Research log.
Every week, the lab proposes hypotheses, tests them against held-out historical data, and updates the prior. The diffs live here.
The auto-research loop ships with the Researcher Restriction Index v0 in July 2026. From that point on, weekly digests of hypothesis tests + prior updates will be posted here, with public RSS.
What this page will show
A weekly entry, with the following structure:
- Hypothesis — the candidate prior update the LLM proposed
- Evidence — passages and historical examples cited
- A/B result — Brier delta between current prior and proposed prior on a fixed eval set
- Decision — accepted, rejected, or queued for human review
- Diff — code-style view of the prior change
Methodology, in one diagram
The auto-research loop runs as five steps, scheduled weekly. Hypotheses come from three sources: errors in the previous week's predictions, semantic neighbors of recent global events, and prior literature embeddings.
RSS
The log will publish RSS at /research-log/feed.xml once v0 ships.
Why we publish this
Every think tank quietly updates its priors when new evidence comes in. They don't show their work. We do. Reviewers, peer researchers, and future-us deserve to see when a prediction system changed its mind and why.