The proof cases that sold this year's biggest theses are failing stress tests faster than the narratives built on top of them can update.

The Pattern

Bhutan has sold 70% of its sovereign bitcoin in 18 months. Druk Holding and Investments went from 13,000 BTC in October 2024 to 3,954 BTC as of this week, with $215.7 million moving out in 2026 alone and most of it going to unlabeled wallets. The last mining inflow over $100,000 was more than a year ago. The hydropower-backed operation that made Bhutan a textbook case appears to have quietly stopped.

This matters because Bhutan was never just a small buyer. El Salvador already proved a sovereign could hold bitcoin as a treasury asset. Bhutan was supposed to prove something harder. That a sovereign could produce it at scale using stranded renewable energy, a sovereign wealth fund willing to experiment, and no legacy financial rails to defend. If sovereign mining was going to work anywhere first, it was there. The rest of the sovereign-mining thesis rested on the assumption that once one small country proved the economics, the wave would start. The country that proved the economics is now unwinding the position under halving pressure, difficulty pressure, and margin pressure. Strategy bought 4,871 BTC last weekend, US spot ETFs absorbed roughly 50,000 BTC in March, and the only visible sovereign seller is the one everyone used to point to.

The structural move here is subtle. The demo did not get debunked. It got worn out. It was real enough to start the narrative and not durable enough to carry the weight the narrative placed on it. If you are building on top of a thesis whose founding proof case is quietly being unwound by the people who built it, you are levered to a story that no longer matches its own footnotes. When we first read Mythos on Tuesday, the argument was that capability is abundant and containment is the new scarcity. That argument stands. What is arriving this week is narrower. The specific capability demo that made containment feel urgent is itself showing stress.

The Tension

AISLE took the Mythos showcase analyses and ran them against cheap open-weight models. All eight models they tested detected the FreeBSD NFS exploit, including a 3.6B-parameter model at eleven cents per million tokens. A 5.1B open model recovered the full chain on a 27-year-old OpenBSD bug. Small open models beat most frontier models on basic security reasoning. Meanwhile the original Anthropic disclosure rested on 198 manually reviewed findings and one publicly named CVE.

So the room where bank CEOs sat with Treasury Secretary Bessent and Fed Chair Powell to hear about a capability gap so wide it justified governance chokepoints was running on a demo whose discovery-grade work was already commoditized in a 3B parameter open model. The demo was real. The extrapolation was not durable. Sophisticated exploit construction remains frontier-dependent. Vulnerability discovery and reasoning, the part that sold the room, is not.

The same stress shows up inside one product line. OpenAI's Advanced Voice Mode still reports an April 2024 knowledge cutoff while the same company's Codex model will work for an hour to restructure a codebase or exploit a system. Andrej Karpathy's framing is that RL training concentrates on domains with verifiable reward functions and strands the domains without them. The ChatGPT brand is carrying a unified demo the actual product no longer matches. Builders pricing their roadmap to that brand are buying a narrative one generation older than the code underneath it.

What This Unlocks

If proof-case durability is the real variable, the winners are the people who check demos at the timescale demos actually fail on, not the timescale demos are marketed on. A sovereign bitcoin position sold as five-year strategy just failed its first eighteen months in its original test country. A capability gap sold as a generational moat was narrowed by commodity open models within a week of disclosure. A unified AI brand is carrying a lagging model underneath its most consumer-facing surface.

The losers are the builders who priced in demo durability. Anyone whose business model assumes Mythos-grade discovery remains proprietary for twelve months has just been told, by AISLE, that the window is closer to days. Anyone whose state-level bitcoin treasury strategy rested on Bhutan as the live reference is now pointing at an empty slot. Anyone who scoped a voice product to the ChatGPT brand rather than the specific underlying model is shipping on top of a label, not a capability.

The builder move is to stop buying theses and start buying the specific mechanism underneath them. When someone pitches you on a structural shift, ask what the proof case is, then ask how it is aging. Ask who is stress-testing it right now and where you can read that test. China's EV exports jumped 140% in March to a record 349,000 units because the oil shock stress-tested the internal combustion thesis in real consumer behavior, not in a deck. That is the kind of proof that does not wear out in a week. Most of what you are being pitched is not that.

Watching Next

Three observables, each with a clear falsification. First, does any other sovereign holder visibly liquidate a bitcoin position in the next 30 days, or does Bhutan remain the only public seller? If a second sovereign joins, the founding proof case is not an outlier, it is a leading indicator. Second, does Anthropic respond to the AISLE analysis by publishing the raw benchmark methodology behind the 181 Firefox exploits and the extrapolation math behind "thousands of zero-days," or does the disclosure stay at its current resolution? Silence is itself a signal about how durable the original framing is.

Third, and this one is for your own business. Pick the single most load-bearing proof case you cite when selling your company. It might be a case study, a pilot, a benchmark number, a named logo. Ask yourself when it was last stress-tested by someone who was not paid to make it look good. If the answer is more than 90 days ago, your pitch is running on a demo that has aged the same way Bhutan's has. Kristalina Georgieva warning of lasting scarring from the Iran war is the macro version of the same question. Scarring is what happens when proof cases from the old regime keep being cited after the regime has moved.

Underweighting

The strongest case against this thesis is that I am confusing two separate failure modes. Bhutan's liquidation, AISLE's replication, and Voice Mode's model lag are cases where each demo performed exactly as designed. The market extrapolated beyond what the test was engineered to support. If that is the right frame, the problem is not proof-case fragility. The problem is a mismatch between what was demonstrated and the scope of the claims built on top.

That distinction matters. Under my thesis, the advice is to stress-test proof cases faster. Under the counter-thesis, the advice is to refuse the extrapolation at the moment it happens, not after it fails. I think the counter-thesis is partially right. The Bhutan case genuinely supports the stricter reading. Druk Holding was cited as proof that sovereign mining could be profitable at scale, and the economics of that specific claim have now failed. The AISLE case is murkier. The extrapolation from 198 validated samples to 'thousands of zero-days' is an Anthropic framing problem, not a capability problem. Those are different failure modes, and I am treating them as one.

There is also an incentive question I did not raise cleanly. AISLE is a cybersecurity firm. They have a commercial interest in showing commodity open-weight models are sufficient for enterprise security work. Anthropic has the opposite interest. Any stress test that settles a live market dispute should be read through that lens. I am publishing this at the level of pattern observation, not proven structural claim.

Bottom Line

The question to ask every thesis this week is not whether the proof case was real, but how it is aging, and who is stress-testing it while you are still citing it. If your own pitch is running on a proof case older than 90 days, you are the narrative infrastructure that has not caught up yet. Go retest your best demo before someone else does it for you.