Paper 1 — A Differentiable Atari 2600

128 B RIOT RAM, per frame, 30 frames of NOOP from the standard ALE boot (60 NOOP + 4 RESET). Max diff 0 bytes/frame on all 64 games.

Script

tools/rom_sweep/sweep_jutari_ram.py

Command

python3 tools/rom_sweep/sweep_jutari_ram.py

Artifact

tools/rom_sweep/results_jutari_ram.md

Runtime

~9–10 s / game

Hardware

M1 Max (CPU)

Verified by

PXC1 + jutari Pkg.test()

210×160 palette-index framebuffer, per frame, 60 frames of breakout_random_actions after the standard boot. Max diff 0 px/frame.

Script

tools/rom_sweep/sweep_jutari_screen.py

Command

python3 tools/rom_sweep/sweep_jutari_screen.py

Artifact

tools/rom_sweep/results_jutari_screen.md

Runtime

~10 s / game

Hardware

M1 Max (CPU)

Verified by

PXC-S screen conformance

The JAX port is swept against xitari with the same paradigm; PXC2 then asserts jaxtari ≡ jutari frame-by-frame.

Script

tools/rom_sweep/sweep_jaxtari.py

Command

python3 tools/rom_sweep/sweep_jaxtari.py

Artifact

tools/rom_sweep/results_jaxtari_ram.md

Runtime

minutes / game (JAX eager)

Hardware

M1 Max (CPU)

Verified by

PXC2 cross-check

The executed straight-through SOFT path is byte-identical to HARD: with relaxation off (the default), soft_rom_peek / soft_ram_peek equal the original one-hot dot product over 5,000 random peeks, and toggling a relaxed run on and back off leaves RAM and the rendered frame unchanged (no state leak). The 4-panel video only illustrates HARD ≡ SOFT-STE while a relaxed (α,T) path drifts — the proof is the regression check, not the video.

Script

tools/relaxation_study/verify_soft_ste.jl

Command

cd jutari && julia --project=. ../tools/relaxation_study/verify_soft_ste.jl

Artifact

tools/relaxation_study/video_out/divergence_si.mp4

Input data

• tools/relaxation_study/dump_divergence_frames.jl — renders HARD / SOFT-STE / relaxed scanlines to raw frame streams
• tools/relaxation_study/make_divergence_video.py — encodes those frames into the 4-panel mp4

Runtime

seconds (check)

Hardware

M1 Max (CPU)

Verified by

verify_soft_ste.jl (RAM + frame byte-identical) + the 64/64 screen sweep

The mp4 is an illustration, generated by the two input scripts. The bit-exactness is proved by the script in this row (verify_soft_ste.jl), not by the video.

Peak forward throughput: 2,947,553 env-steps/s on GTX 1080 Ti (batch 4096) and 3,119,115 on Quadro RTX 5000. 3000 CPU instructions/rollout, 10 repeats, batch sweep 1→65536.

Script

tools/bench_jaxtari_gpu.py

Command

python3 tools/bench_jaxtari_gpu.py --rom pong.bin

Artifact

results/gpu/gtx1080ti.json

Runtime

~10–80 s / batch config

Hardware

GTX 1080 Ti / Quadro RTX 5000 · JAX 0.10.1

Verified by

raw JSON env block (device, jax version, per-batch wall time)

Peak forward+backward (jax.grad) throughput: 2,799,870 env-steps/s on GTX 1080 Ti and 2,911,069 on Quadro RTX 5000 (batch 4096).

Script

tools/plot_gpu_throughput.py

Command

python3 tools/plot_gpu_throughput.py

Artifact

jutari_paper/paper/figures/gpu_throughput.pdf

Input data

• tools/bench_jaxtari_gpu.py — runs the batched soft-mode benchmark on the GPU → results/gpu/*.json
• results/gpu/gtx1080ti.json — raw per-batch wall times (GTX 1080 Ti)
• results/gpu/q5000.json — raw per-batch wall times (Quadro RTX 5000)

Runtime

plot from cached JSON

Hardware

GTX 1080 Ti / Quadro RTX 5000

Verified by

results/gpu/{gtx1080ti,q5000}.json

plot_gpu_throughput.py only reads the committed JSON and draws the curve — every plotted number traces to a measured wall time in results/gpu/*.json.

The heatmap is a per-step likelihood model, P_step(α,T) = p_read(T)^ρ · p_branch(α)^f_b. Its inputs are measured by running the real soft simulator (soft_step) on the Space Invaders ROM for 3,000 steps: ρ (mean instruction length), f_b (branch fraction), the actual branch-offset set and the fetched-address histogram. dump_profiles.jl writes those profiles; make_relax_heatmap.py renders the outer combination. The operating point α=6, T=0.14 is independently bit-exact-verified by verify_soft_ste.jl.

Script

tools/relaxation_study/dump_profiles.jl

Command

cd jutari && julia --project=. ../tools/relaxation_study/dump_profiles.jl

Artifact

tools/relaxation_study/relax_profiles.txt

Input data

• tools/relaxation_study/make_relax_heatmap.py — reads relax_profiles.txt → renders fig_relax_heatmap.pdf
• jutari_paper/paper/figures/fig_relax_heatmap.pdf — the rendered figure

Runtime

seconds

Hardware

M1 Max (CPU)

Verified by

relax_profiles.txt (measured); plotted by make_relax_heatmap.py

The heatmap is a model, not a per-cell brute-force scan: the closed-form p_read/p_branch are evaluated over statistics measured from a real 3,000-step soft run on the SI ROM (dump_profiles.jl). The operating point α=6, T=0.14 is separately bit-exact-verified by verify_soft_ste.jl.

The inverse ∂(move-right)/∂joystick is computed by Zygote autodiff through the paper's bilinear sampler: −35.73 for left, +35.73 for right, 0 for up/down — identical across all three soft variants (Theorem 1), while the naive integer-index path gives 0 in every direction. The forward ∂screen/∂RIGHT is a finite-difference directional derivative through the sampler and lights up the cannon edges. The values live in the committed ji_grad.txt; si_joystick_fig.py only plots them — it computes nothing.

Script

tools/xai_si_gradient/si_joystick_gradient.jl

Command

cd jutari && julia --project=. ../tools/xai_si_gradient/si_joystick_gradient.jl

Artifact

tools/xai_si_gradient/out/ji_grad.txt

Runtime

seconds

Hardware

M1 Max (CPU)

Verified by

ji_grad.txt identical across the 3 soft variants; figure si_joystick_gradient.pdf

The gradients are computed here, not in the figure script. Input is the real space_invaders.bin ROM (not redistributed; obtained via AutoROM), stepped to the 35 s scene for the cannon footprint. Zygote.gradient differentiates the sampler objective; the forward saliency is a central finite difference on the joystick. Outputs land in out/ji_grad.txt (committed) and out/ji_*.raw (regenerable).

Reads the committed gradient fields/values from tools/xai_si_gradient/out/ and draws the 2×2 figure (scene, sampler saliency, naive≡0, inverse bar chart). A plotting step only — no computation. Listed separately so the figure script is not mistaken for the source of the numbers.

Script

tools/xai_si_gradient/si_joystick_fig.py

Command

python3 tools/xai_si_gradient/si_joystick_fig.py

Artifact

tools/xai_si_gradient/out/si_joystick_gradient.pdf

Input data

• tools/xai_si_gradient/si_joystick_gradient.jl — computes the gradient fields/values → out/ji_grad.txt + out/ji_*.raw
• tools/xai_si_gradient/out/ji_grad.txt — the committed inverse-gradient values plotted in panel (d)

Runtime

seconds

Hardware

M1 Max (CPU)

Verified by

inputs from si_joystick_gradient.jl (ji_grad.txt, ji_*.raw)

Pure plotting — no data is computed here. The bar heights in panel (d) are read from ji_grad.txt; the constant wbar = 0.26 (si_joystick_fig.py:73) is only the matplotlib bar width / side-by-side offset (3 bars × 0.26 ≈ 0.78 of the unit spacing), not a data value.

Cumulative commits / active sessions / calendar days computed directly from the repo git log up to the implementation cutoff.

Script

jutari_paper/paper/make_figures.py

Command

python3 jutari_paper/paper/make_figures.py

Artifact

jutari_paper/paper/figures/fig_timeline.pdf

Runtime

seconds

Hardware

any

Verified by

git log (public history)