# basis1

> The Excel engine for AI agents. Let AI agents read, check and edit Excel models.
>
> Basis re-runs every formula in an .xlsx against the values Excel itself cached
> in the file, and reports what diverges — plus the overtyped constants, broken
> references, circular dependencies and stale external links it finds on the way.

## Properties

- https://basis1.tech — this site: what the audit finds, and reference pages on
  Excel's actual calculation and file-format behaviour.
- https://basis1.tech/app — the spreadsheet. Opens .xlsx/.xlsm, runs the audit,
  keeps per-cell history with branching and merge. Requires an account to save.
- https://api.basis1.tech — the parsing API: a typed cell graph, a formula
  dependency edge list, cell-anchored images and a verification certificate.
  Start at https://api.basis1.tech/llms.txt for the machine-readable reference.

## What the audit reports

- Hardcoded plugs: a constant typed over a formula in an otherwise-uniform
  fill-down. Found by bytecode-program identity, not by heuristic.
- Broken references: #REF! cells and the dependents that inherited the error.
- Circularity: cycle membership, and whether the workbook's own iterative-calc
  settings sanction it.
- Staleness: volatile functions and links to workbooks you do not have.
- Verification: every formula recomputed and compared against Excel's cache,
  with divergences reported by coordinate.

## Measured

Across two public corpora (Enron spreadsheets; FUSE, from Common Crawl),
4,207 formula-bearing workbooks:

- 43.2% of substantial models (>=1,000 formulas) contain at least one hardcoded
  plug; 29.9% in the FUSE half, 43.2% in the Enron half.
- 97.31% of the 3,344 FUSE .xlsx workbooks under 1 MB recompute with zero
  mismatches against Excel's cached values (a workbook-level rate), over
  4,033,846 formulas compared.

## Formats

Reads .xlsx and .xlsm. Legacy .xls, .xlsb and .ods are recognised but not read
in the browser; the API converts them server-side.