Map: 518 fingerprints, 38 walks over 31 stretches, 1 outings.

Each walk held out in turn, and its scans located from the rest of the map, each answer
placed along its stretch by the levels, which is experimental:

                                by networks  and by signal  signal, calibrated  settling ties choosing the path  at a walking pace
  scans held out                        518            518                 518            518               518                518
  placed on the right stretch            66             63                  63             69                55                 69
  placed across a mark                  452            419                 420            449               463                449
  landed on the wrong stretch             0              1                   1              0                 0                  0
  not on the map                          0             35                  34              0                 0                  0
  mean error, of a stretch              38%            37%                 37%            40%               42%                41%
  median error, of a stretch            31%            28%                 29%            33%               33%                34%
  mean error in metres                 33 m           32 m                32 m           35 m              36 m               36 m
  median error in metres               29 m           27 m                28 m           31 m              30 m               31 m
  within the band                394 of 518              -                   -     375 of 518        363 of 518         365 of 518
  pulled in, near a corner            -13 m          -13 m               -13 m          -12 m             -12 m              -11 m
  moved faster than a walk          8 of 52        5 of 48             5 of 48        8 of 55           7 of 45            3 of 55

  A scan that landed on the wrong stretch is not a small error, it is a different street,
  so it is counted apart rather than averaged into the distances. One placed across a mark
  is on the stretch next door, a few metres past the mark the two share, and its error is
  measured through that mark. Signal, calibrated compares levels after correcting the card
  by what it learns on the run, as --locate --watch --match signal does: here it corrected
  512 of 518 scans, by +0.0 dB. Settling ties places each scan with the scans before it
  breaking a tie between two stretches. Choosing the path takes the likeliest path through
  them, which can overrule the scan itself. At a walking pace settles ties and keeps each
  answer's place along the stretch to a walk's speed, as --locate --watch does. Pulled in,
  near a corner is how far the answers to scans taken within a fifth of a block of a
  corner landed towards the middle of the block, on average, and negative past the corner,
  into the block next door. Moved faster than a walk counts the steps between two answers
  in a row on the right stretch that nobody walks in the seconds between them. Within the
  band counts the answers no further from where the scan was taken, in a straight line,
  than the band the live map draws around them: 20 m, and 60 m more for all the similarity
  short of a perfect match. An answer on the wrong stretch counts as outside it, and one
  with nothing measured in metres is left out. The band was measured matching by networks
  weighed by rarity, and is not given for the signal or for networks weighed alike.

  518 of 518 answers were backed only by the outing the scan came from,
  which flatters this result: that is the map recognising a walk, not a place.
  Walk these streets again on another day and run this again.
