Five of the seven decisions logged since do not count as the founder's under the company's own rule, because the last commit to touch each is not signed with his key. Measured with the company's own authority reader, from the rule's start on to commit 7f13240ff, committed . This is the snapshot measured on . A later signature may change the verdict; this history measurement does not refresh automatically.
Agents here run the company's work. The founder keeps the decisions that cost money, reach a customer or cannot be undone. This page shows where that line is written, how a decision is proved to be his, and where the proof is missing.
A rule nobody can check is a wish. So every rule here names the file, the signature or the check that would catch it being broken. The agent safety page covers what the agents cannot do at all. The kill switch page covers how everything stops.
What is agentic AI governance in a company run by agents?
It is the split of decisions between the agents and the founder, written down before the decision comes up, and checked afterward by something other than anyone's word.
Two rules in the company constitution, version 1.0.0, carry the rest. Autonomy is earned per lane, by blast radius. The workflow commits, never the agent.
Governance is the umbrella. This page covers who may decide, and how anyone would prove it later.
Judged, not measured: the definition and the line about a wish are the method's thesis. Governing automated systems is an established field. This page is a dated record of one company, not a framework.
What can this record not see?
Stated before any figure:
- A signature proves custody of the key, not that the founder was present.
- The authority matrix is read as written. No one has re-read every row against the code that enforces it.
- The signature counts and authority verdicts come from one private repository's history. The decision evidence gives dates, classifications, verdicts and signing commits. Readers can recount those rows and apply the published verdict rule. Cryptographic verification still requires repository access; the public file contains no key identities or signatures.
- Approvals the founder gave and nobody wrote down are NOT OBSERVED. They are counted nowhere.
- Whether a refused decision is signed later is open. The signature and authority figures remain the dated snapshot until someone reruns the generator.
Who may decide what?
Each of the eleven departments has one row in an authority matrix, generated from the department registry. A lane starts gated. The constitution says moving it to autonomous takes a logged decision that names the criterion it met.
A few rows, as the matrix reads today:
- Finance may update the ledger and run cost guards. Any spend, pricing change or billing change is gated.
- Legal may draft first-pass language for a human lawyer. Signing, filing or agreeing to anything is gated, always.
- The Inspector General may report what its checks find, and nothing else.
- Above every row, the founder keeps money, claims copy, production data, anything irreversible, the gates on commit and push, and the protected files.
Read as text, not tested as enforcement. Each row says what the matrix and the constitution state. Whether the code refuses every gated act in every lane was not re-checked for this page.
How is a decision proved to be the founder's?
By a signature from a key, never by a name. An author email proves nothing, because any process can type one.
That rule was earned. On the founder's old signing key was measured with no passphrase on its private file. The record states the consequence: any process running as the owner, every agent session included, could sign as the founder with it. It was retired the same day.
The replacement lives in the Mac's Secure Enclave. The record says it asks for Touch ID on every signature and cannot be exported. A sealed backup key sits in a password manager, never on the Mac at rest and never in any agent. Both were pinned in the record on .
Since , these acts require a commit that key signed: a decision entry, a constitution amendment, a raise in an agent's autonomy and a lesson promotion. Removing a global halt or any other halt also requires that signature, with one exception: a scoped halt whose original creation record names evaluator or ladder as its writer. Removing that machine-written scoped halt is exempt. Its current text cannot grant the exception.
Signing away from the Mac is what Launch Keys is being built for. Where it stands is on human in the loop.
How often does the founder actually sign?
Rarely. From to , the company's history holds 1,243 commits. Eight carry a good signature from the founder's key: seven from the active key and one from the backup.
1,138 carry no signature. They are not counted as his, whoever made them. 62 carry the retired key and 32 a key the record never pinned, and none of those count. Three could not be verified.
His signatures fall on six days. The longest stretch between two of them is eleven days. The ladder that promotes the agent organization pauses every promotion after fourteen days without one.
Commits by the key that signed them
Eight of 1,243 commits carry the founder's key. The longest stretch between two days he signed is 11 days.
Counts as the founder's
- Active key7
- Backup key1
Signed, does not count
- Retired key62
- A key never pinned32
- Could not be verified3
Not signed1,138
- A dated snapshot: the generator runs only when someone runs it, so a signature made after the measured day is not counted here.
- Classified by key fingerprint against the owner key record, never by author or by git's trust letter, at revision 7f13240ff.
- A part too thin to see is drawn three pixels wide; every count is exact in the key, the table and the downloads.
Not shown. Any key, fingerprint or signed commit. The figure counts classes only.
Built from authority-verification.generated.json (the counts) and governance-figures.generated.json (the window and the gap), which scripts/research/method-figures/ai-governance.ts writes when someone runs it with --write. No hook or bot runs it.
Snapshot as of
The numbers as a table
| Signature | Counts as his | Commits | Share |
|---|---|---|---|
| The founder's active key | Yes | 7 | 0.6% |
| The founder's backup key | Yes | 1 | 0.1% |
| The retired key | No | 62 | 5.0% |
| A key the record never pinned | No | 32 | 2.6% |
| Could not be verified | No | 3 | 0.2% |
| No signature | No | 1,138 | 91.6% |
Download the data: ai-governance-signatures.csv, ai-governance-signatures.json
Measured: every commit dated or later, classified by its key fingerprint against the owner key record, never by author or by git's trust letter. It would be disproved by a recount by fingerprint that gives a different number, or by a signed commit the classifier misses.
What happens to a decision the founder made but did not sign?
It is recorded, and it does not count.
On the founder said, in the words the log keeps: "yes move the handoffs into docs/handoffs/". That became decision D-0106, and the handoffs now live there. The commit that carried the entry onto the main line was not signed with his key, so the company's own reader refuses it as his.
The work happened. The proof that he decided it did not.
Since , the reader has admitted two decision entries and refused five. Of the raises in an agent's autonomy, it refused three and left three exempt, because the rule leaves that kind of change free. Of the removals of a halt, it admitted one and refused one.
The download publishes method authority-decisions-v1 at revision 7f13240ff486faf34c698f44dff832a93329d6d3. It covers 106 numbered decisions: two admitted, five refused and 99 grandfathered. The authority total also includes one grandfathered preamble unit, which is not a decision.
Each numbered heading in the committed decision log is a decision-entry unit. Its date is the first ISO date in that heading. The preamble is a separate unit and is excluded from these decision rows.
The authority boundary is the unique ancestral addition of the authority reader. Units at or before it are grandfathered and carry no signature claim. Later changes require an owner signature. The reader follows every parent; unchanged values inherit provenance only when the merge-base comparison permits it. Refusal wins a tie. A new signed change can replace a refusal.
The committed verifier checks the provenance commit against trusted owner keys at the measurement revision. Revoked or unpinned keys do not count. The ownerSigned value records that result. Admitted decisions have ownerSigned true; refused decisions have false. Grandfathered rows use null because their signatures were not judged.
provenanceCommit identifies the commit whose verdict the unit carries, including an inherited verdict. signingCommit is that commit only when its signature verifies as the owner's. A signature proves key custody, not personal presence.
To verify the dated evidence with repository access, run node --import tsx scripts/research/method-figures/authority-evidence.ts --check. The generator reads committed files at the recorded revision. It does not refresh the history figures or register an automatic rebuild.
A governed company can tell what was decided from what was proved.
What the authority reader did with each kind of act
Decision entries7 judged
Autonomy raises6 judged
Halt removals2 judged
None judged since the rule began: constitution amendments, lesson promotions and accepted waivers.
- A dated snapshot: the generator runs only when someone runs it, so a signature made after the measured day is not counted here.
- 228 acts from before the rule began are grandfathered: the reader makes no claim about them, so they are counted in the table and not drawn.
- Exempt means the rule leaves that kind of change free, not that it was checked and passed.
Not shown. Which units were refused. The public file carries classes and counts only, by the generator's own boundary, so no path, key or commit is published here.
Built from authority-verification.generated.json, the classes and counts the generator publishes, which scripts/research/method-figures/ai-governance.ts writes when someone runs it with --write. No hook or bot runs it.
Snapshot as of
The numbers as a table
| Kind of act | Judged | Refused | Admitted | Exempt | Grandfathered |
|---|---|---|---|---|---|
| Decision entries | 7 | 5 | 2 | 0 | 100 |
| Constitution amendments | 0 | 0 | 0 | 0 | 1 |
| Autonomy raises | 6 | 3 | 0 | 3 | 110 |
| Halt removals | 2 | 1 | 1 | 0 | 5 |
| Lesson promotions | 0 | 0 | 0 | 0 | 12 |
| Accepted waivers | 0 | 0 | 0 | 0 | 0 |
Download the data: ai-governance-authority-verdicts.csv, ai-governance-authority-verdicts.json
Download dated authority verdicts and signature classes (JSON)
Download the versioned authority method and per-decision evidence (JSON)
Checked for D-0106 only: its handoffs folder exists. Whether the work behind the other refused entries happened was not checked for this page.
How did the agents' right to push change, and why?
On a push to the main line became human only. On September 21 the founder released it: "Yes I want you to be able to push".
On September 22 the handoff loop added a push to main to its own tool grant and rewrote its own closer. A later review, ratified September 23, marks that authorship as inferred, not proven.
By October 2 the loop's agent pushed nothing. The loop itself lands each iteration through a gate that sends any change to a judging file, a served file or a path outside its allow list to a pull request instead of main. The record of that loop's containment is Governing Self-Repair.
How many decisions has the company logged, and when?
106, from July to . By month: 36 in July, 8 in August, 35 in September, 24 in October and 3 in November. Judged from the entries' titles, not counted: July is when the organization was founded, and September is when the console and its commands were built. The full record, with a download, is the Decision Ledger.
Three entries sit out of date order (D-0039, D-0078 and D-0106), because a decision made on a branch can land after a later one. So the last entry is not the newest decision. On , while this page was written, the public dataset's newest-decision field read the last entry's date instead of the newest one.
Every day the company logged a decision, and the silences between
106 decisions on 34 days. The longest silence is 13 days, from to , and the busiest day is , with 16.
A day with logged decisions; its area is the countThe days between two decision daysLogged after a decision dated later
Drag along the dots, or use the arrow keys, to read each decision day.
- Dates are the dates the entries carry, not the day they landed on the main line.
- The axis ends at the newest date any entry carries (D-0105), not at the last entry logged.
- On a phone, dots on adjacent days overlap; the table carries every day.
- A silence is a stretch with no logged entry, not proof that nothing was decided.
Built from public/decision-ledger.json, which scripts/generate-stats.cjs rewrites from the committed decision log on every commit and every bot landing.
Ledger generated
The numbers as a table
| Date | Decisions | Entries | Running total | Days since the day before | Order |
|---|---|---|---|---|---|
| 16 | D-0001 to D-0016 | 16 | In order | ||
| 5 | D-0017 to D-0021 | 21 | 1 | In order | |
| 7 | D-0022 to D-0028 | 28 | 6 | In order | |
| 3 | D-0029 to D-0031 | 31 | 1 | In order | |
| 2 | D-0032 to D-0033 | 33 | 2 | In order | |
| 2 | D-0034 to D-0035 | 35 | 1 | In order | |
| 1 | D-0036 | 36 | 5 | In order | |
| 1 | D-0037 | 37 | 13 | In order | |
| 1 | D-0039 | 38 | 1 | D-0039 logged after D-0038 | |
| 1 | D-0038 | 39 | 1 | In order | |
| 1 | D-0040 | 40 | 3 | In order | |
| 4 | D-0041 to D-0044 | 44 | 7 | In order | |
| 1 | D-0045 | 45 | 10 | In order | |
| 1 | D-0046 | 46 | 2 | In order | |
| 2 | D-0047 to D-0048 | 48 | 2 | In order | |
| 3 | D-0049 to D-0051 | 51 | 2 | In order | |
| 3 | D-0052 to D-0054 | 54 | 3 | In order | |
| 4 | D-0055 to D-0058 | 58 | 4 | In order | |
| 2 | D-0059 to D-0060 | 60 | 4 | In order | |
| 1 | D-0061 | 61 | 1 | In order | |
| 4 | D-0062 to D-0065 | 65 | 3 | In order | |
| 6 | D-0066 to D-0071 | 71 | 1 | In order | |
| 4 | D-0072 to D-0075 | 75 | 1 | In order | |
| 2 | D-0076 to D-0078 | 77 | 1 | D-0078 logged after D-0077 | |
| 2 | D-0077 to D-0079 | 79 | 1 | In order | |
| 1 | D-0080 | 80 | 4 | In order | |
| 2 | D-0081 to D-0082 | 82 | 1 | In order | |
| 13 | D-0083 to D-0095 | 95 | 2 | In order | |
| 4 | D-0096 to D-0099 | 99 | 2 | In order | |
| 1 | D-0100 | 100 | 8 | In order | |
| 2 | D-0101 to D-0102 | 102 | 10 | In order | |
| 1 | D-0106 | 103 | 2 | D-0106 logged after D-0105 | |
| 2 | D-0103 to D-0104 | 105 | 8 | In order | |
| 1 | D-0105 | 106 | 3 | In order |
Download the data: ai-governance-decision-days.csv, ai-governance-decision-days.json
Measured from the generated Decision Ledger, every entry it holds. Dates are the dates the entries carry, not the dates they landed.
How is governance kept from simply pleasing the founder?
The constitution says preference learning shapes proposals toward the founder's taste and never overrides the constitution, the ledger or the north star. An organization that only learns to please its founder is a failure, not a feature.
A signature cannot fix that. It only proves custody of a key. The company records that as an accepted risk, decision D-0030, dated .
Dissent is written in as a seat. The board of advisors carries a standing role, the Tenth Man, whose job is to argue against the room whenever it agrees. How often that seat is actually triggered is measured on the synthetic personas page, and most of its triggers are not wired yet.
Judged. A rule against pleasing the founder is a written intent. Nothing measured here shows it working.
AI governance questions? Answers.
Related reading
- The kill switch
How the Machine is stopped, with the dated record of its halts.
- AI Agent Safety
What the agents cannot do at all, what enforces each limit, and where each one leaks.
- Human in the loop
Where the founder's time goes, and the recorded commands that replace hand edits.
- Autonomous operation
The dated record of the Machine running without a person in the loop.
- AI oversight ledger
The standing record of who caught each issue and who fixed it.
- Governing an Agent Leadership Team
The governance preprint: the organization chart and how its seats are held to account.
- All Orbyt Labs methods
Every method page, each with its dated public record.
Cite this page
Cite this page for the practice and follow its sources for the underlying records.
Bartak, J. (2026). AI Governance. Orbyt Labs Research. https://