What I'm willing to claim — and what I won't.
Most of what gets said about AI can't be checked. That's the genre's defining problem: confident sentences, no way to tell which ones are true. I run a lab and a consulting practice on the opposite bet — that the most useful thing I can hand a skeptical reader is a claim you can verify, and a clear account of the claims I won't make. This page is that account. It's the standard every other page on this site is held to.
Four kinds of claim
I sort everything I say into one of four kinds, and I mark which is which so you don't have to guess:
- Demonstrated
- something I've actually shown. Every demonstrated claim points to an artifact that exists: a run log, a test suite, a named client. If I can't show it to you, it isn't demonstrated.
- Plausible
- something built or designed for a purpose, with sound reasoning and partial evidence, but not yet proven. The verbs give it away: built for, designed to.
- Aspirational
- an intention, stated as one: I'm preparing, the plan is. Not a capability.
- Unbuilt
- named precisely so it's never mistaken for something that exists. The things I haven't done are part of an honest picture, so I say them out loud.
The companion note runs these four tiers across my own projects, by name — including the parts that don't work yet. That's the worked example; this is the rule. What we refuse to claim→
What I refuse to claim
A few refusals are absolute, because they name the ways this kind of work usually goes wrong:
- No prediction. The simulation work maps how a decision could go; it does not tell you how it will. Anyone selling a forecast of a genuinely contingent situation is selling a fiction.
- No capability I haven't shown. A thing the method is built for is not a thing it has done. I keep those words apart on purpose.
- No borrowed interest. No "organizations are looking at it" without naming them. If I can't name it, I don't claim it.
- Numbers trace to a source. Every figure on this site comes from a file I can open, re-checked against that file before it ships — not from the last time I happened to say it.
- Status matches reality. When I describe what a project is, I describe what it is today, not what I hope it becomes.
Independence
I take no referral fees or commissions from any AI vendor. I have nothing to sell you on a model provider's behalf, which means I can tell you the truth — including that the right answer is sometimes a person, a spreadsheet, or nothing at all. The strongest bias in this field is the reflex that AI is always the answer; taking none of the vendors' money is how I keep it out of the advice I give you.
When I'm wrong
I would rather publish a finding that didn't work than bury it — a result that falsifies the method is still a result, and it's more useful to you than a tidy story. When I get something wrong in public, the correction is public too. The point of a standard is that it holds when it's inconvenient.