Open Commons ObservatoryEnglish·Русский

Журнал

Как шла работа — что сделано, что рассматривалось и что отвергнуто.

A reference names a row by its key, not only by id

check_rules resolves a [[rules.references]] target (and unique_ids, appears_in) by each table's declared key when that is one field, else id. Why: linux-phone-oco's device base has a parts table keyed by component with no id, and abouts naming a part were refused. The message is unchanged; a test with a target keyed by another field.

Отвергнуто: A new rule field naming the target column per reference: the key already says what names a row.

text-as-data-template hands its scaffold to tad init

tad init DIR starts a bare database. The template's reasons, read from its session_log before it retires (Chesterton's fence): DuckDB and JSON lines (here as 2026-09-21-duckdb-and-json-lines), extracted from a working repository (2026-09-21-extracted-from-a-working-repo; the knowledge base is extracted the same way, from linux-phone-oco/docs), subtree to subrepo and declining to fork git-subrepo (vendoring, superseded 2026-11-11 by released binaries pinned per repository), tagging dist for pinning (the same need, met now by tad-vX.Y.Z releases and bin/install-tad.sh), the RFC-like Pages style (how pages look for people - the face's now), and the backlog as a table (carried to ideal-knowledge-base, whose standing tables it is). Its open render items come here as render-output-guarded.

Отвергнуто: Both hold a domain or a convenience a bare database does not need.

A conditional rule, read right

The canon's rule 'when a year of origin is given, an author is required' had been read as 'an author is always required' - tad reported 20 problems where Python's canon passed. The translator now tells a field's own rule from a rule guarded by another field (a dependency), and refuses any other guarded rule. The equivalence proof caught it.

Presentation, declared as data

tables.toml can say how pages read: [site] (index title, introduction, order, open counts, relation words) and [pages.*] (introduction, groups, order, what each row shows, list layout). The canon's 7 pages rendered from it matched its own renderer's on the first proof run, after the notice line; with nothing declared, three repositories render unchanged. tad 0.3.0 released with no hand in it: a feat landed, land waited for release-please's pull request and landed it, the conveyor built it. The canon switched; Python and DuckDB left the tad backbone entirely. A test expectation of mine was wrong for the third time (a relation's note shows from both sides) - expectations now come from the reference renderer's code.

tad 0.2.0 released by itself and delivered everywhere

release-please opened tad 0.2.0 from a commit; land now lands release-please's release pull requests too (found by branch) and released it; the conveyor built both architectures in the same run; bin/update-pins.sh moved all four repositories to it through their own CI; a weekly timer keeps doing so. On the way: land refused my own 106-character title, so for a few minutes the lab ran a land not yet on GitHub - closed by landing it and checking the installed file is the same.

Canonical form and checked edits, byte for byte

tad canon and tad set, matching DuckDB's own writing exactly: compact JSON, fields in declared order, empty fields as null, rows in their order, new rows at the end. Four proofs passed on the first run.

tad check, in Rust, equal to Python's validation

A tad crate: declarations as data, translated from the four DuckDB schemas by sqlparser (refusing what it does not know), and every row checked. On the four repositories and eight broken copies it reaches the same verdicts as Python's canon. A first comparison against Python's check misled: that command only compares names; corrected by reading the code.

Отвергнуто: That: it does not validate rows; canon does.

tad in Rust, designed

Looked at what tad is (a 127-line runner and a 478-line renderer, in four repositories, 41 tables using only simple constraints) and at what others built (Frictionless Table Schema, CSV on the Web, Dolt, Datasette). Designed tad in Rust with tables declared as data, proven equal to the Python version before any switch.

Отвергнуто: That: a database engine to check text.

tad released; the playbook is the first on it

tad-v0.1.0 tagged by land --tag on merged code; the conveyor built it for two architectures with checksums and provenance. The playbook moved onto it and its CI went green; tad set made its first real change there (a session entry). Mistakes on the way: an edit check that read '.tad' as a pattern (stopped safely on a false alarm); a commit message predicting document changes that did not happen - corrected in the playbook's own log.

Documents rendered by tad, byte for byte

The renderer ported to Rust, with Python's string habits (title, capitalize, str) and its look-around patterns (fancy-regex). Every document of the four repositories, both formats, identical to today's Python renderer. Found: three repositories carry older copies of the renderer.

Отвергнуто: That as the reference: the copies are older; today's engine is the one tad replaces.

Every rule of the four tad repositories, as data

The 13 cross-table SQL checks became four kinds of declared rule; tad agrees with Python on all four repositories and on seven more broken copies. A test of mine once printed an empty verdict for a declaration I had cut wrongly; tad had refused it rightly - the proof now demands pass, caught, or an error shown.

The switch is done: Python leaves tad's backbone

All four repositories switched to tad in Rust, each proven on its own terms (feeds byte for byte; the canon's pages byte-identical). The Python engine retired from this repository, its last version named in the README. Mistakes on the way, each caught: an edit check reading '.tad' as a pattern, a README edit that took a heading with it, a README link to a removed folder, document changes predicted that did not happen, a 102-character title (refused by land), and a job that printed a failed verify without stopping.

Отвергнуто: That: no repository vendors it any more, and a second engine would drift.

It tells its essence

Its idea written for anyone - tables kept as plain text, a guard at the door - and the workings of the step it has reached. It says plainly that it is kept as it is and that its ideas continue in ocoeco.

Отвергнуто: Presenting it as where new work goes.

tad gets a path of its own

tad's backlog and milestones written as tad data and exported through its mark, so the observatory shows how close tad is to being a rhizcut in every repository.

Отвергнуто: A top tab (it would crowd the overall look); tad as a template (a rhizcut instead).