The Moat · Structural Neutrality

The company that builds the AI can't be the company that certifies it worked.

That isn't our slogan. It's the rule the whole category is being rebuilt around. Every other AI-assurance vendor is run by — or paid by — the operator they grade. We're not. And we made that a structure you can hold us to, not a sentence you have to believe.

Monitoring proves your AI ran. It can't prove it worked.

Three layers sit between "the AI ran" and "the AI worked." Only the top one is independent.

Observability
OpenTelemetry · OpenInference · LangSmith — "Did it run?"
Operator-paid
AI governance
"Is the model well-behaved?"
Operator-paid
Independent verdict — MarginSignal OS
"Did it work — and whose fault if it didn't?" The layer on top, not a rival.
Buyer-paid

We don't replace your monitoring. We turn its output into evidence someone outside your building will accept.

Independence isn't our idea. It's becoming the rule everywhere it matters.

EU AI Act, Article 31

A body assessing high-risk AI must be independent of any operator with an economic interest in it — and may not advise the parties it judges.

SEC / PCAOB auditors

An auditor may hold no financial interest in, and provide no conflicting services to, the entity it audits. Independence is the whole job.

The 2008 ratings lesson

When the rated party pays for — and can shop — its own rating, the rating stops meaning anything.

We don't cite these because we have to. We cite them because we already built to them.

Why the funded incumbents structurally can't do this.

The structure most assurance vendors are structured around

  • — Operator-run or operator-funded — the graded party buys and runs the grader
  • — Sells advisory to the same parties it assesses
  • — Issuer-pays warranty or rating — profits from the verdict
  • Each a sound business. None structurally independent.

The structure we chose

  • Buyer-pays — the side that consumes the verdict pays for it
  • No equity, board seat, or advisory from any party we measure
  • An independence covenant in our governing documents
  • Copyable only by unwinding their own model. That's the moat.

The rule (non-negotiable)

1 · No financial entanglement with the parties we measure.

No equity, board seats, lending, or paid consulting from any party whose AI systems, agents, or outcomes we measure — operators, model vendors, BPO/CX providers, or insurers underwriting the risks we attribute. Any such funding is capped below control, granted no board seat and no methodology rights, and disclosed.

2 · Buyer-pays, not issuer-pays.

We charge the side that consumes the verdict. Where a measured party funds a specific assessment, the fee is fixed, published, and paid regardless of the result.

3 · We verify the system and the vendor — never your team.

Our role is a shield for the buyer, not an indictment of their people. If it failed, that's the vendor's or the model's problem to answer for — not your budget to defend.

4 · A structural firewall.

An independence covenant in our governing documents; methodology set by an independent committee, published and versioned; funding sources disclosed.

5 · No fabricated results. No undisclosed conflicts. Ever.

We do not claim clients or results we do not have, and we disclose any relationship a reasonable party would want to know.

We have zero outcomes rated publicly today — and we'd rather say that plainly than fake one. The first reproducible verdict is the catalyst. The honesty is the credential until then.

Don't trust the verdict. Re-run it.

A verdict is only worth something if a skeptic can reproduce it. Ours are externally anchored and recomputable by a party that doesn't trust the operator — the property neither a vendor's seal nor an insurer's badge can claim.

Externally anchored
Not the operator's word for it.
Independently reproducible
A skeptic re-runs it.
Cross-party attribution
Data · model · operator · agent.
Structurally neutral
No stake in the result.

Scope & governance

This policy binds the company, its founder, employees, contractors, and advisors. It is reviewed at least annually and published here. Material changes are versioned and dated.

Version 1.0 · 2026 · MarginSignal OS LLC

Partners and funders evaluating this structure can talk to the founder →

Independence you can verify — not just trust.

The proof holds up precisely because it isn't ours to bias.

Book a Findings Call