Sentimental Graffiti - English Translation Project

An English fan translation of Sentimental Graffiti (NEC Interchannel, 1998, Sega Saturn, T-20106G), a dating sim about twelve girls scattered across twelve Japanese cities. It was never released outside Japan, and as far as we can tell nobody had attempted this before - there was no existing toolchain for any of the game's file formats.

The game ships on two discs. This repository currently covers disc 1: its script is 100% translated - 21,216 player-facing records across all twelve routes - and it builds and boots.

Disc 2 has not been ripped yet. Everything is namespaced by disc so the two never collide: translations live under translation/variants/en/disc1/, and every tool takes --disc=disc1 (the default) or --disc=disc2. Do not assume disc 2 matches disc 1 - check its own TOC, since it may well reuse the same script filenames.

This repository contains no game data whatsoever. No disc image, no extracted script, no font, no executable, no BIOS. Everything here is either our own code or our own English prose.

To use any of it you need a legally-obtained copy of the game - ideally your own physical disc, which you rip yourself. The tools read your disc and write your patched image. Nothing is redistributed.

The same goes for the Sega Saturn BIOS, which Mednafen needs in order to boot anything. We do not link to one and cannot help you find one.

What you supply, and where it goes (all of these are in .gitignore):

originals/disc1/    files extracted from your own disc (script.pak, font.pak, 1st.bin, ...)
iso/                your rip: disc1.bin + disc1.cue, and the patched images the tools write
bios/               your Saturn BIOS (Mednafen wants ~/.mednafen/firmware/sega_101.bin)
extract/text/disc1/ regenerated by tools/scr.py - contains the original Japanese
                    script, which is why it is never committed
build/              patched font.pak / 1st.bin produced by the tools

Everything in this repo is designed around that boundary: the English lives in translation/variants/en/disc1/ as a plain {record index → English} map, carrying none of the original text. It is meaningless without your disc, and that is deliberate.

1. Read docs/SETUP.md and install the prerequisites.
2. Rip your disc (docs/PIPELINE.md - it is a mixed-mode 3-track disc, so a plain .iso is wrong and will not work).
3. Extract the archives, then regenerate the script extraction.
4. tools/textdb.py disc en iso/disc1.bin iso/disc1_en.bin --fullwidth --cols=42
5. tools/uistr.py extract, then tools/patch_1st.py disc en ... for the interface (docs/PIPELINE.md §7 - read the warnings, they each cost a broken build)
6. Boot the result in Mednafen.

Two things this is deliberately set up for.

A different translation. translation/variants/en/disc1/ is a plain {record index → English} map, and a variant can name another as its base and override only the lines it disagrees with. So a new tone, a more literal pass or a different name romanisation is a new directory, not a fork:

tools/textdb.py bilingual en review/     # Japanese and English side by side
tools/textdb.py check mine               # control codes, box limits, budgets

bilingual is generated output for reading - the Japanese and the English stay in separate files so that neither can be edited by accident.

A different dub. You do not need our engine, our cast or an API key. Every clip has a stable filename derived from its id - VO_HON0/002.wav - listed in translation/voice/manifest.json with the line it says. Fill a directory with those names, from any source including real performers, and:

tools/relayout2.py build en iso/disc1.bin out.bin --wav-dir=mydub/

Clips you leave out keep the original Japanese, so you can do one girl and play it. See docs/VOICES.md → Bringing your own audio entirely.

Clone, cd in, and run claude. It reads CLAUDE.md automatically - the formats, the constraints, and the ground rules this project learned the hard way. Four slash commands are set up:

Read-only tooling is pre-approved in .claude/settings.json; anything that writes to a physical disc will ask first.

PROMPT.md has a longer starting prompt and a list of what to tackle next.

The English ships as a bare {record index → English} map, which is safe to redistribute but awkward to proofread - you cannot see what the Japanese actually said. So the tooling generates a side-by-side file from your disc:

python3 tools/textdb.py edit en COMMON          # -> edit/disc1/COMMON.json
#   ... fix the "en" fields in your editor ...
python3 tools/textdb.py absorb en COMMON edit/disc1/COMMON.json

Each record shows jp, en, the record type and the voice cue. Edit the English, absorb it back, and the validator runs before anything is saved - it will refuse the write and tell you which record is wrong if you drop a $C, change a menu's option count, exceed a box budget or introduce a character the font cannot draw.

edit/ is gitignored: those files contain the original Japanese, so they are yours and stay on your machine. That is also why the Japanese is not committed here - anyone with the disc can regenerate it in one command, and anyone without it has no business having the script.

98% translated across script and interface, and playable end to end in an emulator.

Done: the script, fully translated and validated. .pak and .SCR formats decoded with byte-identical round-trip tests. Font format decoded - including a second 16x12 font nobody expected, which is what draws Latin in menus. A four-instruction SH-2 patch that halves the dialogue advance and doubles the line budget, plus narrow Latin glyphs.

The interface too: all 160 location and city names, the main menu, the options screen and its help lines, all 64 profile-card fields, the save/load screens, the calendar and weekdays, the travel map, job names, and all 120 tourist guide blurbs re-flowed into their fixed 16x4 grid.

The full English script is ~2.35 MB against a 1.54 MB slot, so relayout.py rebuilds track 1's layout to make room and trim_avi.py shortens the opening FMV by 2.4 seconds to buy the last 500 KB - both verified, both reversible.

Not done: 68 sound-test track titles, and the 232-line staff roll (deliberately out of scope). Disc 2 has not been ripped.

Cannot be done as text, and documented rather than attempted:

• The girls' given names are three characters each. They live in a 26-byte per-girl record that also holds affection and flag data, and eleven of the twelve are reached by computed address with no pointer to rewrite. Widening it is specified in docs/FORMATS.md; the failure mode is silent save corruption.
• FMV captions are bitmaps in telopcg.pak, a different archive format.
• The opening FMV is Duck TrueMotion 1, which ffmpeg decodes but cannot encode.

Tools and documentation: MIT (see LICENSE). The English translation text is released under the same terms. None of this grants you any rights to the original game, which remains the property of its owners.