Release notes

Every release, newest first. Downloads for each are on the releases page.

Beta

HippoDatamus is below version 1.0. The configuration format and the data a given configuration produces may change between releases. Any change to generated output appears under its own heading below, so an existing schema never changes what it produces without being told.

0.7.0 2026-09-17

Output changes

• A Gaussian column produced different values on Windows and on Linux. The same configuration and the same seed gave different bytes depending on the operating system, which is the one thing this product promises will not happen. It affected roughly 1 row in 125, and only that column type.

  The cause was the Box-Muller transform, which needs a sine. .NET forwards Math.Sin, and Math.Log and Math.Exp with it, to the platform's own C maths library, and the results differ by one unit in the last place. Math.Sqrt does not - IEEE-754 defines it exactly.

  A Gaussian column is now drawn by the Marsaglia polar method, which needs no sine, over a logarithm computed from IEEE-exact arithmetic rather than from the platform. The distribution is unchanged.

  A Gaussian column therefore produces different values than it did in 0.6.0, for the same configuration and seed. No other column type is affected: the two operating systems were already byte-identical on the other 63 columns of the all-column-types sample, and are now identical on all 64.

Fixed

• The application closed itself when two columns had the same name. Adding a second column and pressing Generate was enough to do it: every new column arrived named "New Column", and the dependency graph built a lookup keyed on the name, so the second one raised a duplicate-key error. It surfaced as the window vanishing with no message, and nothing was written.

  This is what failed Microsoft Store certification for 0.6.0, under 10.1.2.10 Functionality: "the product crashes after trying to create a file."

  Three things were wrong and all three are fixed. New columns are now named "New Column 2", "New Column 3" and so on, so the case is not created by using the application normally. Two columns given the same name deliberately are now reported as a validation error that names the column, rather than raising an exception - as are two columns sharing an internal id, which is what a configuration file gets when a column block is copied by hand. And checking a configuration can no longer take the application down: that step ran before the error handling that covers the run itself, so anything unexpected in it reached the dispatcher and terminated the process, where it now reports what went wrong and leaves the application usable.

0.6.0 2026-09-15

Added

• File, Open Sample opens one of the sample schemas that ship with the application, without going through a file dialog. On a Store install the samples live inside the package, which Windows does not let you browse, so they were shipped and unreachable. The list shows each sample's column and row count, simplest first.

Output changes

• Names now come from every English locale, not just the American one. First names go from 3,007 to 3,954 and last names from 473 to 911 - last names nearly double, because the Australian, Indian, Nigerian, and South African lists overlap the American one far less than their first names do. A Full Name column's domain grows from 1.4 million to 3.6 million.

  A First Name, Last Name, or Full Name column therefore produces different values than it did in 0.5.0, for the same configuration and seed. Nothing else about those columns changes: the count, the seed derivation, and the uniqueness guarantee are all as they were.

  Duplicates between locales are dropped, ignoring case, so the pool does not hold both McCullough and Mccullough. en_AU_ocker and en_BORK are left out as joke locales.

Fixed

• Choosing a source file for a File Column now happens from a Browse... button beside the Path field, as it would on a web form. It used to be a single button beneath the whole settings panel, which for this column type sits below six settings and their captions - off the bottom of the window until you scrolled, for the control most people reach for first.
• A Full Name column drew from two different sets depending on its Unique setting: the random draw came from Bogus's own composed name while the indexed one walked the first and last name lists. Both now use the lists.
• Privacy page wording, on the website only.

0.5.0 2026-08-23

First public release.

Added

• Desktop application for building a schema column by column, each showing its type and settings at a glance.
• Command-line tool taking a .yaml configuration, for scripted and CI use.
• 59 column types across identifiers, names, addresses, finance, internet, commerce, text, numbers, dates, and structure. Full inventory in docs/column-types-review.md.
• Reproducible output. The same configuration and the same seed always produce the same data, on any machine. Column ids take part in seed derivation, so sharing a dataset means sharing the .yaml rather than the seed.
• CSV, JSON, and Excel output. Every value is formatted through one invariant-culture path, and each format's line ending is written explicitly - CRLF for CSV, as RFC 4180 defines a record, and LF for JSON. A given configuration and seed therefore produce the same bytes on every platform, so a checksum of generated output is worth taking.
• Pick lists with a per-row count: a cell can hold several distinct values, joined by a separator in CSV and Excel and written as a real array in JSON.
• File-backed columns that draw values from a column of an existing spreadsheet or delimited file, with the same per-row options as a pick list. Paths are stored relative to the configuration and use forward slashes, so a schema and its sources move together across operating systems.
• Column import from a spreadsheet's header row, so an existing file becomes a starting schema.
• Template columns composing values from other columns, with .NET format strings ({price:0.00}) applied in the invariant culture.
• Unique constraints, single-column and composite, validated against the domain size before a run starts rather than failing partway through.
• Helper columns excluded from the output while still driving templates and constraints.
• Configuration bundles (.hippo) sealing a schema together with every file it reads, so a scenario can be handed over as one artefact. The command-line tool takes a bundle directly.
• Dates without a time component, and Excel date cells that render as dates rather than serial numbers.
• Email addresses that cannot be delivered. Email uses real domains by default, because that is what makes generated data convincing. Turning on UseFixedDomain puts every address on one domain, pre-filled with example.com - reserved by IANA and guaranteed never to accept mail - so the option is enough on its own without knowing which domains are safe. Only the domain changes; the local part is generated the same way either way.
• An MCP server exposing the application to an agent, so the UI can be driven and asserted on without a human at the keyboard.

Fixed

Found by a review that verified behaviour rather than reading code, and fixed before release:

• Generating straight after launch failed on a Store install. The default output path was the working directory, which for a packaged app is the read-only package folder. Unpacked from a zip it is wherever you put the folder, so this only ever affected the packaged build, and only until you chose an output path yourself. It now defaults to your Documents folder.

• A template read NULL from any column declared below it. Building the dependency graph reset each column's edge list as it went, discarding the edges a template earlier in the schema had already added. The template then generated before the columns it reads. It looked like "references only work if the column is above the template", but it was less predictable than that: whether a given reference survived depended on where the discarded edge left that column in the queue, so one schema could resolve one reference and null another. Templates now resolve their columns wherever those columns are declared, including through a chain of templates, and a circular reference is reported rather than producing data.

• Generation was far slower than it needed to be. A hundred thousand rows of a twelve-column order table took 102 seconds; the same run now takes 0.66 seconds, and a million rows of CSV takes about four. Output is byte-identical to before - verified against a baseline generated by the old code - so nothing about the data changed, only the speed. Measured numbers are published at hippodatamus.com/docs/performance.

• A template's number format followed the machine's locale. {price:0.00} produced 1.50 on one machine and 1,50 on another where the locale uses a decimal comma - the same configuration and seed giving different data. Number formats are now locale-independent, like everything else written out.

• Password's maximum length did nothing. Every password came out exactly MinLength characters; MinLength: 12, MaxLength: 40 produced twelve, every row. Each row now draws a length from the range.

• A unique combination containing a column with unbounded values was silently unenforced. A combination needs every column in it to have a countable set of values, and one that does not cannot take part - but the rule was accepted without comment and then applied by nothing at all. It is now reported before the run. Marking such a column unique on its own still works, as it always did.

• Currency Code could produce a constant column. Used inside a unique combination it returned USD for every row, and the combination itself came out with duplicates, while validation reported the request as satisfiable. It is now rejected with a message instead. Ordinary use was never affected and always produced real codes.

• Generated line endings followed the platform. The same configuration produced CRLF on Windows and LF on Linux, which made a checksum of output unportable. Both formats now write their line endings explicitly.

Changed

• Excel reading and writing was rebuilt on a different library, for licence reasons. The binaries are about 6 MB smaller each as a result, and every component HippoDatamus ships is now under a permissive licence. The full list travels in THIRD-PARTY-NOTICES.txt inside every download.
• The legacy .xls format is no longer supported, for reading or writing. No permissively licensed library handles it. Use .xlsx, or CSV for anything larger than a worksheet holds.

Known limitations

• Pre-release. The configuration format and generated output may change below 1.0.
• Windows and Linux only. No macOS build.
• The portable Windows download is unsigned, so Windows may warn that the publisher is unrecognised. The Microsoft Store build does not.
• Currency Code cannot be marked unique, on its own or in a combination. Its values cannot be enumerated, so distinct ones cannot be produced on demand. This is now reported rather than silently producing duplicates.
• Base62 Id is sequential, not random. It is derived from the row number, so row 0 is 0 and two Base62 columns in one schema hold the same values. Use Nano Id or Guid for something unpredictable.
• Email still defaults to real domains such as gmail.com, because realistic data is the common case. Turn on UseFixedDomain when the data might reach something that sends mail. Domain, URL and Username have no equivalent option - nothing can be delivered to them.