Blog

Launch posts, each one carrying a figure measured from the corpus in outputs/.

• Hub-and-spoke indexes — Why the 10 KB index rule is a split rule and not a truncation rule: the root keeps one line per section with page and token counts, every section becomes a spec-v2 index of its own, nothing is dropped, and /ldo refuses to improve the prose.
• The lint that gates the estate — llms_lint.py runs the deterministic passes of /ldo and exits 1 on any High; docset_rollout cleanup now runs it across 15 docsets and 652 files at 0 High — and what calibrating it on real docs taught about placeholder keys, PEM headers and quoted injection phrases.
• Keyword plus vector: the cheap path — An FTS5 (BM25) table beside the embeddings: exact tokens like CLAUDE_CODE_SYNC_SKILLS or --append-system-prompt cost no embedding call, and a reciprocal-rank hybrid fixes the queries the vector layer ranks below troubleshooting rows.
• Anchors that point nowhere — 1,124 of 11,965 units on the pilot were anchored to headings the site never renders — MDX <Step> and <Tab> titles that cleaning had turned into headings. The fix anchors every unit to the nearest real source heading.
• Six months of hand-made llms files — What the ecosystem's llms files actually look like when you download 608 of them, what our own V1 pipeline was producing, and why the answer to both was a facts layer instead of a better site dump.
• Abstracting one concept out of many docsets — /lca pulls 'indexing' out of nine database docsets and 'prompt caching' out of three API docs: lexicon expansion, a zero-token harvest, borderline classification, a facet-grouped pack — and what the evals measured.
• A topical llms file from a pool of facts — docset_refine topical builds sections from a concept-tree node's children and files every fact by keyword, then file affinity, then embedding centroid, then ## Shared — the llms.txt family pilot, with the assignment counts.
• Turning a customer's docs into an llms family — A product docset becomes index / full / small / facts, split hub-and-spoke at 10 KB — Cloudflare, PayPal, Claude and LangChain, with the real byte and token counts.