3.1. Comparing implementations🔗

lake exe bench compare A B C [...] runs each named benchmark through its own (auto-picked) ladder, then prints:

the full per-function table (one block per benchmark)
the commonParams intersection (the params at which every function produced an ok row)
agreement over those params: all agreed, diverged at specific params, or hash unavailable (when at least one function's return type lacks Hashable)

A function whose ladder stops early (because its complexity rises faster) doesn't break the comparison — its rows are preserved and the agreement check just skips the params it didn't reach.

The Hashable opt-in: setup_benchmark synthesizes Hashable α once at registration. If it succeeds, every measurement records the hash and compare checks them. If not, runChecked still forces the result (so the benchmark measures real work) but writes result_hash := none, and compare reports hashUnavailable and names the offending registration so you know which return type to fix.

3.1.1. Reading a divergence report🔗

When compare finds disagreement, the summary highlights the earliest common param at which any pair of implementations produced different output, lays the per-function hash table out side-by-side, and tags each non-baseline implementation with the baseline it disagrees with. The first listed implementation (compare A B C → A) is the comparison baseline. Example:

common params (apples-to-apples): 2, 4, 8
agreement: DIVERGED on 2 params — earliest divergence at param=4:
    Sample.linearFn  hash=0xabc    (baseline)
    Sample.otherFn   hash=0xdef    differs from Sample.linearFn
  also diverged at: 8
  (only result hashes are available; see https://kim-em.github.io/lean-bench/Quickstart/ for what to expect)

Hashes are reported as hex literals. The also diverged at: line lists every later common param where any pair still disagreed — useful for telling "one off-by-one bug at small n" from "the two implementations are entirely different functions". Result previews (short string snapshots of the actual return values) are not recorded today — the wire format only carries hashes, and a 64-bit hash gives no way to recover the input. See Quickstart for the recommended debugging workflow when only hashes are available.