July 8, 2026 · ai-sovereignty · local-models · risk · engineering

The Second Hallucination: your model vendor is a probabilistic system too

On June 12, Anthropic launched Fable 5 — the most capable model I had ever put to work. Three days later it was gone: a U.S. export-control order suspended access, Anthropic complied under protest, and every workflow anyone had built on that model stopped cold. The merits of the dispute — a claimed jailbreak, a company saying it was narrow, a government saying it wasn’t — are genuinely interesting and completely beside the point. What matters to an engineer is this: a dependency at the center of the workflow was revoked overnight, by a third party, for reasons that had nothing to do with anything I did.

I’ve written before about the first hallucination: a model hands you a confident, well-formatted, fabricated answer, and nothing in its tone warns you. The fix is a harness — deterministic checks the model can’t talk its way past. My whole system is built on that idea.

June taught me the failure mode has a big brother.

The second hallucination

The first hallucination is the model inventing facts about the world. The second is you inventing facts about the vendor — and holding them with the same unearned confidence:

  • The company will keep existing.
  • The model I depend on will keep being served.
  • The price I planned around will keep being the price.
  • I will keep being a permitted customer.

Every one of those is a probabilistic claim presented as fact. Confident, plausible, unverified — the exact shape of output my harness was built to reject, one level up where no harness was looking.

And the evidence that these beliefs deserve skepticism is public. Fable 5 shows the permission assumption failing. The solvency assumption is on file with the SEC: OpenAI filed confidentially for an IPO in May against a backdrop of multibillion-dollar annual losses, with profitability not projected until around 2030 and a funding gap that analysts measure in hundreds of billions. That’s not a prediction of failure — it’s a reminder that an API contract is, underneath, a bet on someone else’s balance sheet. The pricing assumption fails in the other direction: Chinese labs are serving frontier-adjacent models at a fraction of Western prices, which sounds like good news until you notice what a price war does to the vendors your architecture quietly assumed were permanent.

You cannot unit-test a vendor’s continuity

Here’s what makes the second hallucination nastier than the first: there is no deterministic check for it. I can verify a model’s claim that a file exists. I cannot verify a company’s claim that it will exist. No test suite covers “the Commerce Department changes its mind.”

When there’s no check, the only defense is independence — designing so the belief doesn’t have to be true:

  1. Own the weights that matter. The models doing my system’s bulk work run on GPUs I own. A file of weights on my own disk has no kill switch, no deprecation schedule, and no quarterly earnings call.
  2. Treat every model as replaceable. Everything routes through one chokepoint, and models are certified for roles empirically, by an evaluation battery — so when a model vanishes, requalifying a substitute is hours of compute, not a rewrite. My white paper calls this model-independence: the most capable tiers are the system’s architects and escalation reviewers, never its required operators.
  3. Rent capability; own continuity. I still use frontier APIs — they’re excellent, and for some jobs unmatched. But they get the work that can stop without stopping me.

The uncomfortable symmetry

We spent two years learning not to trust a model’s confident prose. The industry now needs the same lesson about its own dependency diagrams. Your vendor’s continuity is a sampled output from a distribution you don’t control — political, financial, competitive. Treat it the way you’d treat any other unverified claim from a probabilistic system: usable, useful, and never load-bearing without a fallback.

The model lies sometimes. The org chart above the model lies too — not out of malice, but because the future is out of everyone’s distribution. Build for the day the answer is wrong.

The architecture this essay comes from — the local farm, the crucible that certifies replacements, the routing chokepoint — is documented in the HMAS white paper.


All writing