Notes
Short, dated notes from the work: one thing that happened, with the numbers, and what it taught.
- Undetectable, not absent
147 BTC and crypto strategy candidates, none that survives a correction for all of them. How the ledger behind that count works, and why its conclusion is "not seen at this cost", not "no edge".
- Checking the check
The verifier step got five changes in one afternoon, taken from what the field was arguing about that week. Testing and reviewing it found three defects the same day. Two models, same catch rate, zero false positives.
- What a model swap exposed
Faults in four places around Turboscore's local model, from the times it was added, changed or moved. None is a verdict on the model.
- A local LLM on your own server: what it costs, what it needs
Two products answer customers with a language model running on a graphics card from 2016. Measured over 15 days: 12 W on average. The big card sits on the workstation, not in production. The numbers, and what they mean for a small business.
- The verifier that caught the verifier
Before a docs commit, a second agent with no memory of the work tries to disprove every claim in it. On its first live run it found three real errors in one change, including a "correction" that was itself wrong.
- Capping a site build on a 1.7 GB server
This site builds itself on a one-CPU server with 1.7 GB of RAM. The uncapped build peaked at 677–689 MiB; a systemd cap at 450 MiB cost nothing measurable. What the two memory numbers mean, and what did not work.
- What "Who is hiring?" asks for
Twelve months of Hacker News hiring threads, counted by one script: which AI terms rose, which stayed flat, and what the thread itself did.
- A test that can fail
Turboscore's search-quality suite reported green for six weeks while asserting nothing. What it caught once it could fail.
- How I work with coding agents
Several Claude Code sessions at once against live systems: the rules that keep that from going wrong, and the audit that produced them.