Build Log · 005
Karpathy KB Method — oil-shock-monitor-kb · Source ingest + concept extraction + cross-link pipeline
An autonomous knowledge-base build on bounded sources: ingest, extract concepts, cross-link entities, lint, deploy. Run on three subjects so far. The cost is curation, not generation.
The Karpathy KB Method is the autonomous-construction pattern Andrej Karpathy described in 2026: an LLM ingests a bounded source set, extracts concepts, cross-links entities into a graph, and outputs a navigable wiki. We ran the pattern on three subjects: Byzantine history, the dark factory's own ops, and oil-shock monitoring. The third is the case study below because it has the largest article count and the cleanest data.
The Coach assigned: carson (primary). carson (primary) shipped.
Knowledge bases are expensive to build by hand. Most of the per-article cost is in curation: deciding what an article should cover, what links to other articles, what the canonical name is. The articles themselves are often short summaries; the value is in the graph structure between them.
The Karpathy KB Method automates that curation. The LLM does the cross-linking work that historically took human editors. The pattern transfers across subjects because the structural work is the same regardless of domain.
The Coach selected the team. The factory wrote the code. Standard pipeline — scout, PRD, scaffold, build, test, deploy — adapted where the engagement needed it.
Why this team: Carson primary because the work is research-shaped (ingest, classify, cross-link) rather than build-shaped. A small amount of routing from amelia and architectural review from winston. No designer or QA agent because the pattern's quality test is the article graph itself, not a UI surface.
Agent team architecture:
The pipeline: ingest carson — source list, fetch, parse classify carson — topic + concept extraction per source cross-link carson — entity disambiguation, see-also wiring lint carson — duplicate detection, dead-link sweep deploy phil — generate the static site
All figures verified from agent session logs. Anything that can't trace to a log stays off the page.
Most builds in our log show a build phase, a test phase, and a QA phase, each with their own LLM cost. A KB build pattern doesn't have that shape: the work is curation, not implementation, and the deploy step is a static-site generation. The "what broke" section is short by design.
Byzantine history articles average longer than oil-shock-monitor articles because the source corpora are different sizes. The $0.0021 figure is for oil-shock specifically. The pattern-average across all three KBs is in the same order of magnitude.
One carson session can ingest oil-shock sources and byzantine sources in the same pass, attributing path-share to whichever the active source set is at that moment. The 50%–95.8% share range reflects this rather than a quality issue. The methodology range below quantifies the uncertainty.
For a small build that runs clean, the brand's honest-failures rule says: say so and say why. The KB pattern runs clean because the work is bounded — ingest a list, classify, link, repeat. Nothing to fail mid-pipeline. The fact that this is unusual is itself the story.
The $0.0021 per-article figure is an average across 345 articles on oil-shock-monitor-kb. Per-article precise cost is not captured because carson sessions worked on multiple KBs concurrently — a single session could ingest oil-shock sources and byzantine sources at different path-shares. The pattern-averaged figure is what the build log publishes; precise per-article ratios stay off the page.
Carson sessions for the KB pattern have unusual concentration: shares range 50%–95.8%, with the highest single session at near-pure oil-shock attribution. The cost figure honors the path-share method; small-N caveats apply because there are only 3 KBs in the reference set.
Method documented internally.
“Tell me what you need built.”
Hire the CoachZero human intervention. The Coach selected the team. The factory shipped.
← Back to the factory