Glossary

Terms are defined once, in plain language, without repeating the term inside its own definition. If a word isn't listed here, it's either a standard ML/tensor term (see PyTorch's own vocabulary for those) or not yet part of Tensormorph's decided vocabulary.

Address space The full drill-down hierarchy from a whole model to a single scalar. See Model, Architecture, Layer and Module for the exact levels.

Commit pin A reference a loaded checkpoint carries back to the exact Lob commit and content hash it was materialized from. Used to detect when the underlying history has moved past what's currently open (a stale-note).

Compute panel The UI surface for choosing or overriding which backend (CPU, WebGPU, CUDA, ROCm, Metal) serves a given computation. Defaults to Compute: Auto.

Console The panel providing a full TQL REPL, for queries longer or more reusable than a quick filter.

Explicit analysis A tier of computation (SVD/PCA, effective rank, expert similarity) that reads substantially more of a tensor's data than a mergeable tile statistic does, and is therefore requested deliberately rather than computed automatically.

Fallback adapter The generic ModelAdapter used when no architecture-specific adapter recognizes a loaded model. Everything still opens and is inspectable; specialized views tied to a recognized architecture (like the MoE Expert Galaxy) aren't available.

Inspector The panel showing properties and statistics for the current selection, present in every view.

Lob The tensor-native version-control system Tensormorph anchors to via commit pins. Lob is a separate system, not bundled into Tensormorph, and its own branch/merge model is still being defined.

Model Adapter The per-architecture component that reads a loaded checkpoint into the Semantic Model Graph, recognizing that architecture's specific structure (layers, modules, MoE routing, multi-tower fusion points). See Fallback adapter for what happens without one.

Monoid-mergeable A property of tile statistics: a coarser region's statistic can be computed directly from its constituent finer regions' statistics, without re-reading raw values. The basis for level-of-detail streaming.

Outliner The tree view of a model's structure - the primary non-spatial way to navigate the address space, kept in sync with spatial selection in every 3D/2D view.

Problems panel Where flagged diagnostics accumulate across a model - NaN/Inf detections, statistical outliers, import warnings, backend fallbacks - navigable with F8.

RuntimeProbe The trait implemented by anything that can supply live runtime data to Tensormorph (a bundled local runner, a serving-framework plugin, or replay of a recorded .tmrun trace). Exposes capture points and breakpoints on the address space.

Semantic Model Graph (SMG) The typed tree Tensormorph builds from a loaded checkpoint, produced by a ModelAdapter. Every view, selection, and query operates over the SMG, not over the raw checkpoint file format directly.

Stale-note A flag Tensormorph raises when a loaded checkpoint's commit pin no longer matches the current state of the Lob history it points into.

Tile The unit of data Tensormorph streams, statistics-computes, and renders at - a bounded region of a tensor, sized appropriately for the current level of detail.

Tile statistics The set of precomputed, monoid-mergeable summaries (moments, extremes, norms, special-value counts, histogram) available for any tensor region without reading its raw values. See Tensor Statistics.

TQL (Tensor Query Language) Tensormorph's expression and query language, used for selection filters, breakpoint conditions, and saved queries. See Expressions and Queries.

View One of Tensormorph's rendering surfaces for a selection - 3D (Architecture, Layer Stack, Tensor Volume, MoE Expert Galaxy, Parameter Map, Diff, Morph, Runtime Flow), 2D (Matrix/Heatmap), or 1D (Distribution and related). Referred to as an "editor" in some page URLs for historical reasons, but "view" in product vocabulary.

Weight-space pull request A Lob-side reviewable unit representing a proposed weight change, loadable and diffable from within Tensormorph even though Tensormorph doesn't perform the underlying merge itself.

Workspace preset A named, full layout of views and panels suited to a task (Architecture, Compare, Morph, MoE, Runtime, and Performance are among the shipped presets), switchable as a unit.

Related resources

• Model, Architecture, Layer and Module - the address-space hierarchy referenced throughout this glossary.
• Tensor Statistics - full detail on tile statistics and mergeability.