I built BTCDecoded Intelligence to put Bitcoin governance analysis on a data foundation. The first thing I did was run it against my own work. It found things. Incomplete identity resolution across platforms had me undercounting concentration. A few arguments I used to reach correct conclusions were overstated. Some numbers were labeled without enough context to make them comparable. All of it is being corrected and republished. The core findings on merge concentration are stronger after the corrections, not weaker. Analysis from the tool will be released over time at @DecodeBitcoin.

Niklas Gögge wrote something worth engaging with on LLM scanning and Bitcoin security. Two pushbacks. The prescription is right. Fuzzing catches classes of bugs no amount of review finds. But the access argument doesn't hold. Attackers in his examples spent hundreds of dollars in tokens while the defensive response has run into the hundreds of thousands. Defenders can point a new model at their code the same day it ships. There's no gap to lose. He has a better argument sitting right there anyway, which is that scanning is ephemeral and testing compounds. A scan finds a bug and the insight evaporates unless someone encodes it. A harness freezes it forever. That survives regardless of who has which model. Second, a thousand reports that are mostly noise isn't a capability problem, it's a context problem. Point a stronger model at the same thin context and you get more confident noise. A weaker model that knows the protocol's design intent and why specific constraints exist will outperform a frontier model guessing from the code. Same goes for harnesses, since a model that doesn't know what property should hold writes one that tests the wrong thing. That's tractable, and it has the property he's arguing for. Good domain context is durable across model generations exactly like testing infrastructure is. It's his criterion applied to the input instead of the test. It's what I built BTC Decoded Intelligence for.. @DecodeBitcoin #Bitcoin BTCDecoded.org/intelligence