Super Smash Bros. Melee matched about 5,000 functions in each of the last two years, even as the remaining pool shrank by half. The final jump came in linking, the step that makes a repository a buildable program. decomp.dev's owner has now closed the index.
Five weeks after its flagship game hit 100 percent, the site built to gamify decompilation switched off its scoreboard. Decompilation means rebuilding a game's source code from its shipped program. The owner of decomp.dev, who goes by Encounter, has disabled the main index and the addition of new projects.
Existing project pages and their history stay up. What closes is the ranking. The community that built the numbers now has to decide what the numbers are worth.
Melee's history is the clearest test of what the index measured. The project matched almost the same count of functions in two successive years, from a pool that halved. Then it spent 38 days on linking, the step that makes a repository a buildable whole. Matched share counts work done, but linked share shows whether the work can run, and that number moved last and fastest.
The site's own notice explains the motive. It says decomp.dev was built to gamify decompilation, because watching progress rise kept people motivated when every fraction of a percent took real human effort. Today, the owner writes, that idea has warped into "a hollow numbers game". He lists three abuses. People fork projects to "finish" them with AI. Others race to port games using decompilations they did not help write. Some conflate decompilation with ports.
The owner does not blame the tools alone. He writes that AI "is not the issue by itself". He adds that as a sole hobbyist he lacks the energy to define a good decompilation. He also cannot enforce one across hundreds of projects. He dates the shift to early 2026, when Opus 4.6 and GPT 5.4 appeared to know the community's tools.
He says recent models such as GPT 6 Astra are exceptionally good permuters and stay coherent across long tasks. A permuter tries many small rewrites until the output matches. 404 Media reported on 15 October that it has become overwhelming to track which projects are led by humans.
Melee's own build history shows 9,719 matched functions on 30 September 2024. A year later it showed 14,689, and on 30 September 2026 it showed 19,828. That is 4,970 functions in the first year and 5,139 in the second.
The totals look similar, but they describe different workloads. About 10,100 functions were unmatched at the start of the first year, and about 5,100 at the start of the second. The project matched roughly half of what was left in year one and effectively all of it in year two. The second year ran about twice as fast per function remaining.
The monthly path shows when the pace changed. Matched share rose from 53.1 percent at the end of May 2025 to 74.1 percent at the end of September. It crept to 76.2 percent by December. Three quiet months followed, at about one point or less. Then it rose 6.0 points in January 2026, the start of the early 2026 window the owner names. After that it added one to six points a month until it reached 100 percent.
The sharper change came in linking, the step that stitches matched code into a complete, buildable program. The share of code in fully linked units stood at 28.3 percent on 30 September 2025. It was 54.6 percent on 31 July 2026.
Then it reached 77.3 percent on 31 August and 100 percent on 7 September. That day the first build at 100 percent matched appeared, under the commit message "Match mnSnap_80257F24 and link mnsnap". The last 45 points arrived in 38 days. That is about 1.2 points a day, in a project that had gained 26 points across the previous ten months.
Linking had gained 7.5 points in June and 5.3 in July, then 22.7 in August alone. The finish ran several times faster than any earlier month. Matching in the same August moved only 1.4 points.
Matching climbed in bursts across the period. Linking sat near 28 percent for three months and then rose faster than matching ever did. The closing stretch produced a runnable whole.
A fully linked build is the easiest base for a port, and ports are one of the three abuses the owner names. Build records carry a date and a percentage, not the name of whoever produced the last match.
The commit text shows how mixed the work is. By our count of commit messages, 323 of the 2,585 build records in the Melee history name an AI tool, more than one in ten. Another 572 carry no message at all.
A leaderboard built on a byte for byte match rewards whoever spends the most effort searching for permutations that satisfy the check. A matching function can be correct to the byte and still unreadable as source.
That is why the owner can no longer tell the difference. It is also why the closure removes the ranking while keeping the record. A project knows its own contributors, and a hobbyist running an index for hundreds of them does not. So projects that want a milestone to mean human work will have to say so in their own repository rules.
A replacement metric should count what a check cannot fake: functions a named person has reviewed and can explain. By that measure, Melee's 100 percent linked build on 14 October 2026 is a finished program, with the share written by humans still unrecorded. Counting reviewed functions takes more effort than reading a percentage, which is the cost the owner says he cannot carry alone.