Claude Code reviews · October 2–4, 2026

What the reviews found

The calculations reproduce. That does not mean the forecasts are accurate enough to sell or that the research is ready for a journal.

The October 2 review found sparse polling, no tested adjustment for sponsored polls, and incomplete checks against later results. We removed the earlier forecast percentages.

The October 3 review reran the calculations and found that the learned weights added little. A simpler method was at least as defensible. It also found a broken daily-update check and a gap in the website’s access controls.

Our response

October 4 chamber audit: Claude Code reproduced all 18 chamber scenarios and 42 historical checks, but found missed prior-experience credit, candidate-name mismatches, inconsistent Ohio treatment and an underreported directional bias. The new v3 repair addresses the input issues; it does not claim to eliminate bias. Read the chamber audit.

Review scope: The chamber v2 calculations have now been independently checked by Claude Code. Tide, poll-weighting additions and the v3 chamber repairs have not received that same follow-up review.

Fix the daily update’s citation checks. Ask the AI to flag conflicts between a poll report and the data feed.
Restrict operational pages on every hostname and bind the web server to this machine’s local address.
Remove withdrawn forecast margins from the data sent to browsers.
Use the experience regression as the main research method. Keep learned weights as a secondary experiment.
Test better uncertainty ranges for the fixed blends and a stronger polls-only comparison, using only earlier elections.

The later changes are builder-tested and still need another review. Claude Code is an AI reviewer within this project, not an outside human peer reviewer. Its checks are valuable, but they are not a guarantee.

Read the October 2 report · Read the October 3 report · Read the revised working paper

Codex builds and writes; Claude Code audits. People at Lampbotics direct the project and remain responsible for what is released.