What this site ships, and what it only asserts
A memory that cannot distinguish designed from built teaches agents to overclaim, and that is the most common failure found across this estate's stale artefacts. So this page applies the rule to the site making the argument. It is kept as a table rather than a tone of voice, because a tone of voice cannot be checked.
Every claim on this site, by status
| Status | What it means | Where it applies |
|---|---|---|
| Measured elsewhere | A sibling pack took the measurement and published it with its evidence. This site links it and carries no number. | Testing coverage · the dead workflow · the stale index and the fictional README · every cell of the scorecard |
| Position | An argued stance, not a measurement. Defensible, falsifiable, and nobody has tested it. | The memory thesis · design for the failure you had · pre-approve the ladder · the hub contract itself |
| Described from the corpus | A faithful summary of documents in the estate, not verified against current practice. | IFD · how a brief becomes work · the four testing non-negotiables |
| Proposed, not built | Named because naming it is the discipline. See below. | The list in the next section |
| Absent | On the estate's own list, with nothing behind it. | Backups |
What is actually built, on this site
The small honest list of things that exist and run, rather than things argued:
- The pipeline. Validate, then tag, then deploy, on every push to the release branch. Validation also runs on pull requests, so branch work is gated before it reaches the release branch. How it works.
- Auto-increment tagging from a single version file, with the release commit's subject required to agree with it, the next-minor rule enforced, and historical tags backfilled from commit subjects.
- The hub's own rules as CI checks. A discipline page that cites no sibling fails the build. A page carrying figures with no as-of date fails the build. These are small, and they are the demonstration that the freshness-check proposal is a shape that works.
- The agent surface: llms.txt, llms-full.txt, the front page's markdown twin, a sitemap, and the brief pack fetchable as raw markdown at constructed URLs.
- Chrome from one definition. Nav, footer and the version badge are written once and stamped across every page, which is why the version cannot disagree with itself.
Proposed, and not built
Each of these is referred to somewhere on this site as a good idea. None of them exists.
- The scorecard generator. The scorecard is hand-assembled and dated, which is the artefact shape most likely to go stale. Rebuilding each cell from its sibling pack at release time is the fix, and it is not written.
- The reality-document freshness check at corpus scale. The proposal is sound and the version running here is the toy version, on sixteen pages.
- The backups doctrine, or the finding that the vault model already is it. The question is stated and unanswered.
- Confirming IFD against current practice. The guides are summarised, not verified.
- Counting the day-indexes and debriefs. The scale is part of the PM argument and it is stated approximately.
- The memory thesis on the network hub. The page exists here and is proposed for sgit.ai, where it would explain all fourteen sites rather than one.
- Any evaluation at all of whether an agent briefed from these sites performs better than one that is not. The explainability row records that there are no evals anywhere, and this site is not an exception to it.
What is deliberately not published
Stated because redaction discipline is part of the memory model rather than a footnote to it. A memory without one leaks, and publishing here is irreversible.
- Any of this estate's own financial figures. Rates, margins, pricing, provider detail, projections, investor material. The budget discipline is published; the numbers are not, anywhere, in any form.
- Live hostnames, account identifiers and stack addresses, wherever platform material is quoted.
- Any review naming a private individual, and one security review classified internally as an attack roadmap for live code.
- Market statistics from the villagers brief. That brief flags its own market figures as loosely attributed vendor commentary, so this site does not carry them. The market argument belongs to open-source.sgit.ai; this site is the same argument from the inside, without the statistics.
Licensing
All content on this site is CC BY 4.0, consistent with the network. Quoted code and the in-repository methodology guides are Apache-2.0 and retain their notices; this site describes and links them rather than copying them, so the repository stays canonical. That is the same second-source-of-truth rule the versioning section argues for, applied to prose.