A standalone, reproducible study of 44 DOOM ports across ~20 programming languages, measured by structural geometry and call-graph wiring without reading identifiers or comments.
Separation of concerns: gitgalaxy (
tests/doom_rosetta/) owns the engine CI regression gate. This repository owns the scientific study: manifests, keyword-interpretation standards, preregistered follow-up experiments, measurement scripts, results, and the interactive article.
| Directory | Purpose |
|---|---|
manifest/ |
Pinned 44-port corpus (manifest.csv, scope.csv), declared_lineage.csv, subsystems.json, timeline_events.csv |
docs/ |
Canonical measurement contracts, starting with docs/KEYWORD_MODES.md |
preregistration/ |
Registered hypotheses for open questions (DOOM-E1..DOOM-E4) |
measure/ |
Study-side measurement, lineage, fidelity, and report-generation scripts |
results/ |
Generated evidence artifacts (blind_placement.*) and RESULTS.md |
followups/ |
Exploratory & registered follow-up experiments (oop-function-collapse/, composition-lens/) |
article/ |
Interactive HTML article (index.html, data.json, media) and text edition (teardown-of-doom.md) |
tests/ |
Structural integrity & schema checks (pytest) |
- Name-blind twins work within a language:
sfs twinsachieves 93% precision (65% recall) against source-text function diffs (Spearman ρ = 0.83–0.84 across 43–44 programs), while unrelated C controls (Quake, Quake II, Lua) sit at 0–0.6%. - Unsupervised Seed-and-Extend recovers 1-to-1 DOOM DNA across translations: Using
compositionmode, FGW reciprocal-best-hit seeding,deltatranslation-offset calibration, and 3-hop call-graph extension (zero names or paths read), we recover 95.3% (687/721) ofmood(C+COBOL), 89.0% (590/663) ofpuredoom(single-header C), 84.9% (547/644) ofdoomgeneric, 75.9% (426/561) ofgore(Go, even with flat directory layout), and 65.0% (433/666) ofchocolate-doom, while unrelated C controls (lua0.0%,quake0.6%,quake20.6%) and from-scratch reimplementations (room4doom,rust-doom,helionall 0.0%) sit at noise (FINDING.md). - Cross-pair permutations & expanded language/architecture controls confirm zero false DNA: Across 66 pairwise permutations of novel/from-scratch/COBOL repos (
agentic-cobol-doom-cc,agentic-cobol-doom-codex,fps-cob,fsharp-doom,doomf77,cssdoom,sqldoom,rust-doom) and 6 new permissive controls (carddemo-cobol-ctrl,c-raycasting,c-raycasting2,goom3d,bubbletea,macroquad), every non-DOOM pair scores 0–1 SE name hits. Meanwhile, the three COBOL raycasters cluster in FGW cost (0.670–0.706vs.0.748–0.775against same-size business COBOLcarddemo-cobol-ctrl) and align analogous raycaster roles without names (CONTROLS_CROSSPAIRS.md). - Preregistered held-out test (
DOOM-E5): On 5 held-out repos preregistered prior to cloning (EXPERIMENTS.md, DOOM_E5_RESULTS.md), Seed-and-Extend confirmedcydoomgeneric(100.0%Tier-1 todoomgeneric,FGW_cost=0.313),DOOM-style-Game(0.0%), anddoome(0.0%), while falsifications oniron-doom(0.0%— rewritten into Bevy ECS) anddoomarkable(0.0%— 41-function Rust wrapper around externaldoomgeneric-rs) revealed that ECS architectural rewrites erase call-graph DNA even when ported from Chocolate Doom. - Horizontal borrowing is visible without names — and Seed-and-Extend isolates the exact borrowed graph:
eternity(SMMU/MBF branch) andodamex(ZDoom branch) score only 6–12 Tier-1 hits againstlinuxdoom, yet blind Seed-and-Extend recovers 210/276 (76.1%,eternity ↔ chocolate-doom), 220/271 (81.2%,eternity ↔ doom-wasm), 178/212 (84.0%,odamex ↔ chocolate-doom), and 190/228 (83.3%,eternity ↔ odamex) Tier-1 matches directly inside Chocolate Doom's horizontally borrowedtextscreen(TXT_*) UI library (HORIZONTAL_BORROWING.md). - Architecture survives some translations above chance:
doom-scheme(+0.51) andgore(+0.50) preservelinuxdoom's cross-subsystem call wiring on par withchocolate-doom(+0.51), and blind placement hits 60% (κ = 0.50) on Go vs. 20% chance. - Raw wiring and graph-shape distance mislead without nulls: Raw wiring flatters every port (
0.72–0.95), and raw graph-shape distance does not separate DOOM ports (0.30–3.25) from unrelated controls (1.04–2.30).
See preregistration/EXPERIMENTS.md for the registered experiments addressing tiny OOP function splitting (mochadoom, kocoa-doom, managed-doom), cross-language composition matching, and subsystem other % coverage gates.
Measurements only: no third-party source code or WAD/game data is committed. Every port in manifest/manifest.csv carries an explicit OSI or CC0 license. "DOOM" is used descriptively; this project is not affiliated with or endorsed by id Software, ZeniMax Media, or Microsoft.