The moat isn't the model.
It's who's allowed to grade it.
Every funded incumbent in AI assurance is run by — or paid by — the operator it grades. We're the one that structurally can't be. That's a position they can't copy without unwinding their own business.
"The company that builds the AI cannot be the company that certifies it worked."
Enterprises and agencies are deploying AI faster than anyone can prove it works.
Observability tells you the model ran. Governance tells you it was allowed to. Neither answers the question the CFO and the agency actually asks: did it lower the true cost of getting the work done — or just move the cost somewhere the dashboard doesn't show?
Regulation is converging on the answer that this needs to be checked by someone who didn't build it. The EU AI Act (Art. 31) requires conformity assessment bodies to be independent of what they assess. US federal policy (OMB M-25-21) now requires independent post-deployment review of high-impact AI. The demand for a neutral grader is being written into law — and there isn't a clean one yet.
Four reasons an incumbent can't just add this
01
Structural neutrality (Art. 31)
The standard for a conformity body is independence from the thing assessed. A vendor that builds, hosts, or advises on the AI fails that test on day one — no matter how good their tooling is.
02
Buyer-pays, not issuer-pays
2008 taught the market what issuer-pays ratings are worth. We're paid by the party that wants the truth (the buyer), not the party being graded (the vendor). An incumbent monetizing the operator can't flip that without killing its own revenue.
03
Reproducible, not narrative
The verdict is a method, not an opinion — same inputs, same answer, re-runnable by a third party. That's true today, before a single client. It's what makes the rating a falsifiable instrument rather than a consultant's slide.
04
Clean independence story
No platform business to protect, no vendor partnerships to preserve, no operator dollars on the cap table. The neutrality is real because there's nothing structurally pulling against it — and that's a constraint we intend to keep, including in who funds it.
Good companies — structured around a conflict we don't have
The space is real and funded. The point isn't that these are bad products — several are excellent. It's that each is structured around the operator in a way that makes a neutral, buyer-paid outcome verdict awkward or impossible for them to issue.
| Who's in the space | The structural position they're built around |
|---|---|
| Governance & observability Credo AI, Holistic AI, Fiddler, Arize, ValidMind |
Operator-run and operator-paid — sold to and configured by the team running the AI. Strong tooling; not a neutral outside grader of the economic outcome. |
| Security & runtime defense Cisco AI Defense, Protect AI |
Operator-embedded in the stack they protect. Answers "is it safe / attacked," not "did it lower true cost-to-resolve." |
| Independent-leaning auditors Eticas, Resaro |
Closest in spirit — genuinely independent. But they audit process, bias, and conformity, not the dollar outcome, and the work is bespoke rather than a reproducible instrument. |
| Issuer-pays warranty Armilla (Lloyd's-backed) |
Insures the model — paid through the issuer side. Useful risk transfer; structurally the issuer-pays position 2008 already discredited for ratings. |
Capability-level comparison from public information. No disparagement intended; several are strong companies.
The open seat
Independent · buyer-paid · outcome-anchored · reproducible.
No funded incumbent occupies all four at once — because doing so would unwind how they make money.
A falsifiable rating, not a consultant's opinion
The verdict grades an AI-handled resolution across 16 fault lines and attributes any failure to a specific party — data, model, integration, or the human queue. Same inputs, same method, same answer. Here's the shape of one (illustrative, conditional voice):
Did the AI work?
Would not hold
A contact like this would close as "resolved" — then reopen 9 days later.
True cost to resolve
~$214
reported: $38
Fault attribution
Data
✓
Model
✗
Operator
✓
Handoff
✗
One dated event re-rates this company
Now
Non-dilutive in flight
Grants, credits, and accelerator paths that don't touch the cap table.
Unlock — next
First design partner
One named buyer turns the method into a live, real-data engagement.
Proof
One real verdict ships
A reproducible CPR verdict on real data, with a reference behind it.
Then
Re-rate
Category-defining instrument with a proof point — priced off that, not this page.
Pre-revenue. Zero outcomes rated publicly. One dated event away from changing both.
Commercial fastest. Federal first-class.
Commercial
The Exposure Review is live today — a buyer brings their volume, FCR, and cost and gets their own true-CPR exposure, no PHI. Shortest path from interest to a scoped pilot.
Federal
M-25-21 makes an independent post-deployment read a requirement, not a nice-to-have. A 30–60 day Impact read on de-identified data is its own first-class lane — not a thinner version of the commercial one.
Category pull, not revenue: an accepted Federal News Network op-ed on AI accountability (publishing mid-2026), a live dialogue with VA AI leadership, and a caregiver-of-a-veteran founder fit. Signals that the category is forming — stated as demand, never as outcomes.
I measured AI economics as a senior analyst before building this, and I care for a disabled veteran — my father. The thread is the same in both: a system that reports a problem solved when it isn't costs someone real money or real dignity downstream. MarginSignal OS exists to put a reproducible, independent number on that gap — and to stay the one party with no reason to fudge it.
— Brandon Burdin, Founder, MarginSignal OS
The shape of the next step
Non-dilutive first, then a design partner, then a deep-tech pre-seed — on the order of a ~$750K SAFE around a ~$6M cap, priced off a shipped verdict and a named reference, not off this page. The neutrality constraint extends to the cap table: no lead capital from a party we'd rate — no AI vendor, BPO, or insurer in the lead. We're not asking today.
The incumbents grade their own work.
We don't have to.
Talk to the founder
No raise banner. No deck on the public web. Just the founder, the moat, and a dated catalyst.