Prior Work Deep-Dive · self-reported numbers · 2026‑07
A system published 2026‑07 by Impossible Research + UC Berkeley + CMU. Every number on this page is a self-reported value they wrote on their own public page, and the loop signatures are verbatim copies of the schema-level descriptions they published — nothing here was reproduced by us or checked against their code.
Both RHAE(%) figures below are their own self-reports on the Public set (25 games only), not independently
verified by ARC Prize. The code itself is unreleased; the only artifact is the HF dataset
schema-harness/arc-agi-3-schema-traces — so even the function signatures this page
describes come verbatim from the schema-level explanations on their public page, not from source code.
The Fable 5 fallback rule: only games scoring under 80 were re-run with Fable 5 — so 98.98% is not a single-model single-shot score but the output of an ensemble pipeline that “retries weak games with a different model” (honesty notice).
| layer | ours (TTSO) | schema |
|---|---|---|
| ① verification | partial-credit credence (F1 scoring) | exact full-history replay certification (run_backtest) |
| ② body of knowledge | record shards (mechanic/wincond/strategy/metathinking) | one program revised in place (step) |
| ③ state representation | raw grid coordinates, fixed | invented representations · a counterexample can indict the representation itself |
| ④ acting | LLM eyeballing every round (THINK→HYPOTHESIS) | certified BFS queue (run_bfs→commit_actions) |
| ⑤ continuity | briefing reassembly (wake.py render() every round) | notes.md + persistent plan (durable files) |
Click any node in the §02 pipeline below and the matching row of this table unfolds next to its code schema.
The outer four stages are observe → deliberate → execute → record. Inside deliberate, three steps split out — theorize → certify → plan — and those three are the heart of “certify-then-BFS.”
One case, LS20, from their published per-game measurements. They state the discriminating-experiment principle — “commit the action for which the candidate rules predict different outcomes” — applied here too.
36/36 is a measured instance of “certify passed” — their stated figure meaning
world_model_v5 exactly replayed all 36 recorded transitions under run_backtest.
Their certify (run_backtest — a binary certification that passes only if every recorded transition replays exactly from scratch) is stricter than our code-fold’s (α,β) updates, but unforgiving. Our credence, conversely, is partial-credit — forgiving, but with no totality guarantee. The synthesis is to insert one certify-style gate right before promotion: before a skill crosses from park into the store, re-check in full that its predict_next replays every round recorded so far, and only then confirm the credence update. On top of that, what is distinctly ours stays — the cross-game transfer library, Bayesian credence, the leak-free audited constitution — because their memory dies with the game (world_model.py + notes.md live only inside it), while ours carrying over to the next game is the point of departure in the first place.