A missing character. A bad pointer. A convincing citation with the wrong evidence. These are the builds, failures, and fixes behind them.
56 published articles · archive $37
Hayato's Dreamcast art, animation records, collision boxes and palettes now render inside stock Marvel vs. Capcom, driven by Hulk's character code. The v0 video, what the render probe measures, the palette rule the hardware forced, and everything that is still not there: no Hayato moves, no download.
The biggest character in Marvel vs. Capcom 2 now renders inside stock Marvel vs. Capcom on Hulk's character code, from the first build of the port harness that Hayato proved out. The v0 video, the census, the 2,819-tile margin, why every Dreamcast-only character has been silent until now, and everything still missing: no Sentinel moves, no sound, no download.
Which MvC supers are mashable, what to press, and how many hits were demonstrated: character verdicts, baseline comparisons, code limits, reproducible evidence and a CPS-2 chip location map.
DHC v8 is out. The partner handoff's missing background was a stalled Hyper script waiting for an animation event nobody sent; six bytes send it. Level 4 of the meter turns yellow from a tile the game already had. What was verified, at what tier, and the trailer.
The cap is one byte. Four other tables also assume three. A statically verified five-level patch uses Capcom's existing art; a full-match play test is still pending.
Old write-ups made unfinished ports sound ready. I built a read-only MCP server with 21 tools to retrieve the current evidence without granting permission to run or publish.
I wanted to play as the helper I always picked. Rogue took Hulk's slot, but borrowed machinery still blocks a finished fighter—and the download stays withheld.
Six buttons, 21 selectable fighters, 126 captured choices. The downloadable palette patch follows your pick from select to fight; v2 fixes remain in v8 testing.
MvC's own rank table explains the difference. I built a training tool to read those rules live; the code passes headless checks, with a real Fightcade smoke test still ahead.
Lightning Storm lands ten hits. Magneto runs seven gameplay segments. The footage is real; diagnostic setup and unfinished parity keep it from being a release claim.
v8 is out: REMIX and ARCADE in one patch-only installer, with the partner handoff drawing its background, five Hyper levels with a yellow level 4, six colors and the expanded hit display. The release receipt says which checks ran on this build and at what tier.
That is Dreamcast MvC2's measured texture budget. The decoded stage format and conversion plan show what must be cut; a converted stage has not booted yet.
A 90-second control killed the theory. Following the renderer instead produced permanent Storm and Magneto icons in a 2,736-byte MvC ROM patch.
The stock counter really saturates; it isn't hiding extra hits. A separate display counter reaches 999 while the gameplay counter and ROM size stay unchanged.
Storm needs 12,329 tiles. Jin frees 12,873. That 544-slot margin is the difference between a stock-size plan and another expansion—and only the start of the audit.
The 1998 select screen now renders 20 visible boxes. Five new spaces, no erased stock routes—and the names, portraits, and fighters that still have to fill them.
Five fighters passed the editor checks. That did not prove they could fight. The gates that catch wrong builds, misleading screenshots, and missing move behavior.
Left and right can flip the fifth hit's launch—until geometry and the corner win. A stock-ROM experiment explains the rule, and why Zangief's 360 is different.
Mega Optic Blast is now a two-level air super in MvC. Making the stolen beam playable meant fixing its input, palette, damage, and the landing after the light goes out.
His fight tree and Captain America's editor tree told different stories. The missing-animation mystery, the real tile bill, and a gameplay reel that still isn't full parity.
The ending says YOU GOT MAGNETIC SHOCKWAVE. The arcade game never lets you use it. I restored the PlayStation-only reward with Onslaught's own arcade attack.
He appears in SVC Chaos through a custom emulator bridge, with source frames and attack experiments. The ROM, reactions, helpers, and sound still belong to Zero.
Zero's face and name are in Marvel vs. Capcom. Underneath, Mega Man still owns most of the machinery. The crossover prototype exposes what a sprite swap cannot fix.
A Mega Crash-style burst and a hold-to-parry prototype share one damage hook. Finding it exposed a trusted harness that had read the wrong memory for months.
This part has been merged into the single full write up, which carries it forward to the current state of the project.
Ken's sprites run inside Marvel vs. Capcom. The animation tables were the trap: 35 visual routes are now authored, but his Hadouken still throws Gambit's card.
This part has been merged into the single full write up, which carries it forward to the current state of the project.
I wanted Cable with six buttons. His art, four-shot rifle, and imported voice now run in MvC1; his native behavior is still unfinished. The road back from Dreamcast.
An average Dreamcast MvC character is larger than a Saturn's total memory, expansion cart included. The supposedly easy port fails before the first line of code.
A native import needed more space than the game's fragmented gaps could supply. Four emulator changes opened a contiguous megabyte—offline, beyond stock hardware.
Retail CPS-2 ROMs still contain Capcom's HIT, CATCH, AI, and art tools. I automated the hidden lab—and made it reject character ports that only look finished.
Claude agents built an entire working subsystem for the wrong explanation. The real fix was two bytes. How I decide which AI-generated evidence deserves to survive.
A false compression diagnosis nearly sent the port into weeks of codec work. Fixing the read exposed 11,587 tiles and a path to running his real code in MvC.
The street and subway render inside the wrong game. A character bug froze the fight, a second bug moved the underground into the sky, and some art still won't fit.
All 57,370 sprite records round-trip exactly, and character tables match the arcade byte for byte. Four wrong answers later, a running match settled the animation tree.
The game had 23 character slots. Dark Sakura and Shuma-Gorath run beyond them, with their own code—and every table past the old limit waiting to break something.
The HUD decoded every button wrong, an anti-air hid behind if False, and the policy lost control of most frames. A historical audit of the bugs around the model.
Borrowing Ryu's state machine made Dark Sakura fight in MvC1. It also trapped her in his rules. The playable shortcut I retired before building the native port.
One more prompt should not erase a human correction. The artifact IDs, revision checks, and approval rules an impressive AI demo tends to leave out.
A prettier answer can hide a worse production system. Compare against the model you already run, inspect the failures, and make the judge prove itself too.
A manual's storage limit becomes an operating limit, with a perfect footnote. Why real quotes and valid schemas still aren't proof.
A growing catalog turns similar tools into a selection trap. Search first, fetch one schema, and test what the smaller prompt still gets wrong.
The agent picks the wrong tool, a real citation supports a false claim, and the next prompt erases your edit. Four failure stories and one MCP system built in the arcade lab.
Thousands of replay records held his habits. An eight-frame alignment error forced the memorial AI to start over before it could claim to play like him.
Three table formats, a 13-state selector, and a 116-byte interrupt hook. Why adding one secret character became an engine investigation—and why the plan changed.
An unused character ID leads to four guarded load paths. The trail looked like a missing fighter; the later audit showed why it wasn't a ready-made slot.
Bigger ROM files fooled the diagnostics. Making the extra graphics visible took a custom emulator mapper—and cost standard Fightcade compatibility.
Secret developer tools, unused character routes, stolen boss attacks, and fighters in games they never shipped in. The complete CPS-2 research trail, including the dead ends.
A diagnostic menu painted garbage no amount of memory clearing could fix. The text renderer was still drawing because nobody had told it to stop.
Five tracks, 122 audio banks, and a custom FBNeo core made the sound-test demo possible—after a signed integer turned the soundtrack into a speedrun.
A friend's fighting style, characters stranded in the wrong game, and tools that catch my own bad conclusions. Start with the projects that made me keep going.
A design for turning FBNeo into a fighting-game lab: Metal rendering, live debugging, and parallel AI training on Apple Silicon. The architecture before the benchmark.
Six directional codes, their exact byte patterns, and the character IDs behind them—including the unused number that sent the investigation sideways.
A near-zero tile match looked like incompatible artwork. Reproducible ROM analysis exposed the bad assumption—and made the correction testable.
A two-byte search exposed three hidden-character table formats—and a hardcoded position check that kept the selector away from the new entry.
An abandoned Armored Spider-Man experiment found a way around the selector: ROM injection, palette writes, and a VBlank hook the state machine never expected.