The quantified archive
Stats
This site is an instrumented product: every post is chunked and embedded for retrieval, every public interaction is logged, and the numbers below are computed from that pipeline — not typed in.
Corpus
Published posts
67
Embedded chunks
1127
100% of posts searchable by the AI
Curator notes
58%
posts with my own commentary
Topics
178
distinct tags on published posts
Archive growth
Top topics
By format
Usage
The metric framework
North star: answered questions per week. This site exists to compress one person's reading into a queryable knowledge base. A question that gets a cited answer is the unit of value delivered — more of them means the corpus, retrieval, and interface are all doing their jobs. The usage tiles above track it over a rolling 30-day window.
Activation: a visitor asks their first question. Reading the feed is passive; asking is the moment the product's core loop engages. The homepage hero, suggested prompts, and per-post "Ask" entry points all exist to move this number; every question is logged server-side with its retrieval outcome.
Retention proxy: a returning actor asks again within 7 days. There are no accounts, so retention is approximated by anonymized repeat actors (salted IP hash — never the raw address). Coarse, but honest and directionally useful at this scale. Defined but not yet computed: the nightly rollups keep only daily unique counts, so this metric needs actor-level retention tracking before it can appear above.
Quality guardrail: answered rate. Growth in questions is worthless if the archive can't answer them. Questions with zero retrieved sources are logged, surfaced on the evals page, and become the reading list.
Honesty notes: counts are beacon-based (best-effort, bot-inflated, ad-block deflated); days are UTC; usage aggregates come from a nightly rollup of the previous UTC day, so the window always ends at yesterday and today never appears. Small numbers are shown anyway — the framework matters more than the magnitude.
Experiments
Product changes framed as testable bets. A result is only written once the data above can answer it — “running” means exactly that.
- running2026-07-08
Showing the retrieval mechanics ("How I answered this") increases trust enough that visitors click through to sources more often.
- Change:
- Added the per-chunk transparency panel under every chat answer, with similarity scores and body-vs-curator provenance.
- Metric:
- citation_click + source_click per answered chat_query
- running2026-07-08
An inline ask-input in the homepage hero activates more visitors than a chat hidden behind a floating button.
- Change:
- Hero gained a first-class "Ask the archive" input with hint chips; the FAB remains as the secondary entry point.
- Metric:
- activation rate: visitors with ≥1 chat_query (vs post_view)
- running2026-07-08
A one-click "Test knowledge" quiz on post pages converts passive readers into the learning loop.
- Change:
- Public post pages show a quiz CTA (when RAG is public); start and completion are tracked as separate events.
- Metric:
- quiz_start → quiz_complete funnel