Audit every result, down to the script that made it

A reproducibility and provenance dashboard for the two papers built on the differentiable Atari 2600 VCS. Every claim links to the exact code, command, artifact, runtime and hardware — and to the external reference that proves the emulator is real.

Learning explainable AI? The XAI primer and the 30 method guides teach each method on neural networks and then run it, with code, on a machine whose true answer we know.

Space Invaders: xitari vs jutari vs pixel difference

Space Invaders — xitari (reference C++) · jutari (our Julia port) · pixel difference. The difference panel is solid black: byte-for-byte identical output.

Why you can trust this

Validated bit-for-bit against an independent C++ emulator

The strongest defence against “the AI hallucinated it” is that the two differentiable ports are checked against xitari — a separately written, pre-existing C++ Atari 2600 emulator that is not part of this project. On all 64 ALE-supported games the ports reproduce xitari’s 128 B of RAM byte-for-byte and its 210×160 framebuffer pixel-for-pixel. Hallucinated code does not accidentally match an external reference to the bit.

External oraclexitari C++ — 64/64 RAM byte-exact + 64/64 screen pixel-exact — google-deepmind/xitari ↗
Dual-port cross-checkPXC2: the JAX port and the Julia port diverge from xitari identically, in lock-step — a shared root cause, not two coincidences — Read the PXC2 code tour →
Conformance harnessPXC1 / PXC-S / PXC4 replay xitari traces and diff RAM, screen and a full 6502 functional test — Read the harness code tour →
Open dev log769 commits on main with the prompt→code chain; a 7,525-line running bug-fix log — bug_fix_log.md

The two papers

This site covers Paper 1 and Paper 2 only.

AAAI 2027 (submitted) · arXiv preview

Paper 1 — A Differentiable Atari 2600

Two end-to-end differentiable ports of the xitari VCS, validated bit-for-bit

Open the evidence ledger →

 Nature Machine Intelligence (in preparation)

Paper 2 — An Interpretability Ground-Truth Benchmark

Scoring XAI / mechanistic-interpretability methods against the exact causal truth of a known machine

Open the evidence ledger →

How to read a claim

Each row in a paper’s ledger is one auditable result.

FieldWhat it gives the reviewer
Valuethe measured number, read from the committed result file
Scriptlink to the exact source that produces it
Commandthe verbatim invocation to reproduce it
Artifactthe committed output (results table / JSON / figure) it wrote
Runtime & Hardwarewall-clock and the machine it ran on
Verified bythe test / oracle gate that guards it
Statusmeasured vs deferred — we never dress a plan up as a result