Watch it compile

August 2026 · a live demonstration, gated and hash-bound

This is our world model predicting what will happen to a repository, before anything runs. The repository was held out before training began; the model has never seen it. An agent is fourteen steps into fixing a real bug; we pause at step six, where the next edit is chosen but has not yet run. From nothing but the command and the state so far, the model predicts what that edit will do, will the code change, which file, how many lines. Replay already knows the answer, so you can check every prediction yourself:

Six for six on this step, and the compiled grader then pays reward 1.00, from the replayed truth. The predictions above were returned by the saved artifacts (sha dbeb6b02…, 577c647e…); the sealed rows passed a one-shot evaluation on 25 repositories and 8 languages the model never saw. re-run live

What compiles out: a verifier

This is the point of the page. Out of fourteen recorded steps, the world model pruned away the reads, the failed attempts, the wandering, and kept the one edit that mattered. Replay recovered the exact lines that edit wrote, and those lines became checks. The result is a verifier, an oracle for this task, made from the work itself, no human authoring a grader. The oracle has already run: it pays reward 1.00 · 0 errors · model none, and the do-nothing candidate pays 0.

See the oracle itself →, its four checks, the run receipt, and every artifact hash.

The successor reads finer

The sealed model answers whether and how much. Its successor, R3-v0 (gated this weekend, internal, two of three heads passed; the third failed and is published), answers what: the kind of edit and how the command exits, from command text alone. Its actual predictions on repositories it never saw:

EDIT-KIND, what will this command do to the file it touches?
cat <<'EOF' > reproduce_issue.py from pydantic impo…  create 0.91           create           ✓
old=$(cat <<'OLD'         # Handle user-defined gen…  in_place_modify 0.86  in_place_modify  ✓
rm test_nameemail_issue.py comprehensive_test.py te…  delete 0.98           delete           ✓
cat <<'EOF' > test_generic_fix.py from typing impor…  create 0.91           no_op            ✗

EXIT-CLASS, how will this command exit?
ls -la                                                no_error 1.00         no_error         ✓
cd /tmp && python /root/test_nameemail_issue.py       permission_io 1.00    permission_io    ✓
grep -A 10 -B 5 "NameEmail" tests/test_networks.py    permission_io 1.00    permission_io    ✓
ls -la .venv && source .venv/bin/activate && python…  permission_io 0.87    other_nonzero    ✗

The misses are shown deliberately. The best one: a command that writes a file is predicted create at 0.91, but the recorded truth is no_op, the file already existed with identical content. Command text alone cannot know that; replay does. That gap is why the compiler grades against recorded outcomes and uses the model to rank, never to adjudicate.

Why this compiles into RL environments, the essay · the numbers · hello@oneirology.ai