Try everything, then check it exactly

The shortest constructions

The machine tries every legal move, in order of length, until it reaches the goal. The winner is then rebuilt in exact arithmetic, where it either works or does not.

ProblemEuclidHis movesFewestStatusCostChecked
equilateral triangleI.152fewest possible2C1 + 2C2 + 2C3yes
perpendicular bisectorI.10293fewest possible2S1 + 1S2 + 2C1 + 2C2 + 2C3yes
midpointI.10294fewest possible4S1 + 2S2 + 2C1 + 2C2 + 2C3yes
perpendicular at a pointI.11215fewest possible2S1 + 1S2 + 4C1 + 4C2 + 4C3yes
double a segmentI.3132fewest possible2S1 + 1S2 + 1C1 + 1C2 + 1C3yes
square on a segmentI.46375fewest possible2S1 + 1S2 + 4C1 + 4C2 + 4C3yes
equilateral triangle (compass-only)I.152fewest possible2C1 + 2C2 + 2C3yes
double a segment (compass-only)I.3133fewest possible3C1 + 3C2 + 3C3yes
midpoint (compass-only)I.10296fewest possible6C1 + 6C2 + 6C3yes

Why Euclid's numbers are bigger

The two counts measure different things, and the gap is mostly that. His moves counts every line and circle drawn once the earlier propositions he leans on are expanded — his I.10 reaches through three other results, so fifteen things get drawn. Fewest stops as soon as the goal exists and draws nothing to tidy up.

More to the point, Euclid was building constructions that could be proved from what came before. Reusing an earlier result costs moves but buys certainty. Counting moves was a nineteenth-century preoccupation, not his.

Working with fewer tools

A construction is a program, and these are different machines to run it on: compass and straightedge, compass alone, straightedge alone with one circle given, or a compass stuck at one setting. The classical theorems say the compass alone loses nothing, and the search agrees — at a price in moves.

How strong each answer is

Three different things, and they used all to be called fewest possible. Fewest possible means every shorter figure was enumerated in the exact kernel and none reached the goal. No floating point enters that claim, and neither do the bounds below: the exact enumeration applies no scope limit, no point limit and no deduplication at all. It is a theorem.

Shortest found. The search covered every shorter figure it could represent, but in floating point. A point pair merged at a tolerance of 10⁻⁷, a tangency lost at 10⁻⁹, or two different figures sharing a rounded fingerprint would each hide a construction and turn a longer answer into a false minimum. Certifying these depths exactly is out of reach.

Upper bound. The search hit its node budget. Only the construction shown is meaningful.

The float search is what makes the problem tractable at all, and it carries two further limits: points straying far outside the figure are ignored, and a figure of more than twenty points is abandoned. Neither touches a fewest possible row.