Developer Says AI Agents Rebuilt an Unnamed Shooter's Source, 83% Byte-Exact, After DMCA Notices

Developer Maurice Heumann said on October 9 that Claude Code and Codex agents reconstructed the C++ source of a first-person shooter, with 83% of its functions matching the original compiled bytes. He estimated the job consumed 600 to 700 billion tokens.

Heumann wrote the account on his blog under the title "500+ Billion Tokens Later". He does not name the game and says the code will not be shared. He also said he had removed two earlier posts after pressure from a corporation.

What the results claim

Heumann reports that 99% of the game's functions are present in the reconstructed source and 83% of all functions are byte-exact against the original build. He said the game runs with no noticeable bugs and with all original features.

The remaining functions are, in his view, semantically correct but do not match byte for byte. He blames non-deterministic compiler output and COMDAT folding, a linker step that merges identical functions, that he could not reproduce. He declared the project done rather than spend more tokens on the gap.

1. Present in reconstructed source99 %
2. Byte-exact against original build83 %
3. Decompiled before the verifier existed80 %

Source: Maurice Heumann, blog post of 2026-10-09, accessed 2026-10-12

No third party has checked these figures. The project is not public, so there is nothing to build or compare.

The byte-matching verifier

The 80% figure comes from the stage before the verifier, and Heumann said that code contained semantic errors such as wrong signatures, invented or removed logic and costly architectural changes. A script that compared the rebuilt object file against the original executable and its PDB debug symbols fixed that, he said. It excluded relocation bytes and checked that references pointed at the same symbol and offset.

The rules were strict. Inline assembly, naked functions, object patching and embedded bytes were prohibited. CI hashed the verifier and compared it with a stored secret so agents could not edit the check that was grading them.

Setup detail	Reported value
First month	4 agents, 3 workers and 1 reviewer
Later configuration	14 Luna and 2 Opus 5.5 agents
Peak	more than 15 agents
Context compaction trigger	42% full, not the 90% default

Heumann ran Claude Max (20x) and Codex Pro together, with Sonnet 5 used most and Opus 5.5, Luna, Sol and Terra also heavily used. He said Haiku and Luna were poor fits until the verifier gave them strict feedback. An hourly cron job made agents reread the instruction document.

Where the token figure is soft

The 600 to 700 billion figure is an estimate. Heumann said logs were lost when virtual machines were wiped, and that roughly 300 billion of the tokens were Luna tokens. He gave no dollar cost.

On the Hacker News thread, commenter TomatoCo challenged a $100,000 to $750,000 cost estimate, and Heumann, posting as momo5502, said much of the volume was cheap-model tokens and cached input. Commenter sailingparrot asked whether the project was worth its emissions.

The legal question stays open

Heumann wrote in the thread, "I received a few DMCAs and thus stopped." He did not say who sent them. Commenter bitwize argued that a rewrite derived from reverse engineering could still infringe under the Whelan standard, while thway15269037 asked what exactly would be infringed and in which jurisdiction. Those are commenters' opinions, not rulings.

Commenters also split on whether byte-matching is the right goal. aetherspawn called identical assembly a ridiculous target, and Gigachad replied that emulation-grade fidelity and speedrunners need it. Commenter j2kun said functional equivalence depends on how complete the test suite is.

Legal pressure on unofficial software projects is not new: the Nitter maintainer reported a cease-and-desist from X. For another agent build, see the agent running on Cloudflare's free tier.

Heumann has not said whether he will publish the verifier or the instruction document the agents used.