Confidential engagement · Rebuilt for display
A governed AI operating system
Built for a company whose business units had grown apart and needed to speak with one voice again. The client is under a confidentiality agreement, so what you are looking at is a rebuild of the machinery with every fact replaced.
This engagement is confidential. Rather than redact a real repository, which still leaves you holding a copy of one, I rebuilt the machinery from scratch under an invented client called Meridian Works. Every name, number, path and detail below is fabricated. Nothing here can be checked against anything real, and that is deliberate. What is genuine is the method: how the work is structured, how it verifies itself, and what it refuses to do.
Click any folder to open it, any file to see what it holds and why it exists. This is the real structure, with the contents described rather than reproduced.
Eight sections, roughly half the length of the first system I built, governing considerably more.
Who this serves
Read this first and in full, and do not crawl the folders to orient yourself.
That anti-crawl line is in all four of my systems. It is the cheapest performance improvement available: an agent that reads the map opens four files where an agent that explores opens forty.
Load order
Constitution, behavior, hard limits. Then only what the task needs.
The folders
One line each.
Trust order
Approved beats facts beats drafts. Unchanged since the first build, because it was right the first time.
Routing
Request type to exact file list. In this build the table is generated from the installed capability list rather than maintained by hand, so it cannot describe something that is not there.
Hard rules
Never invent. Placeholder over fabrication. Nothing external without approval. Never edit an approved file in place. Archive rather than delete.
Output checklist
Runs before every handover, and terminates in a deterministic check rather than in a person saying they looked.
Governance
Owners, approvers, and the lifecycle.
The guardrails, framed the way they actually work: as things the system refuses to do.
Produce anything shareable while under embargo.The system refuses to generate external material until a stated date. The refusal is a rule in the constitution, not a reminder in a document someone might not read.
Invent a commitment.Service levels, guarantees and terms come only from approved source material. This is the rule the verification fleet caught being broken, inside two finished drafts.
Merge two business units' contradictory facts.When parts of a company disagree on a fact, the system keeps both and surfaces the conflict rather than silently choosing.
Grade its own work.Verification always runs in a fresh context. Given the reasoning, a model will tend to agree with it.
Accept a self-report as completion.Handover is gated on a deterministic check exiting clean. An agent saying it is finished is not a signal.
Use a value that is not in the token sheet.No raw hex, font name or logo appears inline anywhere. Change the token table and the whole system re-renders.
Real source, copied out of the repository that runs. Each excerpt prints the file it came from, and any trimmed lines are marked rather than quietly spliced.
The interesting capabilities are not the generators. They are the ones that check the work.
graph-method
How to design multi-agent work. Fan out, reduce, verify, synthesise, and the tests that tell you the shape is wrong.
acceptance-test
A short set of questions that prove an install works. Several of the correct answers are refusals.
pre-delivery-scan
Deterministic scanner. Reads its rules from markdown and exits non-zero on a violation.
drift-audit
Checks every generated file against its source, including the constitution itself.
twin-sync
Regenerates the rendered twin of every source file so the two cannot diverge.
portal-build
Generates the browsable portal from the system, in multiple variants from one template.
identity-audit
Audits a brand system against its coverage contract and reports what is missing.
decision-log
Files a decision with its reasoning, so the same argument is not had twice.
question-register
Captures open questions with an owner, and refuses to let them be closed by inference.
The design method, which is the part of this build I would most want to be judged on. It is an approach rather than a script, and it applies to any system.
The fake-edge test
If you cannot name the specific artifact that travels along an edge between two agents, the edge is imaginary and the two nodes are really one node.
Most over-engineered agent graphs are a single prompt wearing a costume. This test kills them before they get built.
Fan out, reduce, verify, synthesise
Fan out for coverage. Reduce to deduplicate before anything expensive runs. Verify independently. Then one synthesis pass.
The reduce step is the one people skip, and it is the one that stops you paying to verify the same finding five times over.
Different lenses, not more passes
When something can fail in more than one way, give each verifier a different angle rather than running the same check repeatedly. A handful of genuinely different perspectives will find what a large number of identical ones never will.
Fresh context, always
Verifiers never see how the artifact was produced. Shown the reasoning, a model tends to agree with it.
Deterministic anchors
Every graph terminates in something that either exits clean or does not. A machine check passing is evidence. An agent reporting success is not.
Named failure modes
The recurring ways these graphs go wrong, documented with standing fixes so the next person does not rediscover them: the verifier that agrees with itself, the merge step that loses track of where a finding came from, and the graph that is really a chain.
Meridian Works, the invented client. Real structure, invented values. Every color carries its measured contrast, and where it fails the failure is published as a rule rather than nudged away.
Every change is logged with its reason. A change is not done until the system agrees with it.
LaterAdversarial self-audit
Pointed the verification fleet at my own install. Most of what it flagged was refuted by independent judges. What survived was real.
The worst finding: my own pre-delivery scanner was a rubber stamp. Its rule parser was loading nothing, so it passed every document, and it had been reporting green for days.
Fixed, and the drift auditor extended to cover the constitution itself so the same class of failure cannot recur silently.
MidThe routing table became generated
Marked the routing table as generated output, emitted from the installed capability list. The constitution can no longer describe something that is not there.
EarlyVerification fleet
First real run. It caught a service commitment that appears in no source document, sitting in two finished customer-facing drafts that had already been approved by a human.
That single catch justified the whole build.
StartConstitution and skeleton
Eight sections, adapted from three previous builds rather than written fresh.
Adaptation converges. Invention drifts.
| Path | What it holds | Size |
|---|
Next
Four builds, one architecture
This is the newest. Set beside the first, the distance between them is the argument.
The layer under it
How every one of them runs.
The operating loop, the guardrails, the agent roster and the rest of the machinery, in one place.