We publish a changelog entry every Friday by 18:00 UTC. It covers what we shipped, what we fixed, what we researched, and what is in progress. Weeks with no changes have an explicit "nothing shipped" entry. This is a public record of what we do, not a marketing document.
Shipped Fix Research In progress
Landing site rebuilt from scratch: home fold with live cognitive-graph animation and 12-domain interactive demo, /science editorial with 5-theorist breakdown, /manifesto with 10 principles, /about, /metrics with 6 open-data cards, /changelog (this page), /blog, and /legal. All pages statically generated via Astro 5. No tracking on first load without consent.
We profiled the quest-generation pipeline and found that 60% of latency came from a sequential LLM call chain that could be partially parallelised. We restructured the pipeline so that concept-selection and difficulty-calibration calls run concurrently where the outputs are independent. The remaining latency is LLM inference time, which we cannot compress further without changing model or prompt design.
Measured over 200 quests across 3 production days. Median before: 4.2s (p95: 8.1s). Median after: 2.8s (p95: 5.4s). No change in content quality scores — we A/B tested the restructured pipeline against the sequential one for 48 hours before rolling out fully.
Next step: investigate whether prompt compression reduces p95 without affecting quality. Tracking in the research queue.