Fight Foundry

What this is

Preservation research. A 1998 arcade fighter rebuilt rather than emulated, move by move, then held against the original until it matched.

AI did the rebuilding, from three sources, none of them enough alone.

Where it is

About 70% of the intended game, in features and in feel. Feel is the closer of the two. Far from finished, and shown that way on purpose.

Two fighters, with movesets, cancels, reactions and supers. Missing: most of the roster, much of the systems layer, and the polish of a finished game.

Method

In order of authority.

  1. The ROM. Owned, several times over. Run under a pinned emulator, so measurements repeat. Sprites come from memory at known breakpoints; boxes are measured against the pixels they cover.
  2. Published frame data. Decades of community documentation. Fast, and usually right. Where it disagreed with the ROM, the ROM won.
  3. Direction. A person deciding what counts as correct. Measurement says what the original did, not whether it feels right.

Problems

Sound you cannot record

Isolating a sound needs two replays that differ by one thing. No such pair exists for a punch landing: closing the distance also changes the animation. Four distances, the air and the edge of range all failed. The sound was left out, not faked.

A click on every super

Samples sounded distorted. The obvious cause, too many at once, measured clean at −4.8 dBFS. The real one: a sample ended at full amplitude. A hard cut is a step, and a step is a click.

The round that would not end

The winner stood still but held velocity nothing could clear: only hit motion had drag. The beaten fighter's body blocked the drift, so nothing looked wrong. Two faults lined up; only one showed.

Proof

Tools

What the rebuild was made with. Separate pages, read-only.