ungovr.providers.sgit.ai / decks / How this composes
How this composes
With Risk Mandate, Licence to Operate, and the sgit vault family · 10 slides · September 2026
9dceb805af0aa1df — checked against the vault on every buildThe shape everything shares
Four vaults, one chain, different segments of it:
AIUC-1 standard -> control -> requirement -> attestation -> conformance -> insurability LICENCE grant -> mandate -> DELTA -> policy -> covered / not covered THIS VAULT entity -> instrument -> provision -> obligation -> control -> evidence -> acceptance
They overlap at control, evidence and acceptance. That is not a coincidence — it is the same risk-acceptance grammar applied to three different sources of obligation: a standard, a mandate, and a statute.
This vault supplies the third one, which none of the others had.

What we take: the two-edge rule
The AIUC-1 conformance vault keeps two edges "never traversed in one query without the query naming which it used":
| Edge | Answers | Absent means |
|---|---|---|
evidenced_by | does the standard say this? | a build defect |
attested_by | does this subject do this? | the control is unevidenced — which is the finding |
We do the same thing on a different axis. ung: is what UnGovr publish; sg: is our model; exactly one edge crosses between them and it is drawn inferred.
Their rule is enforced by a test that goes red if a layer edge reaches a source_observation. Ours is enforced by a build check that fails if any page names sg:resolvesTo without marking it inferred.
What we take: unevidenced is a state
From the same vault: "Every control in scope gets a row whether or not anyone has looked at it, so an absent row is never quietly read as compliance."
Their first build of one subject across 53 controls: 2 evidenced, 48 unevidenced, 3 contradicted — "the designed answer, not a failure to finish."
Our evidence node ships unevidenced and our acceptance node ships with no owner. Both are the default, both stay, and both mean nobody has looked rather than nothing is wrong.
What we contribute: a jurisdiction layer
327,138 entities. 398 records laws. 254 of them sub-national.
A policy/v1 that covers an agent acting for a public body has to know which body, and which law binds it. Neither the Licence to Operate vault nor the conformance layer carries that, and neither should have to build it.
This vault is that lookup — with the honest caveat attached, which is that the entity-to-law edge is inferred, reaches 96.6%, and hands 20.5% of matches to a human.

What we contribute: a clock nobody negotiates
The conformance layer's headline mechanism is that time breaks the policy: move the date to 2027-01-15, "with nothing edited by anybody", and one condition becomes 53 exclusions — because the attestations behind them expire.
Our obligations expire on a different clock, and it is not ours.
Gov. Code § 7922.535(a): the agency shall determine and notify within 10 calendar days, extendable by 14 in unusual circumstances. Calendar days — office closures do not toll it.
An attestation expiry is a policy parameter somebody chose. A statutory deadline is a fact about the world. A conformance model that can carry both is strictly better than one that can only carry the first.
The delta, at zero
Licence to Operate's argument, made countable:
| In their demo | ||
|---|---|---|
| CAN DO — the grant | what the agent can technically reach | 12 capabilities |
| MAY DO — the mandate | what is actually expected, and all the policy insures | 4 |
| THE DELTA | inside reach, outside authority. Nothing covers these | 8 |
This vault's app declares "permissions": [].
CAN DO equals MAY DO. The delta is zero — not managed down, absent. It is the smallest possible instance of their argument, and it is why a stranger can open this vault read-only and be asked for nothing.
Where the three meet: an unowned acceptance
Our chain ends on sg:acceptance/unassigned — no named owner, no review interval, no revocation path.
Read through the conformance layer's vocabulary, that is not a gap in the diagram. It is a conformance state, and it is unevidenced.
Read through Licence to Operate's, it is not a condition on a policy. It is an exclusion — and the exclusion carries its reason in the control's own words: an acceptance without an owner is a note.
Three vaults, three vocabularies, one unfilled field, and all three agree on what it means. That is the strongest evidence so far that the grammar is real rather than a house style.

What we took from the family, concretely
| From | What we used |
|---|---|
| Risk Graph Explorer | PUBLIC.md's three rules for a vault whose read key is published. We added a fourth: nothing whose licence forbids republication |
| Regulation Graph | law-as-a-citable-graph, hash-verified to source bytes — the closest analogue to this work |
| AIUC-1 conformance | the two-edge rule, and unevidenced as a default rather than an absence |
| Licence to Operate | grant / mandate / delta, and the policy/v1 shape |
| Risk Mandate | that a real application ships as a vault, with a release history |
| sgit.ai | the embed protocol — the read key posted to a pinned origin after a handshake, never in a URL |
The one thing to build next
Make the obligation layer a policy/v1 producer.
This vault already has the pieces: an entity, an instrument, an addressable-enough provision, a statutory deadline as a scalar, a control, and an acceptance with a named hole in it.
The missing step is emitting the conformance row — one obligation, one control, one state, one valid_until driven by the statute rather than by an attestation — in the shape the conformance layer already reads.
Then a policy could say: this agent may file a records request on behalf of this body, in this jurisdiction, under this law, with this deadline — and say what it does not prove.

What this does not prove
Borrowed deliberately from the conformance layer's own does_not_prove field:
- It does not prove Santa Barbara County complies with anything. The evidence node is
unevidencedbecause nobody has observed the county. - It does not prove UnGovr intend any of this. No conversation has taken place; every
statement about their intentions is inference from what they publish.
- It does not prove the inferred edge is right for any particular entity. It reaches 96.6%
and hands 20.5% of those to a human.
- It does not make anything citable by version. That release channel was withdrawn, and
acceptance test 9 regressed to
unevidencedrather than being left looking satisfied.
These slides are rendered from 03__how-it-composes.md, republished byte for byte from the vault. The live decks — the same file, parsed by the vault's own app — are in the Decks view of the vault, which moves the moment the vault does. This page moves when the site rebuilds.