Exact matrix multiplication, explained.
Compute exact matrix products, follow selected outputs, compare equivalent executions, inspect representations, and exchange reproducible local cases.
A reference workbench for learners, numerical researchers, and codec or integration developers.
- Choose a calculation and load a worked example. Edit A and B in row-major order, then compute the current exact result.
- Select any output cell. Explain its reduction, compare equivalent source executions, or inspect an input, codebook, stored symbol or output representation.
- Use the independent codec and IEEE tools for representation questions. Export a local case to keep its exact inputs, observation and implementation identity; import always validates and recomputes.
Loading the exact calculation engine… Your inputs remain editable.
Advanced protocol editor
Method, limits and guidance
How the method works
Decode the code: a stored symbol represents an integer, a codebook word, an IEEE value, or another declared scalar. The encoded symbol is not necessarily the value used in multiplication.
Accumulate exactly: products and reductions retain their declared integer or dyadic meaning. Finite IEEE products use canonical Laurent atoms and ternary contraction before output encoding.
Encode once: wrapping, saturation, or IEEE rounding applies at the terminal output. The Terminal integer encoding and IEEE examples make this distinction visible.
Four complete research journeys
Learning: load A first 2 × 2 product, compute, select its first output and Explain selected output. Follow 1 × 5 and 2 × 7 through exact prefixes 5 and 19, then inspect another cell to see its own arithmetic.
Numerical research: choose any of the 24 operations, load its example, and Compare equivalent executions. Both source commands and complete responses are available. Change the route or an input and run again; earlier research cannot stand in for the changed draft.
Codec and representation development: inspect B, a codebook entry or an output at a zero-based index and decoded coordinate. The independent codec tool explores all nine stages and their composition; the IEEE tool explains binary32 and binary64 fields, classifications, exact dyadics and canonical atoms without requiring a matrix calculation.
Integration and reproducibility: Export local case, review the exact source-owned JSON and Download. Import the file to validate the closed versioned schema and recompute. If its model identity differs, review both identities and explicitly Recompute with current model. The saved observation is compared with the new execution, including any changed answer.
Protocol & limits
Guided dimensions are 0 through 8; integer input magnitudes have at most 64 decimal digits. Dyadic input scales are 0 through 64. Empty dimensions are valid when their entry arrays are empty.
All 24 canonical families have guided forms. The advanced editor accepts the same version 1 tab-separated protocol; use canonical comma-separated arrays and no trailing newline.
The underlying calculation request retains its 16 KiB protocol limit. Research envelopes have a separate 2 MiB UTF-8 transport budget, complete reports are bounded at 4 MiB, and local case files at 1 MiB. The action budget includes JSON escaping and any submitted draft values; a file below its own limit can still exceed the encoded action budget. Oversized data fails explicitly; exact values remain strings.
Composed codec plans have at most eight stages and all nine stage kinds are inspectable. Stored rows use / separators. Select a stored symbol and every valid decoded coordinate to follow its actual composition path; report tables retain every row through pagination.
Ordinary computation has a 30-second deadline; explicit research and case actions have a 180-second deadline. Cancel, edits, selection changes, route changes and file changes retire stale replies. No saved observation is treated as a fresh answer.
Privacy & recovery
On GitHub Pages, calculations run in this browser. The standalone browser application does not send operands to a calculation service or save calculation history in browser storage. Loading the application and deliberately following documentation links use the network.
Case files are read only when you explicitly import a selected local file; exporting prepares source-owned bytes and Download saves them. GitHub Pages performs parsing, identity checks and recomputation locally. There is no automatic upload, history, browser storage or clipboard write. A downloaded case contains the operands and result, so share it deliberately.
Hologram deployments send calculation requests to their host intent endpoint. The host’s data and execution policies apply; cancellation retires the pending result in the interface.
Inputs stay available through cancellation, timeout, and recoverable loading or calculation failures. Retry reloads the computation engine; Reset restores the current calculation’s source-owned starting values.
Results show the current calculation rather than a saved history. Avoid entering sensitive data on a shared device; the current draft and result remain visible in the page until cleared or closed.
Source & verification
Matrix arithmetic, guided request handling, result interpretation, examples, and this interface are declared in LexLean. PrismPM generates the portable artifacts and the browser interface.
Release acceptance checks source policies, kernel evidence, independent arithmetic oracles, clean-root reproducibility, and deliberate defects. GitHub Pages publishes the accepted browser assets; source-free SDK and actual browser checks then inspect the live release.
Accessibility acceptance covers the declared finite semantic, keyboard, focus, contrast, target, and responsive checks with pinned browser oracles. It is not a claim of complete WCAG certification.
The Pages deployment provenance identifies the deployed model and generated browser artifacts. Compare those identities with the retained acceptance and publication observations when reviewing a release.