research

The design record behind kindspec — the passes, the measurements, and the experiments that produced rowspec 0.1.0 and the design briefs for blockspec and nodespec.

This is a record, not a specification. Nothing here is normative. Where this repository and a published SPEC.md disagree, the spec wins and this is the history of how it got there.

It is published for one reason: the specs cite these measurements by name, and a citation that resolves to nothing is not evidence.

How to read it

DESIGN.md              the overall verdict — start here
RESEARCH.md            the prior-art survey the design pass began from
SPEC-table.md          the table kind's spec as it stood before rowspec
design-findings/       one file per pass: what was asked, measured, decided
findings/              the competitive and market teardowns
experiments/           the code that produced the numbers, and its output

design-findings/STATE-OF-PLAY.md and design-findings/X-open-problems-closed.md are the closest thing to a summary. ADJUDICATION.md records where two independent passes disagreed and how it was settled.

What is deliberately not here

The corpora. 1.9 GB of cloned repositories, downloaded benchmark archives and generated git histories are excluded, along with virtualenvs and build artifacts. Every one is named in CORPORA.md with its source, so a run can be reproduced rather than trusted.

A third-party product teardown. One experiment disassembled a shipped competitor's bundle to see how it stored documents. Neither the vendored bundle nor the write-up is published: the write-up identifies a named individual and their private infrastructure, and the project's argument does not need it. What it concluded is recorded in findings/00-pithy-teardown.WITHHELD.md, and findings/01-direct-competitors.md covers the same ground from public sources.

Two raw output files over 2 MB — D0-dogfood/ledger-iptv-channels.jsonl and E4-if-sum/out/noteval.json. Both are regenerable from the scripts beside them; the aggregated results that the findings actually quote are tracked.

The method, in three lines

Every substantive defect this project found was found by two independent implementations disagreeing, by a mutation gate that had to notice a deliberate break, or by a mechanical check catching a check that could not fail. None was found by review. That is the transferable result, and it has nothing to do with tables.

Licensing

Prose — *.md, design-findings/, findings/ — is CC-BY-4.0. Experiment code and its output — experiments/ — is MIT, with three exceptions listed file by file in NOTICE: output that quotes SpreadsheetBench is CC-BY-SA-4.0 and output that quotes TmplEnron is CC-BY-4.0, because quoted corpus content stays under the corpus's terms; and third-party sample files keep their own licences. Full texts in LICENSES/. See LICENSE.