Lower multiple choice probability floor from 1% to 0.1%

[CodexVeritas]

Summary

• PredictedOptionList clamped every option to [1%, 99%], which is stricter than the 0.1% minimum the Metaculus API accepts. Bot forecasts with options under 1% (or over 99%) were silently renormalized.
• On questions with many options, raising each low option to 1% could move the largest option by more than 5 points, which tripped the existing max-change check and raised a ValueError.
• The floor is now 0.1%. Options below it are pinned to exactly 0.1% and the excess is taken proportionally from the other options, so renormalizing can't push a pinned option back under the API minimum (plain clamp-then-divide would leave values like 0.000999). With 2+ options this also caps every option at 99.9%.
• Float rounding is absorbed by setting the largest option to 1 minus the others, so a 2-option list lands on exactly 0.999 instead of one float step above it. Lists with 1000+ options, where every option cannot keep 0.1%, raise a ValueError.
• The "Sum of option probabilities is not 1" warning is now only logged when the sum actually differs from 1; it previously fired on every PredictedOptionList.

Live API check

Posted multiple choice forecasts to a test question on www.metaculus.com:

Lowest option	Sum	Result
exactly 0.001	1	accepted
0.000999 (plain clamp-then-divide output)	1	400: "Probabilities for current options must be between 0.001 and 0.999"
0.0009	1	400 (same message)
exactly 0.001	1 + 1e-6	accepted

So changing only the clamp constants to 0.001/0.999 would get forecasts with a 0% option rejected.

Tests

• Updated test_clamping_of_probabilities for the new 0.1% / 99.9% bounds.
• Added tests for: sub-1% options passing through unchanged, a 40-option question that used to raise, and normalized outputs always landing in [0.001, 0.999].
• Added a test that the sum warning is skipped for sums of 1 (including float noise like 0.1 * 10) and still logged otherwise.
• Re-enabled a commented-out PredictionExtractor case with 0.1%–0.5% options.
• Checked ~590k inputs (exhaustive 2- and 3-option grids plus seeded random forecasts with many near-0.1% and near-99.9% values, up to 1500 options) against an exact-fraction reference implementation and the old code: no outputs outside [0.001, 0.999], no sums off by more than 1e-12, no reordering of options, re-validating an output never changes it, and no input that the old code accepted with API-valid output is now rejected.
• The new tests fail against the old code. Full unit suite passes (TZ=UTC).

🤖 Generated with Claude Code