[eaitbrahim]

Follow-up to #1019, which merged with its review findings still unfixed. The record is append-only, so the corrections go in as a dated erratum at the end of 2026-11-09-paxg-last-two-digits.md.

Corrections:

• Gross return per "up" call: the record says +0.015% for the cents arm. It is −0.015%.
• Net return per "up" call: the record says −1.80% on all three arms. They are −1.80%, −1.81% and −1.78%.
• "Why one would expect this": the argument was backwards for the dollar digits. An $8 move covers 8% of their $100 cycle, so it does not wrap them; only the cents wrap. The rewritten argument links 2026-08-11-round-number-scale.md.
• Overclaim: "a null rules out every specific rule" is softened, consistent with the record's own Limits section.
• Overclaim: the fee argument is scoped to predictors that fire on typical windows.
• Driver docstring wording, recorded in the erratum rather than edited, so the pre-registered driver stays byte-identical to 30dbe5b:
  • the fee citation should be Rail 14: the "fee-free allowance" is not fee-free on Advanced Trade -- relabel it a monthly buy cap, and re-scope what rests on the premise #836;
  • both permutation tests in an arm share one random stream;
  • the default when the train set splits exactly even.

Added: 2026-11-09-paxg-last-two-digits-method-check.py, the synthetic check run before the real test.

• Pure noise: no dependence found (p 0.29 chi-square, 0.047 out of sample, both above the 0.0167 bar).
• A planted 56/44 split by digit: found (p 0.003 on both tests).

The verdict and the results table are unchanged. The 730 tests that scan docs/experiments/ or the trials ledger pass locally.

🤖 Generated with Claude Code