A specification and conformance suite per artifact kind.

Knowledge artifacts — tables, documents, canvases — are edited by several people over years, and version control is where that happens. These formats are specified so that a repository full of them stays correct, and each specification ships with an executable conformance suite that measures whether an implementation actually conforms.

Two branches insert a row into a spreadsheet exported as CSV. Stock git merge merges them cleanly, with no conflict and no marker anywhere, and the total is 480 where the truth is 660. LibreOffice confirms it. A structural merge tool gets the rows perfectly right and produces the identical wrong answer, because no merge algorithm can see that the string in a cell encodes a position.

That is not a merge bug. It is a representation bug, and it is only fixable in the representation:

• Addressed by name, never by position. A1 is a coordinate; a column name is not.
• Correct or refused, never quietly wrong. A file that cannot be understood is rejected, loudly, rather than degraded into a different file that happens to parse.
• Right under stock git, with none of this installed. If correctness needs a merge driver, then correctness is lost the moment someone clones without one, or the forge merges server-side.

Each is named after its unit of identity, because identity is the hard part of every one of them.

Only rowspec exists. The other two are a named intention and a settled identity decision, not a draft — there is no specification for either, and saying so is cheaper than being asked.

Two repositories support all three:

The research repository is published for one reason: the specs cite those measurements by name, and a citation that resolves to nothing is not evidence. It redistributes no corpora; each one is named with its source so a result can be reproduced rather than trusted.

The specification exists so the conformance suite has something to check. The suite checks out two branches, runs stock git merge, evaluates the merged file, and asserts on the computed number — because a merge that produces a well-formed file with a wrong total is exactly the failure that has no marker, and nothing else measures it.

Behind that sits a mutation gate: the implementation is deliberately broken in dozens of specific ways, and the suite must notice every one. A mutant that survives is a hole in the suite and is reported as a failure, not a warning.

The conformance suite is not written by whoever writes the implementation. This is a role, not a review step. It has been tested the hard way: three separate times, an enumerator checking their own work reported full coverage, and an adversary then found silently-wrong cases in the same code. On the most recent pass the two halves were commissioned in parallel and could not see each other — the independent author's cases killed 7 of 8 mutants written blind, and found a live defect that returned a plausible number in every row.

A check that cannot fail must never report a pass. This project has now reproduced that failure ten times in its own tooling: a runner that walked an empty directory and printed 0 failures over 226 unopened cases, a mutation gate that disarmed itself on a reformat, a canonicaliser that scored 129/131 while being the identity function, a differential harness whose injected defect silently stopped applying. Every one was caught by a mechanical check rather than by review. It is the most transferable thing here and it has nothing to do with tables.

Per directory, deliberately. Specifications are CC-BY-4.0; conformance fixtures are CC0-1.0 so they can be vendored into an implementation in any language under any licence; runners are MIT; reference implementations are Apache-2.0 OR MIT. Contributions are under a DCO sign-off, not a CLA.

Embedding test cases inside a copyleft specification document is the mistake this split exists to avoid.