GridCraft Corpus

Real-file test oracles for GridCraft: small workbooks made in Microsoft Excel by a committed script. Each file carries Excel's own result next to every formula (the cached <v> beside <f>), and GridCraft's check recalculates the file and compares, cell by cell. Every file is generated from scratch; nothing comes from a real workbook.

Contents: What this is · Layout · Checking with GridCraft · Verifying · Adding or regenerating files · excel/ · Metadata and privacy · artcraft-fixtures · Provenance · License

What this is

A workbook saved by Excel stores the value Excel computed for each formula. That cached value is the oracle: GridCraft opens the file, recalculates every formula from scratch and reports each cell whose result differs. The files are kept outside the GridCraft repository so its history stays small, and a consumer should fetch them at a pinned commit and check them against SHA256SUMS, so results can't change without a reviewed pin bump.

Every file covers one feature. Its path names the failure class and the file names the shape and its parameters (power-negative/caret-odd-roots.xlsx), so a failing file says what broke.

Layout

One top-level folder per authoring application:

Folder	Source	Files
excel/	Microsoft Excel for Microsoft 365 (16.0, build 20430, en-US), driven by tools/excel-oracles/	107

Also in the repository:

• tools/excel-oracles/: the generator (generate.ps1), the case list (cases.ps1) and the metadata check (verify-metadata.py).
• SHA256SUMS: one sha256 path line for every corpus file.
• LICENSE-MIT, LICENSE-APACHE, NOTICE: the dual licence.

Third-party corpora are not copied here. This repository holds only files we generate.

Checking with GridCraft

From a GridCraft checkout, build the CLI and point it at the folder:

cargo build --release -p gridcraft-cli
target/release/gridcraft-cli check ../gridcraft-corpus/excel --json
# check collects only .xlsx/.xlsm from a folder; name the .xls files explicitly:
target/release/gridcraft-cli check ../gridcraft-corpus/excel/legacy-xls/*.xls ../gridcraft-corpus/excel/precision-displayed/*.xls --json

check reports calc mismatches (expected = Excel's cached value) and a round-trip comparison (GridCraft's saved copy against the source). Its default tolerance is 1e-9, relative above magnitude 1 and absolute below, so a raw residue such as 1.2e-10 where Excel shows 0 passes on its own and shows up only in the cells downstream (see near-zero/).

The upward chains in dependency-chain/ overflow the stack of GridCraft 0.6.1 (main at 84acf62), which ends the whole check run with no output. Until that is fixed, check that folder file by file, or leave it out.

Verifying

sha256sum -c SHA256SUMS          # Git Bash or Linux; on macOS: shasum -a 256 -c SHA256SUMS

Adding or regenerating files

1. Only generator-made files. Every file is written by tools/excel-oracles/generate.ps1 from a case in cases.ps1. No hand-made, downloaded or found workbooks, and nothing from anyone's disk (see AGENTS.md).

2. Add a case to cases.ps1: a path under excel/, a list of cells (address, number, text or formula, optional number format) and the workbook settings (iteration, precision as displayed, .xlsx and/or .xls). A build step can add what cells can't (a PivotTable, a chart, a deleted sheet).

3. Run the generator on Windows with Excel installed:

   powershell -NoProfile -ExecutionPolicy Bypass -File tools/excel-oracles/generate.ps1               # every missing file
   powershell -NoProfile -ExecutionPolicy Bypass -File tools/excel-oracles/generate.ps1 '^date-text/' -Force   # a subset, regenerated

   The argument is a regular expression on the path (near-zero/sub-reordered-parts-small.xlsx). It prints one line per file: ok (generated), skip (already there; -Force regenerates) or FAIL. At the end it runs verify-metadata.py over excel/; any finding fails the run. Set GC_PYTHON if python isn't on PATH.

4. Refresh the manifest in the same commit:

   find excel -type f \( -name '*.xlsx' -o -name '*.xls' \) | LC_ALL=C sort | xargs sha256sum --text > SHA256SUMS

5. One commit per batch, naming what changed and the Excel version and build.

Excel never writes the same bytes twice (timestamps, document ids), which is why existing files are skipped by default: regenerate a file only when its case changes. Never rewrite this repository's history once it is published; consumers pin commits.

excel/: Excel oracle workbooks

107 files (101 .xlsx, 6 .xls), 1.53 MB in total. The cases are synthetic shapes of failure classes found by running GridCraft over real-world workbooks; only the shapes were kept. Each file with a handful of formulas also has a text listing (C1: =A1-B1) two columns right of its used range, so it explains itself when opened.

Directory	Files	What it exercises	Excel's answer
near-zero/	7	=A1-B1 where A1 and B1 are the same three parts summed in a different order (0.1+0.2+0.3, 1000000.1+0.7+0.2, and 300000000.1+0.1+0.1, whose 5.96e-8 residue is above check's tolerance); SUM variants; the same difference in parentheses, inside IF(...>0) and as a divisor; one copy with iteration on	The final - of a formula snaps to 0, so =A1-B1 is 0 and =1/C1 is #DIV/0!. Inside parentheses, a comparison or SUM(A1-B1) the residue (1.16e-10) survives, so IF(A1-B1>0,…) is TRUE
power-negative/	4	^ and POWER with a negative base: odd roots ((-8)^(1/3), (-32)^(1/5), (-8)^0.2, negative exponents, cell operands), even roots and 14-digit exponents that stay #NUM!, and the exponent limit (1/2147483645 works, 1/2147483647 and 1/1E15 don't)	(-8)^(1/3) = -1.9999999999999998
criteria-ref-error/	10	#REF! in a range argument of SUMIF, SUMIFS, COUNTIF, COUNTIFS, AVERAGEIF, AVERAGEIFS, MAXIFS, MINIFS; a #REF! criterion as a control; the realistic origin, a deleted sheet	#REF! for every range position; a #REF! criterion gives 0
date-text/	11	Date text with month names and without separators: "1Jan", "Jan1", "1 Jan", "Jan 2021", "1Jan21", "15Mar2024", MONTH(1&"Feb"), DAY("15Mar"); numbers read as two-digit years ("Jan 32", "Feb 29"), also with - and /; -// day forms ("Jan-21", "Jan/1"); the two-digit-year window ("1Jan29", "1Jan30"); month prefixes longer than three letters ("Sept", "Janu", "Decemb"); punctuated forms that stay #VALUE! ("4.Mar.2026", "1Jan, 2021", "Jan 15 2021", "001Jan", "Ma") and those that parse ("Mar 4, 2026")	"Jan-21" is 21 January of the current year, not 2021; "1Jan29" is 2029 and "1Jan30" 1930. See the year note below
precision-displayed/	16	Precision as displayed on: =-130+657.91 formatted 0, =A1*10, =A1/7 formatted 0.0000, =A3*7, =-1+2.91, in .xlsx and .xls; the same cells with the option off as a control. One file per format family (percent, currency, 0.0,, scaling, fractions, scientific, 0.00;(0.0) sections, literal text, [<1]0.00;0), with General, date/time and @ as controls; a cross-cell read; a spill landing in cells formatted 0, 0.0, 0.00	528, 5280, 75.4286, 528.0002. Every number format rounds to what it shows (0.0,, stores 1234567.891 as 1200000, ?/8 stores 0.3 as 0.25); dates, times, General and @ keep the full value. Each spilled element is rounded by its own cell's format
legacy-xls/	10	Legacy (pre-dynamic-array) formulas with a range in one cell, =Sheet1!A1:G1 and =Sheet1!A1:A7; a range inside an operator inside SUM, =MAX(SUM(A1:A3/12),0); a one-cell Ctrl+Shift+Enter array; a formula filled down (a shared formula). Each saved as Excel 97-2003 .xls and, as a twin, .xlsx	Implicit intersection: the cell in the formula's column (row); #VALUE! outside the range. =MAX(SUM(A1:A3/12),0) is 2 in row 2 and 6 as an array formula
circular/	16	A1 =SUMIFS(B1:C1,B2:C2,"x") with C1 =A1*2: C1 not selected by the criteria (not circular), selected (truly circular, iteration off and on), in the first or a later criteria range, whole columns, size-mismatched ranges, an array criterion that does or doesn't select C1; the SUMIF, AVERAGEIFS, MAXIFS, MINIFS and COUNTIFS forms; INDEX as a control	Not circular unless the criteria select C1, or C1 is in the first criteria range (SUMIFS and COUNTIFS); size-mismatched ranges are #VALUE!, not circular. The truly circular files cache C1 = 0
dependency-chain/	3	=A{r+1}+1 down 5,000 rows (each cell reads the row below), a 4,000-link chain hopping between two sheets, and the downward chain as a control. Excel stores each column as a shared formula	5000 and 4000 at the top
rate/	2	RATE(n,0,pv,0) and RATE(n,0,pv,tiny fv)	RATE(5,0,-100,0) = -0.9999994091451825; RATE(1,0,-100,0) and RATE(5,0,0,0) are #NUM!
round-trip/	3	A PivotTable from an in-sheet range (row field, sum of values) with GETPIVOTDATA, a clustered column chart over a range, and a two-cell workbook whose author fields are blank but whose docProps/core.xml carries the created and modified dates	For round-trip loss; their formulas are a light calc check
pivot-external/	4	PivotTables whose source is in a second workbook, source-data.xlsx (committed next to them): a range, a defined name in that workbook, and two PivotTables sharing one cache. The link is saved relative (Target="source-data.xlsx"), so it resolves inside the repository	GETPIVOTDATA over them gives 58, 392, 156 (and 107 for the second PivotTable)
getpivotdata/	12	GETPIVOTDATA against an in-sheet PivotTable with two row fields: grand total, one and two field/item pairs, the data field by caption, mixed case, a cell inside the PivotTable as the reference, an item from a cell; a missing item, an inner field alone and an unknown data field; SUMIFS over the same rows as a control; ten more shapes: row and column fields, report filters, totals off, dates and numbers as items, renamed captions, hidden items, several data fields, a stale report (source edited, PivotTable not refreshed) and a tabular PivotTable on another sheet	392, 107, 72; #REF! for the missing item, the inner field alone and the unknown field
pivot-data-model/	3	PivotTables whose cache is a workbook connection or the Data Model, not a worksheet range: a worksheet table (Sales) loaded into the model; an OLE DB connection to a workbook the generator writes into its staging folder (not in the model); the same OLE DB connection loaded into the model. Each has GETPIVOTDATA formulas next to it	cacheSource type="external" with a connectionId and xl/connections.xml; the two model files also carry xl/model/item.data. GETPIVOTDATA gives 392, 107, 166 (model forms use [Measures].[Sum of Amount] and [Sales].[Region].&[north]; the plain "Sum of Amount" and a bare item are #REF! on the model)

Locale. Excel ran as en-US (country code 1, decimal ., month-day-year). Date-text parsing depends on it, and on Windows' two-digit-year window (1930-2029): "Feb 29" is February 2029 and "Feb 30" February 1930, not a day in the current year.

Year. Date text without a year takes the year the calculation runs in. Only date-text/month-name-no-year-y2026.xlsx and date-text/month-name-day-this-year-y2026.xlsx depend on it: their cached serials are right during 2026. From 2027 on, a correct engine will mismatch every cell in those files; regenerate them with the new year in their names. MONTH/DAY extractions don't depend on the year.

RATE. Excel's RATE is an iterative solver, so its answer for a degenerate case (pmt 0, fv 0) is wherever its iteration stopped, about -0.9999994. A correct-in-spirit implementation may not match to 1e-9.

Pivot sources. The pivot-external/ files read their source from source-data.xlsx in the same folder. Keep the four together; regenerating a pivot file needs source-data.xlsx to exist first (the generator copies it into its staging folder and opens it from there, which is what makes Excel write the link relative).

Iteration. circular/sumifs-selected-cell-iterate-on.xlsx and near-zero/downstream-if-positive-iterate-on.xlsx are saved with iterative calculation on (100 iterations, max change 0.001). The circular one converges towards 10 and stops where Excel's iteration stopped, so a near match is the expected result there.

Files added in the second batch

File	Reproduces
pivot-external/source-data.xlsx	The source workbook: 8 rows of Region, Category, Amount and a workbook-level name SourceRows over them. No formulas
pivot-external/range-source.xlsx	PivotTable whose cache reads [source-data.xlsx]Sheet1!A1:C9; GETPIVOTDATA against it
pivot-external/defined-name-source.xlsx	The same PivotTable with the cache reading the defined name source-data.xlsx!SourceRows
pivot-external/shared-cache.xlsx	Two PivotTables (by Category, by Region) on one external cache
getpivotdata/field-item-forms.xlsx	Ten GETPIVOTDATA forms, six with values and four that are #REF!
getpivotdata/sumifs-control.xlsx	Control: the same PivotTable, totals from SUM/SUMIFS only
getpivotdata/row-col-fields.xlsx	Region by Category cross-tab: grand total, one and two items in either order, an empty intersection (0), a cell inside the data area as the reference; a reference off the PivotTable and an unknown item are #REF!
getpivotdata/report-filter.xlsx	A report filter showing one item, one showing (All), and one with a single item hidden: the totals follow what is shown, and items filtered out are #REF!
getpivotdata/totals-off.xlsx	Subtotals and grand totals turned off: #REF! for the grand total and for Region alone, values for Region with Category
getpivotdata/date-number-items.xlsx	Dates and numbers as items: DATE(), a serial and the text 3/10/2026 give values, 2026-01-15 and a date with no data are #REF!; the numbers 10 and "20" give values, 10.5 is #REF!
getpivotdata/captions.xlsx	A renamed data field, field and item: the caption (Area, North Zone) is what Excel accepts for the field and the item; the data field answers to its caption and to its source name, and the source field and item names are #REF!
getpivotdata/hidden-items.xlsx	Items hidden by a filter (a Region and a Category): a hidden item is #REF! and the grand total covers what is shown
getpivotdata/multi-data-fields.xlsx	Four data fields (sum, count, average, % of total), each asked for by its caption; the plain source name answers with the first
getpivotdata/stale-report.xlsx	Source edited and the PivotTable not refreshed: GETPIVOTDATA returns the displayed (old) value, SUMIFS the new one
getpivotdata/stale-report-sorted.xlsx	The same with the source sorted by Region, which separates a stale report from an item-order difference
getpivotdata/tabular-other-sheet.xlsx	A tabular PivotTable on another sheet: a field/item pair, a missing field and a reference given as a wide range
date-text/month-name-day-this-year-y2026.xlsx	"Jan-21", "Jan/21", "Jan-1", "Jan/1", "Jan-30", "1Janu", "Janu 1", "1Febr", "01Jan", "Sep 1": days in the current year
date-text/month-name-number-as-year-separators.xlsx	"Jan-32", "Jan/32", "Jan-0", "Feb-30" and "Jan 32"+0-style coercion: the number is a year
date-text/two-digit-year-window.xlsx	"1Jan29" is 2029, "1Jan30" 1930, with and without separators
date-text/month-name-long-prefix.xlsx	"Sept", "Janu", "Septe", "Septem", "Decemb" with a year parse; "Janux 2021" and "Sept" alone are #VALUE!
date-text/month-name-punctuation-value-errors.xlsx	"4.Mar.2026", "Jan.21", "1Jan, 2021", "4 Mar, 2026", "2021Jan", "Jan 15 2021", "001Jan", "Ma", "Sept 1 2021", "Mar.4": all #VALUE!
date-text/month-name-punctuation-control.xlsx	Control: "Mar 4, 2026", "Jan 15, 2021", "4-Mar-2026", "4/Mar/2026", "Jan-2021", "Jan/2021" parse
precision-displayed/pad-on-percent.xlsx	0%, 0.0%, 0.00%: =1/3 is stored as 0.33
precision-displayed/pad-on-currency.xlsx	$#,##0.00, a negative section in parentheses, "$"#,##0.0
precision-displayed/pad-on-thousands-scaling.xlsx	0.0,,, #,##0,, 0.00,: rounded at the scaled digit (1200000, 12000, 9880)
precision-displayed/pad-on-fractions.xlsx	# ?/? and ?/8: stored as the fraction shown (0.3 is 2/7 and 2/8)
precision-displayed/pad-on-scientific.xlsx	##0.0E+0, 0.00E+00, 0.0E+0
precision-displayed/pad-on-multi-section.xlsx	0.00;(0.0) and 0;-0.000: the negative section's decimals apply to negatives
precision-displayed/pad-on-literal-text.xlsx	0.0" u", "x"0.00, and 0.0"%" (a quoted % doesn't scale)
precision-displayed/pad-on-conditional-sections.xlsx	[<1]0.00;0: rounded to two decimals whatever the section shows (=2+1/3 shows 2, stores 2.33; =-10/3 stores -3)
precision-displayed/pad-on-general-control.xlsx	Control: General keeps the full value
precision-displayed/pad-on-date-time-control.xlsx	Control: m/d/yyyy, h:mm, m/d/yyyy h:mm, [h]:mm keep the full value
precision-displayed/pad-on-text-format-control.xlsx	Control: a formula in a cell formatted @ (set after entry) keeps the full value
precision-displayed/pad-on-cross-cell-read.xlsx	=A1 formatted 0.0 reading a General =10/3 stores 3.3; a General cell reading a formatted one reads the rounded value
precision-displayed/pad-on-spill-mixed-formats.xlsx	=SEQUENCE(3,1,1/3,1/3) spilling into cells formatted 0, 0.0, 0.00: stored as 0, 0.7, 1; SUM over them is 1.7
circular/sumifs-first-criteria-range.xlsx	C1 in the first criteria range: truly circular (C1 cached 0); the twin of sumifs-later-criteria-range.xlsx
circular/sumifs-size-mismatch-value-error.xlsx	Criteria range one cell wider than the sum range: #VALUE!, not circular
circular/sumifs-array-criterion-unselected.xlsx	=SUM(SUMIFS(B1:C1,B2:C2,{"x","z"})): C1 not selected, not circular
circular/sumifs-array-criterion-selected.xlsx	{"x","y"} selects C1: truly circular
circular/averageifs-unselected-cell.xlsx	AVERAGEIFS, C1 not selected: not circular
circular/maxifs-unselected-cell.xlsx	MAXIFS, C1 not selected: not circular
circular/minifs-unselected-cell.xlsx	MINIFS, C1 not selected: not circular
circular/countifs-later-criteria-range.xlsx	COUNTIFS with C1 in the second criteria range: not circular
circular/countifs-first-criteria-range.xlsx	COUNTIFS with C1 in the first criteria range: truly circular
legacy-xls/cse-single-cell-array.xls, .xlsx	One-cell Ctrl+Shift+Enter arrays {=SUM(A1:A3*B1:B3)}, {=MAX(A1:A3-B1:B3)} stay arrays (32, -3); SUMPRODUCT as a control
legacy-xls/shared-formula-fill-down.xls, .xlsx	Two formulas filled down six rows, stored as shared formulas (SHRFMLA in .xls)
legacy-xls/implicit-intersection-inside-sum.xls, .xlsx	=MAX(SUM(A1:A3/12),0) as a legacy formula: 2 in row 2, #VALUE! in row 5; 6 as an array formula
round-trip/core-properties-blank-author.xlsx	Author and Last Author blank, created and modified dates present: a save must keep docProps/core.xml

Files added in the third batch

File	Reproduces
pivot-data-model/model-from-table.xlsx	Worksheet table Sales (Region, Category, Amount; 8 rows) added to the Data Model; a PivotTable on the model (Region rows, sum of Amount) on a second sheet. Five GETPIVOTDATA formulas: grand total 392, north 107, south 166, and two #REF! forms
pivot-data-model/oledb-connection.xlsx	OLE DB connection (ACE provider, type 5) to a generated workbook, not in the model; a PivotTable whose cache reads the connection, with records saved. GETPIVOTDATA gives 392, 107, 166. The source path is C:\gc-corpus-stage\source-data.xlsx, which does not exist after the run
pivot-data-model/oledb-connection-in-model.xlsx	The same OLE DB connection loaded into the Data Model (type 100 connection plus ThisWorkbookDataModel); a PivotTable on the model. GETPIVOTDATA gives 392 and 107

Metadata and privacy

• The generator saves every workbook with Excel's own remove personal information on save (Workbook.RemovePersonalInformation). Excel for Microsoft 365 otherwise writes the signed-in Office identity into docProps/core.xml (lastModifiedBy) and the .xls WRITEACCESS record, whatever Application.UserName says, and the save folder into xl/workbook.xml (x15ac:absPath). With the option on, Author, Last Author and Company are saved empty, absPath is left out and workbookPr carries filterPrivacy="1".
• Files are saved through a neutral staging folder (C:\gc-corpus-stage) and then moved.
• verify-metadata.py checks every file: the author fields are empty (or "GridCraft contributors"), there is no docProps/custom.xml, and no part (or, for .xls, no byte run in Latin-1 or UTF-16LE) contains a user folder or the generating machine's user name, display name, computer name or domain. It reads those from the environment and from Excel, so they are never written into this repository.
• No Excel setting is changed for good. Calculation settings (iteration, precision as displayed) are set for every case; SheetsInNewWorkbook is put back at the end.

Relation to KNUFFIXX/artcraft-fixtures

KNUFFIXX/artcraft-fixtures holds small synthetic files behind bug reports to the ArtCraft projects, with a gridcraft/ folder of openpyxl-written workbooks whose expected values (measured in a de-DE Excel through COM) are written into cells next to the test cells. Where it is cheap we follow its conventions, and the two could be merged later:

• Same shape: one folder per target, every file written by a committed script in the repository, no real-world content, MIT OR Apache-2.0.
• Self-describing files: like their expected-value cells, our files carry a text listing of their formulas next to the used range. Our expected values are Excel's own cached results, which gridcraft-cli check reads directly, so no expected-value cells are needed.
• README table: the excel/ table above has their File / Reproduces columns; a Report column can be added as issues are filed.
• Differences: hyphens instead of gc_ underscores in names, a subfolder per failure class, the authoring application (Excel itself) instead of openpyxl, en-US instead of de-DE, and a SHA256SUMS manifest for pinning. To contribute a file there, copy it under gridcraft/gc_<name>.xlsx and add the case's formula to their table.

Provenance

Everything here is original work by the GridCraft corpus contributors. The files are generated from scratch by the scripts in tools/, with Microsoft Excel used only as the authoring and calculating tool. No Excel templates, add-ins or sample files are included. Every number, text and formula is synthetic.

License

Dual-licensed under MIT or Apache-2.0, at your option. Copyright (c) 2026 the GridCraft corpus contributors. See NOTICE.

_{Microsoft and Excel are trademarks of the Microsoft group of companies. This project is not affiliated with, sponsored by or endorsed by Microsoft; the names are used only to describe the files' origin.}