WebTug fairness evidence v1 — published October 20, 2026
Report: https://webtug.com/obstacle-fairness

Files
  reproduce.mjs      Dependency-free evidence runner.
  obstacle-path.ts   Exact production generator snapshot.
  transitions.csv   All 20,000 transitions, using round-trippable JS numbers.
  summary.json      Parameters, counts, full-precision measurements and SHA-256 hashes.

Reproduce with Node.js 24 or newer
  1. Save reproduce.mjs and obstacle-path.ts into the same directory.
  2. Run: node reproduce.mjs --output ./results
  3. Compare results/summary.json and results/transitions.csv with these published files.

If all four files are in the same directory, run:
  node reproduce.mjs --check
The command throws an error unless the reproduced CSV and JSON match byte for byte.
No packages, browser, network access, or WebTug account are required to run it.

Fixture
  Seed: 20260726; 20,000 consecutive transitions; 700 px high portrait arena.
  Scale: clamp(700/680, 0.8, 1.08); gap height: 174*scale.
  Initial gap center: 700*0.48; no previous-previous center at the first transition.
  PRNG: state=(1664525*state+1013904223)>>>0; value=state/4294967296.
  Two random values are consumed for each generated transition.
  A large move is strictly greater than 65% of the maximum step.
  After a prior move strictly greater than 78%, a move in the same direction
  must be no greater than 55% of the maximum (comparison tolerance 1e-9 px).

CSV columns
  transition: one-based index
  previous_previous_y: blank only for the first transition
  previous_y, next_y, delta_y: px; positive delta moves down the arena
  large_move: 1 if above the 65% threshold, otherwise 0
  continuation_limit_applies: 1 if the prior severe move and current direction
    meet the protected-pattern condition
  forbidden_continuation: 1 if that protected pattern exceeds its limit

Scope
This tests the real generator's bounds, maximum step, and protected pattern.
It does not simulate player motion, test every viewport, establish human reaction
times, or prove recovery from every possible position and velocity.
It is automated evidence, not a human playtest or a measured learning study.
The source commit recorded in summary.json identifies the production snapshot
from which this unchanged generator was copied; file hashes identify exact inputs.