Study 04 Exhaustive solver test

Possible answer—or
the most useful guess?

We played every word in a curated 2,315-answer corpus to compare familiar candidate ranking with guesses chosen for maximum information.

✓ Generated by the regression testList version 2026-08-29 · SHA-256 eed49f95c7ac

Result Every answer played

More information. Fewer guesses. No failures.

Current-style baseline

Candidate frequency

3.65average guesses
Solve rate
99.22%
Worst case
8 guesses
Failures
18
Opening word
SLATE
UnscramblerHQ information engine

Entropy + probe guesses

3.43average guesses
Solve rate
100.00%
Worst case
6 guesses
Failures
0
Opening word
TARSE

The information strategy saved 0.22 guesses per puzzle on average. It solved 823 answers faster, tied on 1,014, and was slower on 478.

Method Same engine as the live tool

What the test actually does

The test treats each of the 2,315 curated words as the hidden answer, applies Wordle’s exact green, yellow, gray, and repeated-letter rules, and continues until the answer is solved.

The baseline may choose only from answers still possible and ranks them by position-specific letter frequency. The new strategy may also choose from a separate pool of 14,855 accepted guesses when a non-answer separates the remaining possibilities more efficiently.

Both the live solver and this test import the same feedback and ranking engine. The published numbers come from the generated benchmark artifact; the page contains no manually entered performance totals.

Distribution Information strategy

How many guesses did it need?

2 guesses67
3 guesses1,233
4 guesses962
5 guesses52
6 guesses1

Scope Honest limitations

A reproducible model—not tomorrow’s private answer list.

The answer corpus is the original curated Wordle solution set, not a prediction of The New York Times’ private editorial choices. It excludes dictionary-wide proper-name noise such as Aaron, Abbas, and Abdul.

The accepted-guess pool is a dated compatibility snapshot. The New York Times can add or remove accepted guesses, so UnscramblerHQ versions the list and publishes its checksum instead of claiming permanent access to a private dictionary.

Normal mode permits information probes. Hard mode restricts suggestions to still-possible answers.