GOVERNATLAS / PRODUCT DOSSIER

Captum

Captum documents a model interpretability library for PyTorch, including attribution methods for inputs and model components.

Explore the product ↗Add to comparison ↗

01 / CONTEXT & SCOPE

Understand the proposition.
Define the boundary.

Where it fits

Evaluate the attribution method together with the PyTorch model, target and baseline it actually uses.

The decision to make

Identify the prediction target, attribution method and baseline before comparing saliency views. Input, layer and neuron attribution expose different parts of a PyTorch model. A visually concentrated heatmap does not establish correctness or human meaningfulness, especially if preprocessing and the selected target are not retained.

What this record does not establish

The current official repository is meta-pytorch/captum; the older path redirects. Interpretability and perturbation capabilities do not establish causal truth, production robustness or universal model compatibility. No code was installed or run here.

1 current source · 4 published assertions
Source coverage is a research boundary, not a quality score.

02 / YOUR EVALUATION

Turn interest into a useful trial.

Open the evaluation playbook ↗

A test worth running

Use a toy synthetic network with a known output path in an authorized environment. Compare two justified baselines under the same method, retain the prediction and attributions and inspect a method-appropriate sanity check. Keep any adversarial perturbation experiment outside the initial explanation review.

GovernAtlas editorial proposal. We have not performed this test.

Evidence to bring back

A PyTorch/model/target/baseline/attribution-version record.
A retained attribution and baseline-sensitivity or sanity-check observation.

THE OPERATING DETAIL

Make the attribution experiment reconstructible.

GovernAtlas analysis: practical review considerations, not findings from a product test.

Pin model and target

Record the model checkpoint, preprocessing, evaluation mode and output target for a synthetic example. Retain the tensor shape and feature mapping. A convincing visualization cannot explain the deployed application if it was generated for a different output, transformed input or earlier checkpoint.

Challenge the baseline

For a method that requires a baseline, choose and justify one before inspecting the attribution. Compare a second defensible baseline while keeping the model and case fixed. Preserve both results and the numerical diagnostics appropriate to the selected method. Variation should remain visible in the interpretation.

Separate attribution and perturbation

If the exercise also generates an adversarial or minimal-input perturbation, retain its objective and constraints separately from the attribution. Ask whether the perturbed input is meaningful in the application’s domain. A changed prediction under an artificial perturbation does not establish a feasible real-world intervention.

03 / DOCUMENTED EVIDENCE

Read the claim. Inspect the basis.

Each statement retains its classification and preserved source. A vendor assertion is not an independent test result.

first party asserted

Its documentation describes input, neuron and layer attribution, as well as adversarial and minimal-input perturbation capabilities; these do not establish causal explanations.

Source, dates & preserved provenance

Captum — primary scope and current identity ↗

Authority: first party
Observed 2026-10-16 · Reviewed 2026-10-16 · current until 2026-11-15

Artifact SHA-256: aa1b58a5d4a327dcd0ef4b824c578af6cde5a8bc41f52c8e97c0f609a96d5ca6

The hash identifies the preserved capture. The live source may have changed.

first party asserted

The project describes concept- and training-example analysis using TCAV and TracIn influence functions.

Source, dates & preserved provenance

Captum — primary scope and current identity ↗

Authority: first party
Observed 2026-10-16 · Reviewed 2026-10-16 · current until 2026-11-15

Artifact SHA-256: aa1b58a5d4a327dcd0ef4b824c578af6cde5a8bc41f52c8e97c0f609a96d5ca6

The hash identifies the preserved capture. The live source may have changed.

first party asserted

Captum documents PyTorch interpretability methods including Integrated Gradients, saliency and noise-based attribution methods.

Source, dates & preserved provenance

Captum — primary scope and current identity ↗

Authority: first party
Observed 2026-10-16 · Reviewed 2026-10-16 · current until 2026-11-15

Artifact SHA-256: aa1b58a5d4a327dcd0ef4b824c578af6cde5a8bc41f52c8e97c0f609a96d5ca6

The hash identifies the preserved capture. The live source may have changed.

first party asserted

Captum documents a model interpretability library for PyTorch, including attribution methods for inputs and model components. Scope: primary documentation only; no independent model test, fairness determination, causal finding or compliance approval.

Source, dates & preserved provenance

Captum — primary scope and current identity ↗

Authority: first party
Observed 2026-10-16 · Reviewed 2026-10-16 · current until 2026-11-15

Artifact SHA-256: aa1b58a5d4a327dcd0ef4b824c578af6cde5a8bc41f52c8e97c0f609a96d5ca6

The hash identifies the preserved capture. The live source may have changed.

Source register

Captum — primary scope and current identity ↗first party · captured 2026-10-16

Brand asset provenance

Organization identity asset from its official website; inclusion does not imply endorsement.

Official website ↗ · Original asset ↗

04 / BEFORE YOU COMMIT

Better questions.
A clearer decision.

Resolve these with the provider and your own evaluation team.

Can you provide a PyTorch/model/target/baseline/attribution-version record for the selected configuration?

Can you provide a retained attribution and baseline-sensitivity or sanity-check observation for the selected configuration?

Which exact version, entitlements, data paths and responsible roles apply?

Regulatory applicability requires a separate, current determination. Inclusion is not certification.

COMPLIANCE / CLAIMS, SCOPE & GAPS

Start with the requirement.
Inspect the support.

Prepare your vendor review ↗

These labels describe this dossier’s current evidence coverage. A published mention may be a framework mapping, a qualified assertion or a limitation; it is not a compliance finding.