Why I am building Nebula

I came to Australia in 2018. No family here, no network, nobody to call.

I had an engineering degree and years of work behind me. Neither counted for much. Australian practice is regulated, and my qualifications were not written in the terms that system reads. So I learned it while I worked. Two Masters. I am RPEQ now, and MIEST.

Same person, same ability. What changed was whether it counted.

The system was not the problem. Regulation is the reason one engineer's signature means something to a stranger who has never met him. But I know what it costs to become readable to one, and I know how much real capability is sitting outside one right now, doing nothing.

This happens on every project

Documents die the day they are signed. The work keeps moving.

So somebody rebuilds the forecast by hand every cycle, patches the tracker, writes the report. The agreement comes off the shelf only when there is an argument.

In nearly all of those arguments, somebody knew. The supervisor knew the sequence had changed weeks before the claim came up short. It did not count. It was not in a form the other side had to accept.

That is 2018 again.

The part nobody owns

     ORGANISATION A                        ORGANISATION B

     +--------------+                      +--------------+
     |   finance    |                      |   finance    |
     |   documents  |                      |   documents  |
     |   schedule   |                      |   schedule   |
     +--------------+                      +--------------+
            |                                     |
            | knows what                          | knows what
            | happened in here                    | happened in here
            |                                     |
            +------------[ AN ACT ]---------------+
                              |
                              v
                      was this allowed?

                    nobody answers this

Microsoft and Salesforce both built serious agent governance this year. Microsoft keeps credentials out of the agent completely and issues a fresh single-use token for each tool call. Credentials never enter the model's reasoning at all. Salesforce governs an organisation's whole agent estate from one control plane, third-party agents included, and makes an agent act with a named user's permissions rather than broad access.

Both are good work. Both stop at the same line. They settle whether an act was authorised inside the organisation that ran the agent. Neither settles whether the counterparty has to accept it.

An act can be perfectly proper inside the company that performed it and still be worth nothing to the party on the other side of the agreement, who runs different systems, never saw that policy, issued none of those permissions, and may be about to dispute the lot.

The receipt

The thing I want to build is a receipt. After a consequential act, by a person or by an agent, Nebula says who did it, for whom, under what grant, on what evidence, and why it was allowed or refused.

Invented engagement, real shape:

ACTION             Mark package EL-204 complete
REQUESTED BY       Commissioning agent, for J. Whelan, Site Manager
PRINCIPAL          Northline Energy Pty Ltd
ENGAGEMENT         Kwinana storage, stage 2
GRANT              completion:write
                   issued 14/09/2026, expires 30/09/2026, not revoked
POLICY             Completion policy v4
EVIDENCE REQUIRED  Test record accepted by the principal's engineer
                   Energisation confirmed
                   No unresolved hold point
EVIDENCE HELD      Test record supplied 21/09/2026, NOT ACCEPTED
                   Energisation confirmed 20/09/2026
                   Hold point HP-11 unresolved
OUTCOME            REFUSED
REASON             Evidence supplied but not accepted. One hold point open.
                   The grant was valid and the act was within scope.
AT                 22/09/2026 09:14 AWST
READABLE BY        Northline Energy, Kwinana Civil

Read the refusal again. Valid grant. Act within scope. Still no, because supplied and accepted are different states.

Nebula does not do this yet. I would rather write that than fudge it.

Five cases have to pass, four of them refusals: no grant, evidence not accepted, grant stale or revoked, the authorised act, and authority changing while the act runs. On a real engagement, both sides able to read it. When it passes I will post it here.

Where it stands

Nebula holds no customer engagements today.

What runs: Nebula reads source documents into a model where a term traces back to the document, revision, page and character span it came from, and to which model read it. Observed events are held apart from agreed facts. Where two parties record conflicting positions on what was agreed, both are kept and the matter is not called settled. A model cannot author a party's intent or position; that is refused at the writer, not asked for politely in a prompt.

What does not run yet: the authority layer, and the receipt above.

What I think happens next

Dated, so they can be held against me.

By 27/08/2027. An agent takes a legitimate-looking but commercially wrong action in a real engagement, publicly, and the argument afterwards is about who authorised it rather than whether the model was accurate.

By 27/08/2027. Selling context to AI stops being a business. Models will reconstruct project state from documents on their own. What they cannot reconstruct is who held authority, what was agreed rather than asserted, and what evidence stands behind a claim.

By 27/08/2029. Counterparties start asking for an authority record before they accept an agent-performed act, in the same paragraph of the contract as the insurances.

By 27/08/2036. Acting without a machine-readable record of the authority behind it looks the way an unsigned drawing looks now. Not illegal. Just not acceptable practice.

If I am wrong, this is how: legitimacy turns out not to matter commercially, and everyone keeps accepting acts on trust and arguing later, because it is cheaper. I have taken the other side of that.

What I will be measured on

Targets, not probabilities. I took the confidence column out on purpose.

DATE           OPERATORS  PAYING ORGS  ENGAGEMENTS     RECEIPTS
27/08/2026             0            0            0            0
27/02/2027            25            5           30            1
27/08/2027           250           50          400        2,500
27/08/2029         6,000          800       12,000    1,000,000
27/08/2036     1,000,000       60,000    2,000,000  250,000,000

First row is today.

An active operator acted on at least four separate days in the month, on an engagement holding a document their organisation supplied and a requirement read from it. Supplying a source, correcting a reading, moving an item, recording evidence. Signing in is not an act. Staff, test organisations and the worked example never count.

I want this in the hands of the people who hold the work together and are still fixing the tracker at eleven at night. That was me.

I will post against this table on 27/02/2027, met or not.