articles/ai-classifier-categorical-choice-multilabel-membership"
last-updated: "2026-10-21"
---

# A Single Class Choice Cannot Express Co-Occurring AI Tags

Leviathan Systems — Technical content
Published 2026-10-21; updated 2026-10-21.

A classifier interface must be able to express the label set the application requires. A return restricted to one winner among individual tags A, B and C cannot return A and B together. A declared multilabel interface with separate membership decisions can express that combination, and potentially no tags. Output width and agreement on the leading tag do not establish equivalent deliverables. Check the label meanings, score interpretation and final selection rule together before assessing prediction quality.

## Specify whether tags can coexist

Start with the complete result required for one input. A document may mention both delivery and installation. If its tags are A for delivery mentioned, B for installation mentioned and C for relocation mentioned, returning only its dominant topic loses a distinction that a request for all applicable tags needs.

The scikit-learn 1.7.2 guide distinguishes a single multiclass property from multilabel classification, where an item can receive zero through all available labels. In the latter task, each label has a binary membership. A sample-by-label indicator matrix retains those memberships in separate columns.

Throughout this comparison, A, B and C name individual tags. An interface returning exactly one of those three identifiers has only three possible answers. If every combination is permitted, three binary memberships have eight patterns, including no tags and all three tags. Adding a categorical none class would accommodate an empty set but still leave A and B together unrepresented.

A categorical design could instead give each permitted combination its own class, or a system could use structured decoding to return sets. Such designs need their own explicit label map and decoding contract. The limitation here belongs to the stated one-winner interface over individual A/B/C tags; it is no general prohibition on softmax within a multilabel system.

References: [S4](https://scikit-learn.org/1.7/modules/multiclass.html)

## Follow the score denominators

PyTorch 2.14 defines softmax along a specified dimension. For this one input, hold the A, B, C class positions fixed. A logit z(c) is a supplied real score before transformation. Class-axis softmax gives p(c) = exp(z(c)) / [exp(z(A)) + exp(z(B)) + exp(z(C))]. The three entries share one denominator and sum to one.

PyTorch 2.14 defines sigmoid elementwise: s(c) = 1 / [1 + exp(−z(c))]. Each entry uses its own logit, without a denominator pooling the other labels. Several resulting scores can therefore be large together, and their sum need not be one.

These formulas explain numerical constraints. The task, trained model and declared interface determine what the scores mean to the application. Applying sigmoid to an arbitrary score vector does not by itself establish meaningful membership probabilities.

The related softmax-axis guide addresses which entries belong in a normalization denominator. Here the input and class dimension are already fixed; the question is whether the task calls for one categorical choice or several memberships.

References: [S1](https://docs.pytorch.org/docs/2.14/generated/torch.nn.Softmax.html), [S2](https://docs.pytorch.org/docs/2.14/generated/torch.nn.Sigmoid.html)

## TAG-42: derive both complete score vectors

Fictional case TAG-42 uses one document with the stipulated reference memberships A and B present, C absent. Its required set is {A, B}. Supply the fixed dimensionless logits z = (ln 4, ln 3, −ln 3) in A, B, C order, where ln denotes the natural logarithm. Their exponentials are (4, 3, 1/3).

For sigmoid, the equivalent expression exp(z) / [1 + exp(z)] gives s(A) = 4 / (1 + 4) = 4/5 and s(B) = 3 / (1 + 3) = 3/4. For C, s(C) = (1/3) / (1 + 1/3) = 1/4. The complete vector is (4/5, 3/4, 1/4), or (0.8, 0.75, 0.25). Its sum is 9/5.

For class-axis softmax, the shared denominator is 4 + 3 + 1/3 = 22/3. Dividing each exponential by that value gives p(A) = 6/11, p(B) = 9/22 and p(C) = 1/22. The complete vector sums to 12/22 + 9/22 + 1/22 = 1.

Both calculations use the same invented logits to isolate their mathematical meanings. They describe no executed classifier, trained-model comparison or measured outcome. Replacing an existing model's activation is a separate design change whose suitability this arithmetic cannot establish.

References: [S1](https://docs.pytorch.org/docs/2.14/generated/torch.nn.Softmax.html), [S2](https://docs.pytorch.org/docs/2.14/generated/torch.nn.Sigmoid.html)

## Read the final selection as well as the leading tag

Declare two illustrative returns. Offer M supplies separate label scores and selects every tag whose score is at least 7/10. This cutoff is stipulated solely for TAG-42. Offer C supplies the categorical vector and returns only its argmax tag: the position with the largest score.

Under Offer M, A's 4/5 and B's 3/4 both meet 7/10; C's 1/4 does not. Its selected set is {A, B}, represented in the declared order by (1, 1, 0). Offer C returns A, because 6/11 exceeds both 9/22 and 1/22. Its single winner cannot express the stipulated two-tag set.

A hypothetical application of the same 7/10 cutoff to the categorical vector would select nothing: even its largest entry, 6/11, is below the cutoff. That hypothetical selection differs from Offer C's promised argmax return. It also shows why carrying a cutoff unchanged between these different score meanings leaves the intended decision unsupported.

A ranks first in both complete vectors. A comparison recording only that agreement would miss B's selected membership in Offer M. If an illustrative downstream consumer keeps one review entry per selected tag, M supplies A and B entries while C supplies only A. The difference arises before any assessment of how well a real classifier identifies the document's contents.

Offer M can express the required set and does so under this supplied example. An interface that computes separate scores but then exports only its argmax would again deliver one tag. The final consumed return, including its selection rule, determines which memberships survive.

## Interpret targets and empty results by their declared role

The reference target (1, 1, 0) means that A and B apply and C does not in this multilabel task. Its sum of two is the count of positive memberships. It is not a unit-sum categorical probability target.

PyTorch 2.14's BCEWithLogitsLoss documents a formulation for multilabel binary targets, combines sigmoid with binary cross entropy, and requires targets shaped like the inputs with values between zero and one. This is a named example of the per-label target role. It supplies no recommendation to change a loss or training setup for an existing application.

The related cross-entropy guide concerns categorical target representations and weighted mean losses. Its probability-target requirements answer that different question. The category-feature guide likewise examines zeros in encoded inputs. A vector's dimensions and zeros need its role attached before either guide's interpretation can be applied.

For the present binary membership convention, a fully known target (0, 0, 0) can validly mean that none of the three tags applies. A predicted empty set means that the declared selection chose no tags; its correctness remains to be assessed. Unknown membership, abstention and an absent result need separately specified meanings, so a consumer can distinguish them from a completed no-tag prediction.

References: [S3](https://docs.pytorch.org/docs/2.14/generated/torch.nn.BCEWithLogitsLoss.html), [S4](https://scikit-learn.org/1.7/modules/multiclass.html)

## Separate membership scores from joint claims

Computing each sigmoid separately imposes no sum-to-one competition across the three positions. It does not establish that the underlying labels are statistically independent. The scikit-learn guide explicitly includes multilabel methods that account for correlations between labels.

Consequently, the two selected memberships in TAG-42 supply no measured co-occurrence probability. The score vector alone does not justify multiplying A's and B's values and calling the result the probability that both apply. Whether scores support a probability interpretation, and whether predictions are correct, require evidence beyond their activation formula.

References: [S2](https://docs.pytorch.org/docs/2.14/generated/torch.nn.Sigmoid.html), [S4](https://scikit-learn.org/1.7/modules/multiclass.html)

## Compare the delivered interface against the required set

A useful supplier response identifies the model and label-map revisions, the input item associated with the result, and the meaning and order of every output position. Establish whether labels are individual properties or composite classes, which combinations are permitted, and how no applicable labels, unknowns and unavailable results are represented.

Then follow the return into the consuming application. Identify whether the supplied values are logits, categorical scores, per-label scores or already selected memberships. Ask which exact rule produces the final set, including any cutoff, tie handling, forced winner or cap on the number of returned labels. A membership-capable score array can still feed an interface that discards required memberships.

For TAG-42, a promised single A/B/C winner fails the representation requirement for {A, B}. Offer M's declared scores and selection produce that set. This resolves the example's interface decision; deciding whether an actual delivered classifier performs the task well needs applicable prediction evidence.

Once the positive decision for each label is defined, the related alert-precision guide explains why a performance headline needs its reference population and counts. Output semantics come first: neither a three-position shape nor a matching top tag settles what was predicted. The appointed application owner should assess the complete declared return against the required label sets before relying on quality claims.

## Discuss a separately needed physical installation

If the wider project independently requires selected equipment to be mounted and physically connected, Leviathan's Rack-and-Stack & Colocation Installation service covers agreed equipment installation, IT cabling, labels and physical handoff records. Use the related service and project inquiry with a nonconfidential equipment list, rack layout, site and required installation records to discuss scope and availability.

The appointed application team retains classifier design, software integration and prediction acceptance. The label-set comparison establishes no additional equipment requirement.

References: [L1](https://www.leviathansystems.co/services/rack-and-stack)

## Primary sources

- [S1 — PyTorch 2.14 — Softmax](https://docs.pytorch.org/docs/2.14/generated/torch.nn.Softmax.html). Checked 2026-10-21. Defines shared-denominator normalization and unit-sum slices along the specified dimension. TAG-42 fixes one input and its three individual-tag positions.
- [S2 — PyTorch 2.14 — Sigmoid](https://docs.pytorch.org/docs/2.14/generated/torch.nn.Sigmoid.html). Checked 2026-10-21. Defines the elementwise transform used in TAG-42. The formula imposes no cross-label unit-sum constraint and supplies no calibration or independence guarantee.
- [S3 — PyTorch 2.14 — BCEWithLogitsLoss](https://docs.pytorch.org/docs/2.14/generated/torch.nn.BCEWithLogitsLoss.html). Checked 2026-10-21. Documents combined sigmoid and binary cross entropy, a multilabel formulation, matching input/target shapes and targets between zero and one. Used to clarify target roles.
- [S4 — scikit-learn 1.7.2 — Multiclass and multioutput algorithms](https://scikit-learn.org/1.7/modules/multiclass.html). Checked 2026-10-21. Distinguishes one multiclass property from nonexclusive label memberships, including zero-label samples, binary indicator targets and methods accounting for label correlations.
- [L1 — Leviathan Systems — Rack-and-Stack & Colocation Installation](https://www.leviathansystems.co/services/rack-and-stack). Checked 2026-10-21. Defines agreed physical equipment installation, IT cabling, labels and installation records for an equipment project.

## Related Leviathan resources

- [Softmax axes and the normalized score population](/articles/ai-softmax-class-batch-axis-normalization)
- [Cross-entropy target representations and weighted mean losses](/articles/ai-cross-entropy-target-representation-weighted-mean)
- [Alert precision, positive definitions and class prevalence](/articles/ai-alert-precision-positive-class-prevalence)
- [Categorical input features: dropped baseline and unknown values](/articles/ai-onehot-unknown-dropped-baseline-feature-meaning)
- [Rack-and-stack physical installation scope](/services/rack-and-stack)
- [Discuss a separately needed physical equipment installation](/contact?service=rack-and-stack&from=%2Farticles%2Fai-classifier-categorical-choice-multilabel-membership)

## Does selected equipment need physical installation?

Share a nonconfidential equipment list, rack layout, site and required IT connections to discuss physical installation scope and availability.

[Discuss the scope](/contact?service=rack-and-stack&from=/articles/ai-classifier-categorical-choice-multilabel-membership)
