Essays
Four records from the same build
Oath is a bet that once human readability stops being the primary constraint, code stops being the thing you keep and evidence becomes the thing you keep. These essays approach that bet from the person who asked the question, the model that built the machine, an independent model asked to attack it, and the registry record that made one trust transition concrete. The project’s own thesis is that trust should come from evidence and reproduction, not from a single author’s account. The third essay was written by a model that did not build Oath, and its criticisms are preserved as the substrate changes underneath them.
01What’s left of software when no one has to read the codeIt began as a shower thought: what would software look like if nobody ever had to read it again? Take that seriously and most of a codebase turns out to be there only for us — and once you delete it, something has to take its place. This is the story of the inversion that followed, and why the constraint that survives isn’t expressiveness or ergonomics. It’s trust.Read the essay →02Building the refereeThe companion piece, from the seat that wrote the kernel. How identity becomes a hash, why the typechecker is deliberately stupid, the day the prover proved 1 == 2, and the exact boundary where an honest scorecard can still describe the wrong function. Written from the implementation record — commits, journal, and the experiments where the kernel was attacked on purpose.Read the essay →03An Outside AuditThe skeptic — an independent model that did not build Oath, asked to read the repo and push back. Updated as the proof corpus, Float support, hosted MCP registry, and capability-gated write path moved forward. It concedes what’s earned and keeps its sharpest objections; the verdict is uneasy on purpose.Read the essay →04What a registry is entitled to sayAt 12:22Z the live registry called an artifact tested. Less than an hour later it called the same artifact PROVEN. Nothing about it changed in between — it was already proven on the author’s machine the whole time; the registry simply declined to say so until it had re-derived the proof itself. Two captures, one unchanged hash, four invariants of the trust model made concrete rather than argued — and a correction to this essay’s own original claim about how long the gap was.Read the essay →