DNA is copied semi-conservatively: helicase opens the double helix at two replication forks and a polymerase, held on the strand by a sliding clamp, adds the complementary base to each template base: A with T (2 hydrogen bonds), G with C (3).
Each copied base is wrong with probability p = 1 - aeff/100, where aeff = a · (99.5/100) - 0.15 · UV%. The number of errors in n copied bases is therefore Binomial(n, p): mean n p, variance n p (1 - p). A wrong base is drawn from the 3 non-matching ones and shown as a red-marked mismatch with no hydrogen bonds. Time to finish with forks starting at bases 25 and 75 is t = 1 s + 76 / rate.
UV light joins neighbouring pyrimidines (T-T, C-T, T-C, C-C) into cyclobutane dimers. The page draws dimers on a fraction of those sites in proportion to the UV slider; the accuracy penalty of 15 points at 100 % is a teaching parameter, not a measured value.
Simplifications: 100 bases (a real human chromosome has ~108), constant fork speed of 5-30 bp/s here versus ~50 bp/s in humans and ~1000 bp/s in bacteria, no leading/lagging strand or Okazaki fragments, no proofreading or mismatch repair (which cut real error rates to ~10-9 per base). The drawn helix twists 0.32 rad per base (about 20 bp per turn) so the bases stay readable; real B-DNA has ~10.5 bp per turn.
Running numerical checks...