Text encodings

There is no single FF7 character table. The engine ships three decoders and they do not agree, so a string is only meaningful together with the section it came from. Decoding field dialogue with the kernel's table produces plausible-looking wrong text, which is the worst failure mode available.

Field text

The field-mode codec is its own. It round-trips all 51,693 retail field strings byte-identically across the 702 maps in flevel.lgp, which is the check that establishes it is actually the right table rather than a close one.

Field strings live in section 0 alongside the script bytecode, which is why growing a script moves the string table — see [[technical/opcode/field-script]].

Kernel text

KERNEL.BIN and kernel2.bin use back-reference compression on F9: a marker followed by an offset and a length pointing backwards into text already emitted. The mechanism is not documented in the corpus and was recovered empirically; it decodes and re-encodes all 1,680 retail entries.

Practical consequence

Any tool reading FF7 text must know which of the three it is holding. A codec parameter is not optional politeness — picking the wrong one silently yields readable nonsense.