--- deck: What we plan to build subtitle: The security-standards graph — specified, nothing built, and honest about which date: September 2026 accent: teal --- # What we plan to build > **Deck 4 of 4.** The dev pack, in ten slides: the gap, the design, and the four things that > could still make it wrong. > > The pack itself is eight files in `packs/security-graph/`, republished on the site at > `/packs/security-graph/`. This is the markdown source of the deck; the **Decks** view is a > projection of it, and if the two disagree this one is right. --- ## Every standard says what good looks like. **None says who has to do it.** A control says *maintain an audit log*. It does not say **Santa Barbara County must, under Gov. Code § 7922.535(a), within 10 calendar days.** The obligated party and the jurisdiction that binds it are the two node classes every standards graph gestures at, and none of them identifies. UnGovr publish **327,138 identified bodies** and **398 jurisdiction-keyed laws**. Those are the anchor nodes our graphs are missing. > **Notes.** This is the entire pitch in three sentences. If you say nothing else, say this. The > gap is not in their data and not in the standards — it is between them, and it is one edge. --- ## The number that carries the argument AIUC-1's conformance vault holds **1,126 crosswalks** from its controls out to **489 distinct control items across thirteen frameworks.** **Ninety-five of those crosswalks point at law** — EU AI Act 62, Colorado AI Act 18, California SB 53 9, NYC Local Law 144 6. **Three of the four laws are sub-national United States jurisdictions.** > **Notes.** Counted from graph/edges.json in vault 2wzct4k7, not read from its prose. A > crosswalk ending at "the Colorado AI Act" is a statement about a text, not about which bodies > in Colorado it binds. UnGovr's Atlas is keyed on exactly that axis. --- ## Two findings from opening their vault, and the second changed the pack **One: all thirteen frameworks are AI-specific.** No NIST CSF, no 800-53, no ISO 27001 or 27002, no SOC 2, no CIS, no PCI DSS. That is the right scope for an agent standard — and it means the general-security half of the map is **absent from this estate**. **Two: they already use `anchor:` nodes**, reached by an `anchors_to` edge, 489 of them. > **Notes.** The second finding forced a rewrite. File 01 had implied the estate had not applied > the anchor pattern. It had. So the pack proposes a different AXIS of anchor — theirs anchor > provisions across standards, ours would anchor obligations to the body and the place — rather > than proposing the mechanism. Two anchor axes compose; they do not compete. --- ## One label had to be read rather than trusted One of the thirteen is labelled **"OWASP Top 10"**, which reads as the web application list — a general security artefact. If it were one, the finding above would be wrong. Its members are `LLM01:25 Prompt Injection` through `LLM10:25 Unbounded Consumption`: the **OWASP Top 10 for LLM Applications.** **The finding survives only because the members were listed instead of the label believed.** > **Notes.** A shorthand that collides with a different, better-known standard is exactly what an > anchor node exists to disambiguate. Worth reporting upstream rather than working around. --- ## Where the rabbit hole actually stops Follow the evidence to a law and UnGovr hands you a **structured summary in JSON** — 31 to 34 fields — and **a link to the official publisher**, on 392 of 398 laws across 339 hosts. What it does not hand you: **the text**, in any format. `.md` `.html` `.txt` `.pdf` `.jsonld` `.ttl` all 404, and `Accept:` headers return **byte-identical** responses. **And 254 of the 398 laws — every sub-national one — have no reachable detail document at all.** > **Notes.** 24 of 24 sub-national slugs returned 404 against 26 of 26 national ones at 200. The > jurisdiction carries a slash and the route is a single path segment. Sub-national is where the > variation lives, and it is exactly the half the crosswalks point at. --- ## The rule the whole design rests on > **Don't merge vocabularies — merging erases the disagreement. Keep them intact and bridge them > through anchor nodes.** *graphs.sgit.ai, thesis sentence 6* Merging UnGovr's entity vocabulary into a standards vocabulary would destroy both. Bridging them costs **one node class and one edge**. > **Notes.** This is inherited, not invented, and it is why the pack proposes two anchors — > ung:entity and ung:jurisdiction — rather than a schema everyone must adopt. --- ## Fifteen edges, and the four that caught us out Every edge is a verb with a **distinct inverse**. Six come from graphs.sgit.ai's established set, four from the AIUC-1 conformance layer, **five are proposed here and marked as proposed**. The pack originally said eight were *reused unchanged*. **Four of the six had been quietly renamed** — `observed_on` given the inverse `observation_of` where the published set says `bears_observation`. **Renaming another vocabulary's inverses and calling it reuse is a fork.** > **Notes.** Caught by fetching the published edge set and comparing row by row, not by rereading > our own table. It is the exact failure rule 6 exists to prevent, committed by the file that > states the rule. The four are now theirs. --- ## What we would supply, and what we would not | They supply | We supply | |---|---| | The instrument, found and summarised, with its publisher URL | Parsing it into provisions, each with a hash | | 254 sub-national jurisdictions, identified and slugged | The anchor binding a body to a place | | A controlled vocabulary — 22 fields, 17 enums | Edges with named inverses | **None of it requires their corpus to be redistributed**, which matters because part of it cannot be. > **Notes.** The AI-law corpus licence says "No license is conveyed by receipt of this file". The > graph can carry the anchor, the hash and the link, and leave the text where its licence > requires it to stay. We have built the parsing half twice already — Regulation Graph from > Formex, and the EU AI Act current text with a sha256 per provision. --- ## The state of it: nothing is built, and the tests say so **Thirteen acceptance tests. Five pass — and all five are inherited** from work this vault had already done. One is `asserted`. **Seven are `unevidenced`.** Four blockers are open. One, S1, closed on the day it was written by doing what the pack said to do first. > **Notes.** This slide is the point of the deck. A specification published before its > implementation, with the count of what it has not earned, is checkable against whatever gets > built. The count itself was wrong first time — it said "two of twelve" — and was caught by > parsing the table with a script rather than reading it. --- ## What would make this wrong **S3 is the one that kills it.** If UnGovr's entity ids churn between releases, the anchor thesis fails and nothing else in the pack survives. **It is unmeasured.** Two snapshots, compared, is the next cheapest thing to do. **S4:** if a standards body already publishes machine-readable jurisdiction scope, their binding replaces our inference and should. **S5 is not closeable by us:** the AI-law corpus cannot be redistributed. > **Notes.** End here rather than on a roadmap. The pack's own rule is that every claim carries a > state and the default is the weakest one — that applies to the pack about itself. Nobody should > leave this deck thinking any of it is built. --- This document is released under the Creative Commons Attribution 4.0 International licence (CC BY 4.0).