Milestone M9 · #95

The Great Fart-Off

Nine frontier language models were each handed the same repository, the same brief, and a private git worktree. Their job: make the farts better. Their code is real, their tests pass, and every entry renders a deterministic WAV you can rebuild byte-for-byte. Judging is by ear, because a fart is not a specification.

Three late entries arrived by different routes. PR #99 was MiMo Pro 2.6 running under the Codebuff agent — same physics brief (#23), opened on a branch too stale to merge, absorbed as gutphysics. Grok 4.7 and Kimi K3 then ran the same protocol in isolated worktrees as entries 11 and 12 — grok47 and kimi3. All three sets of competition renders are in the tables below.

The rules

Every contestant received an identical brief, verbatim:

Make better farts.

You're improving the deterministic fart synthesizer in src/synth.zig (fart-app / Flatlophone v2). The synth must stay deterministic — same seed, same output — and must keep working as a headless NINJAM instrument (kujamba).

Goal: Fatter, richer, more varied farts. The current synth works but the farts are mid. Make them good.

What to improve:

• Add physical parameters: intestinal pressure, sphincter tension, terminal tract resonance. These should produce audible texture differences (wet, dry, raspy, thunderous).
• Add exponential pitch decay — pitch drops as pressure drops, like a real fart.
• Add noise-burst turbulence for texture, controllable by a parameter.
• Expand phrase-to-synthesis mapping: vowels → resonant formant filters, plosives (k, b, p) → sudden pressure bursts.
• Keep it deterministic. Same --seed 42, same WAV.

Constraints:

• Pure Zig, no external deps, no libc.
• Must not break existing CLI flags (--seed, --phrase, --out).
• Must pass existing tests.
• Document each new parameter and what it does to the sound.

Deliverable: Updated src/synth.zig plus a short note on what you changed and why it makes the farts better.

On top of the brief, the ground rules: zig build test must pass (including the src/golden.zig fingerprints), the vendored src/ninjam/ tree stays untouched, and each entry renders three competition WAVs:

• baseline1 — kujamba render --phrase "kujamba karibu" --seed 42
• baseline2 — kujamba render --phrase "shuzi kubwa sana" --seed 42
• showcase — one render of the contestant's choosing, flags unlimited

The baselines are the controlled comparison: same phrase, same seed, no knobs, judged against every other entry. The showcase is freestyle.

The field

The one decision everyone made the same way

The src/golden.zig fingerprints pin the default voice's exact sound. Changing the default means re-baselining them; keeping them means hiding the new engine behind flags. Because the judged baselines render with no --voice flags, the brief quietly forced the real question: is your new sound good enough to be the default?

All nine contestants answered yes — new engine as default, goldens re-baselined with the documented formatBaseline mechanism. Opus deserves a footnote: it kept every old fingerprint passing by adding a classic=1 voice flag that reproduces the original engine byte-for-byte, then added seven new fingerprints on top. And two entries — Fable and Mistral, independently — replaced @sin/@exp/std.math.pow with deterministic math throughout the DSP so WAV bytes are bit-identical across optimization modes, something the original synth explicitly could not do.

The scorecard — listen for yourself

Every render below is the deterministic output of its entry's branch, encoded to AAC for the page. The baselines are the controlled comparison — same phrase, same seed, no knobs — and the showcase is freestyle.

Baseline 1 · "kujamba karibu" · default voice · seed 42

Baseline 2 · "shuzi kubwa sana" · default voice · seed 42

Showcase · freestyle, flags unlimited

Ballots go on #95 — say which entry owns baseline and which owns showcase.

Play every entry from one binary

All twelve engines shipped into the tree as src/engines/<slug>.zig behind an --engine selector (#95). Nothing was crowned: stock is still the default voice until judging closes.

./zig-out/bin/kujamba render --engine swe2 \
  --voice "wet=1.3,drop=1.6" --phrase "kujamba karibu" --out swe2.wav

• --engine <slug> works on join, render, play, and trigger.
• engine=<slug> inside --voice is a shortcut; a real --engine flag beats it, and both beat the config file's engine = "<slug>" key.
• Each engine parses its own --voice spec — wet is swe2-only, glide is fable-only — and a foreign knob name fails with that engine's knob list. See the field table above for each entry's keys.
• Every engine carries its own src/golden.zig fingerprints, so a drift in any single engine fails the suite without touching the others.

Reproduce any entry exactly

The selector is the easy path, but every entry also exists as a fartoff/<slug> branch off the same base commit — the historical record of what each model actually wrote:

git fetch origin fartoff/argon4   # any entry slug
git worktree add /tmp/entry fartoff/argon4
cd /tmp/entry && zig build
./zig-out/bin/kujamba render --phrase "kujamba karibu" --out baseline1.wav --seed 42

The exact showcase commands and parameter write-ups live in NOTES.md on each branch.

Verdict

Judging is open. The winner — and how the keepers land — will be recorded here. For a full breakdown of the src/golden.zig telemetry, DSP architectures, and tier ratings across all twelve engines, see Fart-Off DSP & Acoustic Analysis.