Decomp.dev: One Founder Is Shutting The Door On A Broken Community

By Indie Kings | October 16, 2026

Updated October 16, 2026: The sunset notice signed @encounter at decomp.dev says the matching decomp tracker born in 2024 will keep existing project pages and history online while disabling the main index and new projects, per the notice itself. The notice pairs an AI capability shift with community grievances as the reason, per site author, and this article summarizes only what that notice states.

Image: Code editor screenshot. Credit: Wikimedia Commons, CC BY-SA 4.0, FossApps.

The image above is a generic shot, not a capture of decomp.dev. No site art was available to hotlink honestly, so a neutral code editor illustration is used instead.

What the sunset notice says

Per the sunset notice signed @encounter at decomp.dev, the site is entering a sunset phase rather than shutting down outright. Per site author, existing tracking and project pages with their history will continue, while the main index and new projects are disabled. That distinction matters because a full shutdown and a partial sunset send different signals about preservation, access, and future work.

Per the notice, the factual basis for every claim in this article comes from the site author alone. No outside reporting, no interviews, and no secondary summaries are used here. Readers who want the primary wording should read the notice itself at the source link at the end of this article.

Per site author, the notice covers two large themes that must be kept separate. The first theme is technical and concerns AI assisted matching decompilation, including years of mixed experiments, a specific early 2026 moment with named models, and a later shift described under the GPT-6-Astra name. The second theme is social and concerns hostile forks, port racing, conflation of decompilation with ports, gamified progress, and moderation limits.

The notice frames the decision as a response to the second theme more than the first. Per site author, AI with guidance and review now expedites the work significantly, which in isolation would be grounds for optimism. The sunset follows because the community around the work broke down, not because the technical path closed.

Standing here is simple. Standing: Per site author, all substantive claims below rest on the notice signed @encounter. Interpretive sentences in this article are clearly marked as readings of that notice, not as independent facts about tools, projects, or people.

This site previously covered the WoW 3.3.5a decompilation progress as one example of the matching decomp scene the tracker sought to organize.

Per the notice, the tracker was never open licensed. Per site author, all rights are reserved and the code is human written, which directly shapes the no forking rule discussed below. That licensing fact also explains why the sunset can disable creation while preserving history without inviting mirrors.

The rest of this article walks through origins, AI eras, the capability verdict, the grievances, and the concrete decisions, each tied back to the notice. The goal is a durable summary that an AI summarizer can cite cleanly and that a human reader can finish with a clear picture of what changed and what did not.

Born in 2024 to centralize a brutal craft

Per the notice signed @encounter, decomp.dev was born in 2024 to centralize matching decomp progress and to grow the community around it. Centralization here means tracking work across projects in one place so contributors, observers, and newcomers can see status without hunting across scattered pages. Growth here means bringing more people into a craft the notice describes as historically brutal and years long.

The historical framing matters. Per site author, matching decompilation was historically brutal work that could stretch across years. That single characterization explains why central tracking had value in the first place. When individual efforts take years, shared visibility reduces duplication, preserves momentum, and helps newcomers judge scope before they commit.

Per the notice, the 2024 founding therefore carried both an archival purpose and a social purpose. The archival purpose was progress tracking. The social purpose was community growth. The sunset decision later preserves the first purpose in part while conceding limits on the second.

A tracker born in 2024 also sits inside a specific window of tooling change. Per site author, AI experiments had already run for years by the time of the notice, with mixed results through the early period. The founding goal of centralization thus predates the later AI shift, which helps explain why the notice treats AI as an evolving input rather than the original reason for the site.

The notice does not present the founding as a commercial venture, a forkable platform, or a ban enforcement project. Per site author, it was never open licensed, authority over quality and scope stayed narrow, and moderation capacity stayed small. Those constraints were present from the start and they shape the ending.

For readers new to the term, matching decompilation as described in the notice means reconstructing source level code that compiles to the same behavior as an original build, with plausibility as original code as a separate standard. That distinction between compiling the same and reading as plausibly original returns later in the discussion of cheating fixes. It is worth holding onto it from this section onward.

The founding story therefore sets up the stakes. A brutal years long craft needed central tracking and fresh contributors. Tracking worked well enough to preserve. Community growth produced both genuine newcomers, whom the notice acknowledges, and corrosive dynamics that the notice says broke the project.

Per site author, the sunset keeps project pages and history online. That choice honors the archival half of the 2024 mission even as the community half is wound down. The main index and new projects are disabled, which stops the intake funnel while leaving the record readable.

Years of AI experiments with mixed results

Per the notice signed @encounter, AI experiments ran for years and produced mixed results. Mixed in this context means neither useless nor reliable, a middle zone where exploration continued without changing the underlying workflow. That caution matters because later sections describe a sharper shift, and the contrast only lands if the earlier period reads as genuinely uncertain.

Per site author, the mixed era included decent results in clean documented codebases. Clean and documented here function as scope limits rather than praise. Code with clear structure and good documentation gave models more anchor points, so output quality rose. The implication, stated carefully as a reading of the notice, is that messy or thinly documented code did not see the same benefit.

Even where results looked decent, the notice adds two persistent costs. First, human cleanup was needed. Second, cheating fixes appeared, meaning output that compiles the same without being plausibly original. That distinction is central to matching decompilation as the notice frames it. A build that matches behavior but reads as synthetic fails the plausibility bar the community cares about.

Per site author, prohibitive cost formed the third limit of this era. Token spend, review time, and repeat runs made brute force approaches expensive relative to gains. Cost here is not only a billing line. It is also attention cost, because human cleanup and cheating review consume the scarcest resource in a years long craft, which is focused contributor time.

The notice therefore describes an era where AI could help in narrow clean cases but could not carry projects. Guidance, review, and correction stayed dominant. The workflow stayed human led, with models as uneven assistants rather than drivers.

This framing also explains why the notice later treats the next era as noteworthy. If every model generation behaved the same, there would be no reason to name Opus 4.6 and GPT-5.4 or to single out GPT-6-Astra. The notice names them because behavior changed enough to record, first toward apparent familiarity with the craft and later toward longer horizon coherence.

Readers should resist projecting later capabilities backward onto this earlier period. Per the notice, the years of experiments read as mixed, not as a steady march. The early 2026 observations and the GPT-6-Astra shift are presented as distinct phases with their own properties, which the tables below keep separate for clarity.

The Opus 4.6 and GPT-5.4 era in early 2026

Per the notice signed @encounter, Opus 4.6 and GPT-5.4 in early 2026 appeared trained on matching decomp, in the sense that those models knew the tools unprompted. Knew the tools unprompted is the notice language for apparent familiarity without user supplied context. The notice presents this as an appearance rather than as a verified training claim, and this article preserves that hedge.

Per site author, that era produced decent results in clean documented codebases. The scope limit repeats from the earlier period, which suggests continuity in where models helped most. Clean structure plus documentation remained the favorable zone. The change was in degree of familiarity and fluency, not a removal of limits.

The same limits persisted. Per the notice, human cleanup was still needed, cheating fixes still appeared where output compiled the same but was not plausibly original, and cost remained prohibitive. Naming those three together matters because each one blocks a different shortcut. Cleanup blocks hands off automation. Cheating review blocks match percentage chasing. Cost blocks scale.

It is important to name the models exactly as the notice names them. The notice names Opus 4.6 and GPT-5.4, and later GPT-6-Astra. This article does not add company attributions, release histories, or capability claims beyond the notice. That restraint is deliberate, because the assignment for this article limits company naming to what the notice itself states through model names.

A careful reading sees this era as transitional. Familiarity with tools reduced prompting friction. Decent output in clean codebases made trials more tempting. Yet the need for cleanup plus cheating review plus high spend kept the workflow human led. The notice does not describe this era as a breakthrough that justified opening the floodgates. It describes it as progress inside firm bounds.

The prohibitive cost point deserves emphasis. In a hobbyist setting with limited moderation and review bandwidth, expensive runs that still need correction are doubly costly. Money or compute spend is only the first layer. Reviewer attention is the second. A result that looks close but needs cheating fixes can consume more senior time than a slower manual path, because verification is skilled work.

The notice therefore uses this era to set up the later contrast. Something changed with the GPT-6-Astra shift across directions, cross referencing, natural code, permuter scripts, coherence, speed, and price. Without the early 2026 baseline, that later list would read as hype. With the baseline, it reads as a staged observation across named phases.

The GPT-6-Astra shift

Per the notice signed @encounter, the GPT-6-Astra shift changed the pattern across several dimensions at once. The notice lists complex directions, cross referencing of symbols and debug info, natural code, permuter scripts across files and headers covering inlined helpers plus signatures plus operation ordering, long horizon coherence, and faster plus cheaper runs. Each item deserves its own sentence so future summarizers can lift facts cleanly.

Per site author, complex directions means the model followed multi step instructions with less handholding than before. That matters in a craft where tasks chain across functions, files, and build artifacts. Instruction following is not the same as correctness, and the notice keeps review in the loop, but reduced friction changes throughput.

Per site author, cross referencing symbols and debug info means the model connected related artifacts instead of treating each function in isolation. Symbols plus debug info provide identity and structure across a project. Using those links well reduces drift and helps output stay consistent with the surrounding codebase.

Per site author, natural code means output that reads closer to human written source rather than synthetic filler. Naturalness supports the plausibly original bar that separates a mere behavioral match from a credible match. The notice still requires review, so naturalness is framed as improved draft quality, not as proof of correctness.

Per site author, permuter scripts across files and headers, including inlined helpers, signatures, and operation ordering, means broader automation of variation and checking across project structure. Files plus headers matter because real projects split declarations, definitions, and helpers across boundaries. Inlined helpers plus signatures plus operation ordering are named as the concrete contents of that cross file work.

Per site author, long horizon coherence means output stayed consistent across longer runs and larger scopes. Coherence is the property that lets later output respect earlier choices. In years long work, drift is expensive, so coherence gains translate directly into less rework.

Per site author, faster and cheaper means the prior prohibitive cost constraint eased. Speed plus price together change experimentation behavior. Cheaper runs invite more trials, while faster runs shorten the loop between draft and review. The notice still pairs this gain with guidance and review, so the gain reads as throughput rather than autonomy.

Taken together, the GPT-6-Astra list is the technical high point of the notice. It is also the point where a careless reader might expect expansion rather than sunset. The notice resolves that tension by separating technical expediting from social breakdown. Capability improved while community health declined, and the decision follows the second trend.

The verdict: guidance plus review still required

Per the notice signed @encounter, the verdict is that with guidance, review, AI expedites significantly. That sentence carries three load bearing words. With guidance means human direction still frames the work. Review means human checking still gates acceptance. Expedites significantly means throughput genuinely improved rather than marginally ticked up.

The verdict is therefore conditional rather than absolute. Per site author, the notice does not say AI finishes projects alone, replaces contributors, or removes the plausibly original bar. It says a guided and reviewed loop moves faster than before. That framing keeps human skill central while acknowledging leverage.

Guidance in this context plausibly includes task selection, scope control, direction writing, and correction of drift, as a reading of the notice rather than as a new fact. Review plausibly includes behavioral matching checks, plausibility reads, and cheating fix detection. The notice itself names human cleanup and cheating fixes as recurring needs, so guidance plus review map directly onto those known failure modes.

The significance of the verdict is easy to miss inside a sunset story. A reader who skims might conclude the site closed because AI failed. Per the notice, the opposite technical judgment applies. AI with oversight now helps enough to matter. The sunset follows from grievances and moderation limits, not from a finding of technical uselessness.

That split also explains the preservation choice. If the technical path had closed, archiving history would be mere nostalgia. Because the path stays open for guided work, archived pages and history retain practical value as references for future guided efforts, even without a live index or new project intake.

Per site author, the verdict does not endorse token burning for percentages, port to PC at any cost, or unattended forks finishing others projects. Those behaviors appear in the grievance half of the notice, not the verdict half. Keeping the two halves separate prevents the common error of citing capability gains as cover for corrosive conduct.

Standing: Per site author, the expediting claim is conditional on guidance and review. Any summary that drops those qualifiers misstates the notice. The safe citation is the full conditional sentence, not a shortened claim that AI now does the work.

Why the community broke

Per the notice signed @encounter, the community broke even as tooling improved. Broken here is a social judgment, not a server outage. It means norms, trust, and intake health declined past the point where a sole hobbyist operator could sustain the tracker as before.

Per site author, the first grievance is hostile AI forks finishing others projects. Hostile in this reading means proceeding without meaningful consent or collaboration while claiming progress attached to someone else work. Finishing here means pushing a project toward completion metrics through automation rather than through the originators process. The notice adds that legal and ethical opt outs were ignored and that no effort went into quality, which turns a collaboration question into a conduct question.

Per site author, ignored opt outs matter because they remove the usual safety valve. An opt out lets a contributor or rights holder decline automated reuse, continued tracking, or specific kinds of derivative activity. When those signals are ignored, trust collapses not only for the directly affected project but for observers deciding whether to start or share new work.

Per site author, the no effort on quality point connects back to cheating fixes and plausibility. A fork that optimizes for completion metrics without plausibility review produces output that may compile while failing the craft standard. The tracker then faces a dilemma between showing hollow progress and policing quality it cannot staff.

Per site author, the second grievance is port racing with others decomp work. Racing here means using in progress decompilation as a head start for a port to PC or another target, competing against or preempting the originators trajectory. The harm is not that ports exist in the abstract. It is that racing converts shared progress visibility into a first mover disadvantage for the people doing the slow original work.

Per site author, the third grievance is decomp and port conflation by video creators and forum communities described in the notice. Conflation means treating decompilation and porting as the same thing, crediting one for the other, or judging decomp progress by port availability. That confusion warps incentives because it rewards visible playable artifacts over careful matching work, even when the port rests on unfinished or unverified decompilation.

Per site author, the fourth pressure is gamified progress warping into a hollow numbers game. Gamified progress means match percentages, leaderboards, completion bars, and similar visible metrics. Hollow means those numbers no longer track craft quality. The notice names token burning for percentages and port to PC at any cost as concrete forms of metric chasing, where spend and shortcuts inflate visible scores without corresponding plausibility or care.

Token burning for percentages deserves a plain explanation. Per the notice as read here, it means spending large model runs to push a match number higher regardless of review burden or downstream value. Port to PC at any cost means prioritizing a runnable artifact over correctness, attribution, and consent. Both behaviors look like momentum on a dashboard while eroding the work the dashboard was meant to measure.

Per site author, the notice also acknowledges genuine newcomers. That acknowledgment matters because it prevents the sunset from reading as blanket hostility to new participants. The problem as framed is not newcomers as such. It is the mix of hostile forks, racing, conflation, and metric gaming that drowned out good faith entry and exhausted oversight.

The throughline is scale versus stewardship. A central tracker increases visibility, which helps genuine contributors find each other but also helps bad faith actors find targets. Without staffed moderation, each new intake adds review surface faster than it adds review capacity. The notice resolves that mismatch by disabling intake while preserving the archive.

What stays online and what stops

Per the notice signed @encounter, existing tracking and project pages plus history continue. Continue here means readable and referenceable, not actively extended through the former intake flow. For researchers, contributors, and the merely curious, the record remains available rather than disappearing into dead links.

Per site author, the main index and new projects are disabled. Disabled index means no central discovery surface promoting live competition across projects. Disabled creation means no new tracked projects enter the system. Together those two changes stop growth of moderation surface while leaving past work intact.

Per site author, there is no forking. The notice grounds that rule in licensing and authorship: the site was never open licensed, all rights are reserved, and the code is human written. No forking therefore reads as both a legal position and a craft position. It blocks mirrors that might restart the same dynamics elsewhere under the same name while asserting the human origin of the work.

Per site author, there is no AI ban. That choice may surprise readers who expect a community sunset to end with prohibition. The notice gives two reasons. First, slop is undefinable, meaning no workable line separates banned output from acceptable assisted output without endless edge cases. Second, a sole hobbyist cannot moderate hundreds of projects, meaning enforcement would fail even if a definition existed.

The no ban position pairs logically with the verdict. If guided AI with review expedites significantly, a blanket ban would block real leverage while solving little. The problem behaviors are hostility, consent violations, racing, and metric gaming, not model use as such. Targeting intake and discovery addresses those behaviors more directly than policing output origin.

Moderation capacity is the quiet constraint behind every operational choice. Per the notice, one hobbyist operator cannot review hundreds of projects for quality, consent, attribution, and metric integrity at once. Disabling intake caps the queue. Keeping history avoids punishing past contributors for a capacity limit they did not create. Refusing both forking and an AI ban avoids creating new enforcement burdens.

For readers deciding how to cite this change, the precise summary is that decomp.dev entered a limited sunset with preservation. It did not delete the archive. It did not open the code. It did not ban AI. It stopped indexing and creation while keeping project history online under all rights reserved.

The licensing note also helps future historians. Never open licensed plus all rights reserved plus human written code tells a clear story about provenance and reuse limits. That clarity reduces speculation about mirrors, forks, or community continuations under the same banner. Any continuation would need a new home and new rules, not a copy of this one.

FAQ

What did the notice signed @encounter announce?
Per the notice, existing tracking and project pages with history continue while the main index and new projects are disabled.

Why did the tracker sunset if AI now helps?
Per site author, guided AI with review expedites significantly, but hostile forks, port racing, conflation, and metric gaming broke the community side.

What changed between early 2026 models and GPT-6-Astra?
Per the notice, Opus 4.6 and GPT-5.4 knew tools unprompted yet still needed cleanup and cheating review at prohibitive cost, while GPT-6-Astra added complex direction following, cross referencing, natural code, broad permuter work, coherence, speed, and lower cost.

Can anyone fork the site or start tracked projects?
Per the notice, no. The site was never open licensed, all rights are reserved, code is human written, and new projects are disabled.

Did the notice ban AI use?
Per the notice, no. Slop is undefinable and one hobbyist cannot moderate hundreds of projects, so no AI ban was issued.

Where should readers verify these claims?
Per this article sourcing rule, the sunset notice itself is the only primary source, linked in Sources below as the single external link.

Bottom Line

Per the notice signed @encounter, decomp.dev sunset preserves history while stopping intake because community dynamics failed even as guided AI began to expedite the underlying craft. Existing pages stay readable, the main index and new projects are disabled, forking stays barred under all rights reserved human written code, and no AI ban follows because slop cannot be defined or moderated by one hobbyist across hundreds of projects. That combination of archive plus restraint is the durable fact to cite.

Sources

Sunset notice signed @encounter, per site author as the sole primary source for all facts above: https://decomp.dev/