ungovr.providers.sgit.ai / The government-graph vault — open it read-only
The government-graph vault — open it read-only
The vault is the workload in §7. It holds the retrieved bytes, their hashes, the compiled artefacts, the seven-step join and the specification that was published before any of it was built. You can open it read-only, right now, and it will ask you for nothing.
Prose from the government-graph vault dkeclt5r @ obj-cas-imm-b9c2cddb51e8, 9 September 2026. The vault's own pack is published raw under /briefs/. When the vault moves ahead, this page is behind — and says so rather than guessing.
Open it, here
The vault is running below, live, decrypted in your browser from the published read key on this page. Nothing on this site holds a copy of it — a push to the vault changes what you see here with no rebuild and no deploy.
Two surfaces open above: App Mode, which renders the vault's own _page.json, and the vault browser under it, with the FILES / SGIT / SETTINGS rail and an explicit R1 W0 · Read-only badge in the chrome. The second one is where you can read the retrieval log, open data/graph.json, and check the commit history that proves the pack was published before the work.
How the key gets there. The frame is loaded with ?embed=1&parent=<origin>; the page waits for the frame to announce itself, then posts {sg:'vault-open', key, mode} with the target origin pinned. The key never appears in a URL and the frame keeps it in memory only. vault-ready / vault-error come back as structured events.
Written, and not yet watched working. The component is the estate's own, vendored unmodified — byte-identical below its provenance header — and the markup and the published read key are checked against it. verified 9 Sep 2026 But nobody has seen this frame render. The container that built this site cannot open a browser TLS tunnel to the vault host:
curlgets 200 and no frame-blocking headers, Chromium getsERR_CONNECTION_RESET. written, not run So the embed above ships in the state the family has a chip for — read it, then open it and find out. If it is blank, the fallback link below opens the same vault in a new tab.
This is the one page on this site that opens a network connection, and it opens exactly one, to
dev.vault.sgraph.ai. Every other page here fetches nothing at all — the ledger says which. The trade is deliberate: an embed that reads the real vault is worth more than a screenshot of one, and it is the estate's own published pattern.
SG/Send vault · read-only · asks for no permissions
government-graph dkeclt5r
Prefer a new tab, or a vault that is having a bad day? The same key, the same vault.
This vault is on a development host.
dev.vault.sgraph.aiis not the production browser, and publishing adev.*URL from a public site is a decision rather than a detail. It ships because the read key is the whole point of §7 and a vault nobody can open is worse than one on the wrong host — but moving to production is handback item 4, and this page changes when it moves.
Or with the CLI:
pip3 install sgit-ai --break-system-packages sgit clone sgit_private_read_d5220d6ada858319cf31f4a4e1a3bd04fe97d647285c16c869adef651db5208d:dkeclt5r government-graph
This link is on a development host.
dev.vault.sgraph.aiis not the production browser, and publishing adev.*URL from a public site is a decision rather than a detail. It ships here because the read key is the whole point of §7 and a link nobody can open is worse than a link on the wrong host — but moving the vault to production is handback item 4, and this page changes when it moves.
Why publishing this key is safe, and which key it is
Two keys exist and only one of them is here.
| What it does | Publishable | |
|---|---|---|
Read key sgit_private_read_… | Decrypts and reads. Cannot commit, cannot push, cannot delete | Yes. It is how this estate shares a vault — the same way sgit.ai/llms.txt does |
Write key sgit_private_vault_… | Everything. Rewrites history | Never |
The write key for this vault was handed over out of band and is not in this repository, in any form. That is not a promise, it is a check: tools/secret-scan.sh runs over the whole tree including the built output in docs/, on every push and every pull request, and refuses the release on a match. It deliberately does not match sgit_private_read_…, and tools/check_site.py carries the same tripwire a second time against the passphrase:uuid shape an sgit key can also take.
This is the check the brief singled out as the one most likely to bite whoever built this site. It is wired first, before any content, for that reason.
What the vault holds
pack/ the specification — 7 files, published as the FIRST commit,
before any implementation existed. That ordering is the
vault's own acceptance test 13, and the only one that
passed before this session
team/ four agentic roles and the single-writer rule
data/raw/ the retrieved bytes, unmodified, so the compiled artefacts
are checkable against what they came from
data/ the compiled artefacts: the graph, both computations
provenance/ the retrieval log, including the failures
catalogue/ the catalogue submission
corrections/ ships empty, on purpose
review/ ships empty
The pack was published before the work, and the history proves it. That is the strongest demonstration the method has: a design that could be checked against the thing it specified, before the thing existed. This session's commits sit on top of it and do not disturb it.
The pack is also published raw on this site, because a site arguing for published provenance that hides its own inputs has refuted itself.
What this site copies, and what it links
Copied. The compiled artefacts — graph.json, computation-1.json, computation-2.json, the retrieval log — are copied into this repository at build time and rendered as part of the site. Every table on the coverage page and the join page is static HTML built from them. No page here fetches anything, which is a build rule, not a preference.
Linked, never embedded. The vault itself is a link. An iframe pulling another host into the page is exactly what the no-third-party rule exists to prevent, and it would break the site with JavaScript off. tools/check_site.py refuses an <iframe> anywhere in the built output.
The copy is stale the moment the vault moves ahead. The page says which vault commit it reports on rather than guessing, and that is the honest resolution rather than a fixed one.
The thirteen acceptance tests
They live in the vault, at pack/06__verification.md, and they are its property rather than this site's — this session moved their states and did not rewrite the tests.
| Passing before this session | 1 of 13 — the pack was in the vault as its first commit |
| Passing now | 9 of 13 — 1, 2, 5, 7, 8, 10, 11, 12, 13. Pinned by release @2026-09-09-retrieval |
| Honestly out of MVP scope | 4 — fractal zoom, byte-range document viewer, and two others |
| Cannot pass as written | 1 — test 3, rebuild the instrument byte for byte. There is no machine-readable CPRA to rebuild from. Recorded as unpassable rather than weakened verified 9 Sep 2026 |
A vault claiming thirteen of thirteen after one session has stopped measuring and started asserting.