Level 500: Governance, safety and research · Module 500.7

Four colleagues join two columns of blank cards with lengths of string.
Map it both ways

OVERT and the SAIL crosswalk · Plain gear

Facts as of 18 November 2026. Reviewed and published, version 1.0.0. Big Widget Company, its people and every card below are made up for this course. They are teaching examples, not real records. The practice files for these cards are in this module's bundle/ folder. Every "yes" or "no" below is what the author got when running them. Run them yourself too. Every real document is cited. Where we do not know something, we say so. This is not legal advice.

The sealed envelope

Attestation by construction: Controls make cryptographic proof as a side effect of running, not as paperwork done afterwards. Privacy by architecture: Protected content stays inside the operator's environment; only hashes and signed receipts go across trust boundaries. Independence by structure: Whoever attests to governance is structurally separate from whoever is governed, and an operator attesting to itself does not comply. Statistical rigor by default: Safety claims come with confidence intervals, sample sizes and methods an auditor can reproduce.
The four design principles highlighted by the lesson are design aims, not guarantees; the course paraphrases the home-page summaries, not a conformance or effectiveness finding (s2).
Figure data as text

Four of OVERT's six design aims

Design aims, not guarantees

PartCourse wordingSources
Attestation by constructionControls make cryptographic proof as a side effect of running, not as paperwork done afterwards.[s2]
Privacy by architectureProtected content stays inside the operator's environment; only hashes and signed receipts go across trust boundaries.[s2]
Independence by structureWhoever attests to governance is structurally separate from whoever is governed, and an operator attesting to itself does not comply.[s2]
Statistical rigor by defaultSafety claims come with confidence intervals, sample sizes and methods an auditor can reproduce.[s2]

The four design principles highlighted by the lesson are design aims, not guarantees; the course paraphrases the home-page summaries, not a conformance or effectiveness finding (s2).

Imagine a letter in a sealed envelope. The wax seal is unbroken. The stamp in the wax is your bank's.

What do you know? You know nobody opened the letter after it was sealed. If you trust that only the bank has that stamp, you know the bank sealed it.

What don't you know? Whether the letter is true. Whether the bank did what the letter says. Whether the bank's rules are any good.

That gap is the whole of this lesson. We look at two real frameworks and one kind of table that joins them:

You will learn to:

Please note: the Foundation's OVERT page is written by Ken Johnston and Bob Rapp, who also lead this course.3 Their Foundation is part of OVERT's planned stewardship.1 So we check OVERT's own pages too. We name OVERT, SAIL, Pillar Security and OWASP only because they wrote or published the documents we are reading. Bob Rapp's rule (20 November 2026) is that a company may be named only as the author or publisher of a source we cite. Other companies and products are not named. Naming a document is not advice to use it, and not a promise that any product meets it.

Whose words these are. When we quote OVERT's standard, we are using short pieces of OVERT — Observable Verification Evidence for Runtime Trust, Version 1.1 (June 2026), by Glacis Technologies, Inc. OVERT's governance page says that text may be reused under the Creative Commons Attribution 4.0 licence (CC BY 4.0), if you give credit, name the title and version, link the licence and say what you changed.4 What we changed: we picked out short pieces and put them in our own order. We have not checked whether that licence covers OVERT's other web pages, so we tell you what those pages say in our own words. You can read them yourself at overt.is.

What OVERT is, and what it is not

OVERT stands for Observable Verification Evidence for Runtime Trust.1 Version 1.1 came out on 11 June 2026.1 Its records are meant to show that the rules, security checks and human checks a company says it has really ran.1 Anyone should be able to check those records.

OVERT lists six design ideas.1 Four matter here:

Now the important part. OVERT's own text says a conformance claim "does not constitute a general representation that the system is secure, compliant, or safe."5 It also says OVERT does not vouch that what an AI says is true.5 And its governance page makes clear, in our words, that using the OVERT name is not a stamp of approval for any product.4

So "it meets OVERT" does not mean "it is safe". It is a claim about records.

There are levels, too. OVERT conformance is a level from 1 to 4 plus a scope.5 At Levels 1 and 2, the standard says, an independent notary is not required.5 The company may check its own conformance at those levels.5 Some of the evidence may come from the company itself, and must be labelled that way.5 OVERT grades evidence, too. Only its top grade means someone independent made the proof.5 So always ask four things. What level? What scope? Who checked it? Who signed each piece?

The Foundation's page links to a tool it calls "an OVERT-conformant runtime".3 The page does not say which level or scope.3 Now you know to ask.

Who runs OVERT? On 21 September 2026, OVERT's site said three groups had signed a shared stewardship agreement.1 The site also said it had not started yet.1 Its governance page says a separate certificate must be signed first, and that it has not been signed yet.4 The Foundation's page says "the agreement is signed."3 Both can be true: signed, but not started yet. We do not claim the handover has happened.

Big Widget's blocked order

Big Widget Company has an AI helper called parts-buyer. It orders parts from suppliers. Big Widget's plan says every time it wants to use a tool, it should go through a checkpoint called the gateway. Big Widget's own team runs the gateway. Nobody has yet shown that every route really goes through it.

A rule called TS-2 says parts-buyer may only use certain tools, and may only place small orders.

One day parts-buyer tries to place a big order. The gateway says no. It writes a record and seals it with its own digital key. We call this record Card R7. Card R7 says:

The course gives you four rule cards. They are modelled on OVERT's ideas, but they are the course's own rules, not OVERT's.

CardQuestionCard R7 (author's run)
RC-1 SealIs the seal good, with a key we trust?Yes
RC-2 Who signedDid the company sign, or someone separate?The company signed
RC-3 PrivateIs the private part locked, and has a named person checked what is left showing?Locked, yes. Checked by a person: not on file, so no
RC-4 WorksIs there a test, by someone else, that a stranger could repeat?No

You can run this yourself. The bundle/ folder has Card R7, the rule, the two evidence files and a small checking program. It needs Python and nothing else, not even the internet. VERIFY.md in that folder walks you through it. The checker only does the machine parts: the seal and the fingerprints. It does not decide RC-3 or RC-4. A person has to.

Open EV-17, the test report. It says Big Widget's own gateway team tried 100 big orders, and none got through. That sounds good. But the same team runs the gateway. They only tried big orders. They did not say beforehand how many slips they would accept. So RC-4 is still no.

Why a fingerprint is not always private. Suppose an order could only be one of a few hundred part numbers. Someone with the fingerprint could make a fingerprint of every guess until one matches. Then they know the order. A locked fingerprint mixes in a secret only Big Widget knows, so guessing does not work. The bundle shows this. Out of 6,000 possible orders, a plain fingerprint was matched after 1,653 guesses. The locked one was not matched at all. Even then, the record still shows the time, the tool and where the test report is kept. So the course adds its own rule: a named person at Big Widget, the privacy owner, has to check those too. The Foundation says its own receipts never carry "prompts, outputs, payloads, personal data".6 A fingerprint is not permission to share what it stands for.

So what does Card R7 prove?

Following one risk id

SAIL comes from a company called Pillar Security.2 It splits an AI system's life into seven phases, from planning to retiring it.2 Each risk gets an id like 1.1 or 5.4.2 The first number is the phase.

Risk 3.13 is in the third phase.2 That phase does not build AI helpers. It looks closely at how they are set up: who they act as, and which tools and connections they can reach.2 In our own words, 3.13 is about an AI helper having more tool access, and more connections to other systems, than its job needs. (We use ids and our own words. Read SAIL's own wording on its page.)

Big Widget keeps a made-up crosswalk. Its row for 3.13 says: the control is TS-2, and the evidence is the gateway's records plus EV-17 and EV-18. Big Widget's governance lead signs off the row. It is that team's judgment, not a fact about the world.

Follow the trail:

  1. Risk 3.13: too much tool access.
  2. Control: TS-2 says what parts-buyer may do.
  3. Record: Card R7 says TS-2 blocked one order because it was too big.
  4. Evidence: EV-17 and EV-18.

One blocked big order shows one limit working once. It does not show that the other tools are limited, or that the connections to suppliers are no wider than they should be.

A crosswalk row is like a sign that says "the spare key is under the mat". It tells you where to look. It does not prove the key is there, or that it opens the door.

What is still missing? To know TS-2 really works, you would want:

What can the record never see? This part is our own reasoning. The gateway only writes records for orders that pass through it. Suppose parts-buyer has a second, secret way to reach a supplier. Orders that way make no record at all. A missing record does not mean nothing happened. OVERT itself says it does not promise to see everything.4 Also, a small order is allowed even if it is the wrong part. TS-2 checks size and tools, not whether the order is right.

The Foundation's rule is: "No direct model-to-tool path for consequential actions."7 The gateway is Big Widget's way to follow that rule. Whether it works for every route is still an open question.

When the test report is missing

Now Card E says: EV-17 cannot be found. An earlier check of Card R7 passed. Since then, nothing has changed: not the record, not the seal, not the list of trusted keys. What now? (This is case X in the bundle.)

This is different from a broken seal. A seal check can fail for more than one reason. The record may have changed. The seal itself may be wrong. Or the list of trusted keys may have changed. So when it fails, say which of these the check found, not just "it changed". Case M in the bundle is the first kind: a single word in R7 was changed, and the seal fails. In case X the seal is fine. It is the evidence the record points to that is gone. Call it "missing proof", and name who will find it.

Now try the third kind. Run case X with the newer key list, card-T-refetched.json. That list no longer has the gateway's key. Now the seal check fails, though the sealed part of the record did not change. The cause is the key list, not the record. Yesterday's "yes" only holds while everything the check uses stays the same.

One more list. Another group, OWASP, has its own list of ten risks for AI agents, dated 9 December 2025.8 Two of them, ASI02 "Tool Misuse and Exploitation" and ASI03 "Identity and Privilege Abuse", are close to SAIL 3.13.8 Putting them side by side is our idea, not OWASP's or Pillar's. Every crosswalk is someone's judgment.

A house rule is not the licence

Big Widget's crosswalk has a house rule: "We cite SAIL risks by id only." That is a choice. It is careful and simple. But it does not tell you what the publisher allows.

For that, read the licence. Pillar's sail-skill code library has a licence file. It is a Creative Commons licence called "Attribution-NonCommercial-ShareAlike 4.0 International".9 It lets you copy and share the material it covers, with conditions:9

Some things nobody here has shown:

When you don't know, write "not checked", and name who will find out. At Big Widget that is the person who handles legal questions. Don't guess, and don't treat the house rule as the licence.

StatementWhat kind?
"Cite SAIL risks by id only."House rule (made up for this course)
"Give credit when you share."Licence condition9
"This licence lets you use Pillar's logo."No: this licence does not give trademark rights. Any other permission: not checked9
"The licence covers the SAIL download."Not checked

The three yeses for parts-buyer

The Foundation names three steps: get to yes, stay at yes, recover to yes.7 For parts-buyer:

The AI never decides any of this. People write the rules, release holds and say when things are fixed.

Remember

The Foundation says "I don't know yet" "is always an answer, and it always pairs you with someone."6 So when you don't know, say who will find out.

Words to know: seal (signature), fingerprint (hash), locked fingerprint (keyed commitment), claim, test, crosswalk, risk id, licence, house rule.

Sources

  1. s2 Glacis Technologies, Inc. (2026). OVERT — Observable Verification Evidence for Runtime Trust (v1.1). Standard. Checked 2026-11-20 by source-checker role, separate Claude session cd3e777d (not the writer, not any earlier checker), pre-publish re-check 2 at 1004583 (home page opened 20 Nov 2026; all pre-publish paraphrases). ↩ ↩ ↩ ↩ ↩ ↩ ↩ ↩ ↩ ↩ ↩ ↩
  2. s5 Pillar Security (2026). SAIL (Secure AI Lifecycle) Framework V2. Official doc. Checked 2026-11-20 by source-checker role, separate Claude session cd3e777d (not the writer, not any earlier checker), pre-publish re-check 2 at 1004583 (landing page opened 20 Nov 2026; phase 3 description and 3.13). ↩ ↩ ↩ ↩ ↩ ↩ ↩
  3. s1 Ken Johnston; Bob Rapp, AiGovOps Foundation (2026). What is OVERT?. Official doc. Checked 2026-11-19 by source-checker role, separate Claude session 10396a05 (not the writer), round-3 re-check at 17ce1d5. ↩ ↩ ↩ ↩
  4. s9 Glacis Technologies, Inc. (2026). OVERT governance. Official doc. Checked 2026-11-20 by source-checker role, separate Claude session cd3e777d (not the writer, not any earlier checker), pre-publish re-check 2 at 1004583 (sections 1, 7, 8 and footer opened 20 Nov 2026; revised paraphrases). ↩ ↩ ↩ ↩
  5. s8 Glacis Technologies, Inc. (2026). OVERT — Observable Verification Evidence for Runtime Trust, Version 1.1 (June 2026) (full text). Standard. Checked 2026-11-20 by source-checker role, separate Claude session cd3e777d (not the writer, not any earlier checker), pre-publish re-check 2 at 1004583 (full markdown opened 20 Nov 2026; title heading and every quoted span). ↩ ↩ ↩ ↩ ↩ ↩ ↩ ↩ ↩ ↩
  6. s11 AiGovOps Foundation (2026). The AiGovOps Foundation café — start with your problem (Practice home page). Official doc. Checked 2026-11-19 by source-checker role, separate Claude session 10396a05 (not the writer), round-3 re-check at 17ce1d5. ↩ ↩
  7. s10 AiGovOps Foundation (2026). The AiGovOps Practitioner Manifesto. Official doc. Checked 2026-11-19 by source-checker role, separate Claude session 10396a05 (not the writer), round-3 re-check at 17ce1d5. ↩ ↩ ↩ ↩ ↩ ↩
  8. s7 OWASP GenAI Security Project (2025). OWASP Top 10 for Agentic Applications for 2026. Official doc. Checked 2026-11-19 by source-checker role, separate Claude session 10396a05 (not the writer), round-3 re-check at 17ce1d5. ↩ ↩
  9. s6 Pillar Security (pillar-labs) (2026). SAIL skill licence — Attribution-NonCommercial-ShareAlike 4.0 International. Record. Checked 2026-11-19 by source-checker role, separate Claude session 10396a05 (not the writer), round-3 re-check at 17ce1d5. ↩ ↩ ↩ ↩ ↩ ↩ ↩ ↩ ↩ ↩ ↩ ↩

Next: the quiz