Notes From Several Frontiers at Once.
We work agent execution, verification, machines, operating memory, and capital allocation with one method. What gets published is what the work taught us — where the method held, and where it did not.
Safety cannot be a policy the model talks its way past
Software that moves information can be wrong cheaply. A machine that moves mass cannot. That difference belongs in the architecture, not only in the test plan.
4 min readA gate that passes on mutated input is not a gate
Every green suite eventually turns out to be green for reasons unrelated to correctness. The cheapest defence is borrowed from experimental practice: break the thing on purpose first.
4 min readYour linter cannot tell you that you built the wrong thing
Static analysis answers whether an implementation is likely defective. It is entirely silent on whether this is still the thing you set out to build — and that question compounds.
5 min readMost of what we build is meant to be dismantled
A factory that can only build has no discipline. Archival is a result — provided the work returns its primitives, evidence, and market knowledge on the way out.
4 min readStop building ventures. Build the thing that builds them.
The decision that reorganised everything: treat the factory as the product, and accept that it looks like procrastination for as long as it takes.
3 min readEvidence is append-only, or it is not evidence
A record that can be quietly edited is a record of what someone currently believes. Contradiction should require counter-evidence, not a revision.
4 min readAgents propose. Systems decide.
The question is not how much an agent is allowed to do. It is what an agent is allowed to decide — and the answer should never include certifying its own work.
5 min readAn abstraction with one consumer is a guess
Writing the interface down before building it turned out to be the cheapest experiment available — and the migration guide turned out to be the honest test of whether a boundary exists.
3 min readThe first thing we built to be reused
Starting a second product from an empty directory made the case for shared infrastructure obvious. Building it revealed that reusable and working are different disciplines.
3 min readWe kept building features while the product was elsewhere
A month spent on lesson content rather than software revealed which half of the build anyone would actually miss. The answer was not the half we had been optimising.
2 min readTwo languages is not a translation problem
We built a bilingual product for a regulated exam and assumed the second language was an interface concern. It was the product, and the mistake cost a rebuild.
3 min readWe started by selling our time
The first version of this company was a services business. It worked, and it taught us the one thing services cannot do — which is the reason none of it exists now.
3 min read