A documentation governance system
Headwater types your corpus, checks it as a graph, and accounts for every file it saw. Nothing else can answer “is this corpus still true?”
--- id: HW-DR-0037 relations: governs: - site/index.html - site/how-it-works/index.html - site/compare/index.html - site/proof/index.html - site/tutorial/index.html ---
$ headwater check census 742 files under the corpus root 525 typed 2 untyped 0 unaccounted for 404 findings 16 suppressed, by dated directive taxonomy headwater/standard 4.15.0
Where a vendor site puts logos and percentages, this one puts the run — including the ugly number. A full board with a name against each item beats a clean one.
13944 check instances ran over those 525 documents, from 48 wired rules, of which 7 found anything. The obligation register behind them holds 51 obligations: 48 verified, 2 gap and 1 unverifiable, and each gap names its owner.
A measurement campaign ran on 2026-09-30, and it measured this repository against itself with parts removed, not against the projects it is compared with. So this row stays empty. Principle 11 forbids publishing a number that no run produced.
Shelves, kinds, facets, and relations, in one versioned schema. Your overlay adds what a package cannot know.
shelves:
decisions: docs/decisions/**
kinds:
decision:
purpose: rationale
facets: [status, summary, last_verified]
relations: [supersedes, governs, traces_to]
Documents are typed nodes. Front-matter references are typed edges, and a required edge owes both ends. No language model issues a verdict.
A finding names the rule, the obligation it discharges, the location, and the repair. Mechanical repairs are one command.
0002-deliver-at-least-once.md:9:7 error relation.reciprocity.missing (OB-REL-1): `supersedes` requires both ends, so 0001-store-attempts-in-postgres.md owes `superseded_by` fix (mechanical): headwater check --fix
Indexes, site navigation, an agent instruction file, a JSON export. A hand-edited projection fails the build.
What shelves exist, what kinds live on them, what metadata they carry, and how they may reference each other — all declared in one versioned schema. Customizing the taxonomy never means forking the tooling.
Documents are typed nodes and front-matter references are typed edges. Every validation rule is a constraint on that graph, and every derived artifact is a projection of it.
The corpus feeds agents at intent time, read time, write time, and review time. A probe records whether that context changed what the agent did — not whether it was shown.
commitment II, drawn — real documents from the tutorial corpus
Every edge above was declared by an author and checked by the engine — supersedes owes both ends, and the engine wrote the far half. Edges name targets by identifier, so a rename kills nothing.
Authors declare the graph here. A language model extracts it there.
Read the full comparison, including the arguments against us →
The engine runs, and one campaign has measured part of what it claims. This repository types its own corpus, checks it on every commit, and scaffolds, explains, routes, generates and exports it. The counterfactual campaign of 2026-09-30 measured it by component. The documents under docs/ have a large measured effect on whether an agent gives a sufficient answer. The hooks and skills showed no detectable gain. The evaluation states each figure, with its interval and its cost, and this page repeats none of them. No arm measured the MCP server or the quality of governance itself, and the efficacy claims stay open until one does. The work still open is on the milestones page.
The command surface is 26 verbs in 7 groups, and all 26 carry a written contract. The export layer names 7 emitter targets, of which 2 are built and 5 parse and report the consumer they wait on. The self-assessment states the rest.
Run it against your own corpus, and read what it can't yet answer.
Run the tutorial →