1985 · Melbourne House
Every byte of the C64's 64 KB, one pixel each, 512 to a row. Lit pixels are what this game is: its code, graphics, level data, sound, text, tables and variables, exactly where they sit in memory. Dim areas are screen and working memory; dark is unused.
| Tier | silver |
|---|---|
| Coverage | 100 % of the bytes the game uses have a description |
| Comments checked | Audited & high error rate. Every comment was audited once, but the last sample of 60 comments had 4 wrong, more than one in 30, so between 3 and 16 % of the listing’s comments are likely wrong. How listings are checked |
| Build analysed | |
| Copy | agent-draft |
| Tools | emulator: release v3.13.1, v3.13.1-linux-x86_64-gui.zip (vice-mcp), run under Xvfb; check-emulator 56 of 57 (pause-at-instruction failed), disassembler: regenerator2000 0.9.20, host: Linux 6.18 x86_64, Ubuntu 24.04 cloud container with four cores and no display, python: 3.11.15, widget tests: node 22.22.0 (the page's ports checked against the snapshot and the game's code in kit/cpu/cpu6502.js); Chromium through Playwright 1.56.1 for the page check, subagents: six annotation agents on disjoint address ranges, started with the harness model alias opus |
| Model | claude-opus-5-5 |
| Kit version | 0.0.54 |
Every step that measured the Source tab’s comments or made them more accurate, in the order it happened, with the models that did it. Accuracy of the Source explains each step and what the row above makes of them.
| Step | By | What it found | Of the listing |
|---|---|---|---|
| Research | claude-opus-5-5 | Looked the game up online (its manual, reviews or wiki pages) before reading its code. | |
| Annotation | claude-opus-5-5 | Wrote the listing’s comments. 100 % of the program is described, in 884 comments. six annotation agents started with the alias opus; the lead did $4000-$FFFF and zero page | |
| Sample 1 | claude-opus-5-5 | 12 of 80 comments wrong, so between 9 and 24 % of the listing is likely wrong. All 12 were then corrected. 60 hand-written comments over every annotator's range and 20 from the comments written to a template; 10 and 2 wrong, 14.6 % weighted by stratum size | |
| Audit 1 | claude-opus-5-5 | Checked all 884 comments and corrected 121. two checkers, one after the other, on $0000-$198F (45 corrected) and $1990-$FFFF (76 corrected, 38 of them one wording slip on the multiplexer's self-modified operands) | |
| Sample 2 | claude-opus-5-5 | 4 of 60 comments wrong, so between 3 and 16 % of the listing is likely wrong. All 4 were then corrected. after the audit: 41 hand-written comments over every annotator's range and 19 from the template families, by the checker of the first sample, which wrote and audited none of them; all four wrong were hand-written |
likely wrong the sample’s range corrected right, or likely right not yet measured
Read this before annotating code. What the game is documented to do, with verification status against the binary. Statuses: open (documented, not found yet), traced (in the code, could not be exercised; say what was tried), confirmed (in the code, consistent with the emulator), live (observed directly), differs (the code does something else). "Absent" is not a status.
Sources:
reference/), and the code (facts.md).| Feature | Status | Where |
|---|---|---|
Attract mode: the computer fights itself under DEMO, cycling through the backdrops | confirmed | $1CD2; live, reference/demo*.png |
| F1 starts a game; F5 abandons it and returns to the attract mode | confirmed | $1E64; live |
F3 switches between 1 PLAYER and 2 PLAYER | confirmed | $1EA8, $AE; live, reference/options-2player.png |
F7 switches between JOYSTICK and KEYBOARD control | confirmed | $1EC5, $C6; live, reference/options-keyboard.png |
| Fire starts a one-player game from the attract mode | confirmed | $1E64; live on port 2, and on port 1, which shares $DC01 with the keyboard columns |
| 18 movements from the joystick, eight directions with and without fire | confirmed | stick_to_move $286C; live, all sixteen inputs read back as move numbers (facts.md, Controls). Which named kick or punch each fire move is was judged from screenshots only for some |
| Keyboard controls, the alternative to the joystick | traced | $28B2-$29B3: Q W E / A D / Z X C and S or left SHIFT for player 1; P @ * / L ; / , . / and : or right SHIFT for player 2. Not typed live |
| Two players fight each other, one stick each | confirmed | $1B09; fighter 0 on port 2, fighter 1 on port 1 ($2851); live |
| Scoring with yin-yang symbols: half or whole by the quality of the blow; two whole symbols win the bout | confirmed | quality is distance: $2B70-$2C40 against the reach profiles at $7200; four halves ($8A,X) win ($1D5F); live, a whole point at a distance of 12 |
| A bout lasts 30 seconds on a counter; when time runs out the referee decides | differs | the counter ($212F) counts 30 units of 50 passes of the exchange loop, measured at 50 to 66 PAL frames a unit, so 30 to 40 seconds; the decision ($1C72, $1F93) goes to more halves, then more points; live, the judge's sprites at the end of a two-player game |
| Points scored per blow, shown for each player, and a high score | confirmed | $12A5, $206C, $200A; a high score per mode and a five-place table with names ($146A); live, 200 points for a whole point with move $0C |
| Ranks from novice to tenth dan, each opponent harder than the last | confirmed | $B4, $22C3, $1C2D; the computer's level rises with each bout won and sets its timing masks ($2593); two bout wins per rank ($B5) |
| The backdrop changes with progress: Fuji with a pagoda and a torii, a lake under a volcano, a dojo, a Buddha statue | confirmed | $AF AND 3; all four packed in memory ($3F5F-$71FF) and unpacked by $17BB/$1849; decoded and rendered from the image |
| A bonus round in which a bull charges and must be felled with one blow | confirmed | bull_round $132C, after a rank cleared on the Buddha backdrop; only move 7 (down-forward from a crouch) at a distance of $19-$1D fells it, for 3000 points; live, reference/bull-charge.png, reference/bull-felled.png |
| A computer opponent that fights back | confirmed | ai_think $22CF: blocks, nine scripted plans, attacks by distance; live, reference/play-novice-knockdown.png |
| A shadow under a fighter in the air | confirmed | the line-$D9 handler draws fixed sprites from $D480 at Y $DA; live, reference/jump-shadow.png |
A shout ("kiai") during loading; the disk carries m.spchtbl and m.tsound | confirmed | live, on Melbourne House's 1985 cassette (the Ultimate Tape Archive's) and on this disk after 1 at the menu: the loader puts up the loading picture (M.PICTURE, bitmap $E000, screen $CC00; on the tape $C36F), then its routine at $9000 (the disk's file LOADER) plays $2000-$35D6 of M.TSOUND through $D418, each byte's time from the timer table M.SPCHTBL ($9800/$9900), 9 x (count + 1) cycles; 5,591 flips, 2.24 seconds by the table and 0.7 % longer measured (2,726 of 4,479 gaps timed live within 4 cycles of it, the rest moved by interrupts held up), loading stopped meanwhile. The tape loads the same three files' bytes. Graphics and Music and speech show both |
| Music by Neil Brennan | confirmed | the driver at $09A5 and four songs; song 2 live in a one-player bout; tuned for NTSC, so two-thirds of a semitone flat on PAL (facts.md, Sound) |
| Sound effects on and off with DEL (the menu program's text) | confirmed | DEL flips $CF ($1ED8), which switches the speech samples off ($3186); live, $CF 0 to 1 |
Found in the code, not in the manual.
...... ($1472).$2CAE; live).$1C65).blow_reaction tests CIA 2's timer A ($2CAE) is not known.$15 lines between the fighters' tops ask two row changes of the same raster line ($32EA, $32FD); whether a row then shows a frame late was not tested.The symbol map for this game is symbols.json; the listing behind the Source tab is listing.json. Write-ups, facts and symbol maps are CC BY-SA 4.0. The game itself is not hosted here.