Retrodict

Retrodict

Retrodict is an agent for [ARC-AGI-3](https://three.arcprize.org). It solves every level of all 25 public games. Its [official scorecard](https://arcprize.org/scorecards/9c403765-db5b-40b1-beab-6fa3f40119b0) reports 99.86% mean RHAE (relative human action efficiency) at a $654 API-list-price cost. Only [Tycho](https://github.com/NIMI-research/Tycho) scores higher in the public harness review: 100.00% at an estimated API-equivalent cost of $2,986. This places Retrodict on the reported cost-performance frontier. Retrodict uses [ThinHarness](https://github.com/ryanbbrown/thinharness) and gpt-5.6-sol at `max` reasoning effort.


*Selected public harnesses above 75% RHAE. Cost methods and run-selection rules differ. See the [comparison methodology](docs/arc-agi-3-harness-comparison.md) for sources and qualifications.*

Further reading: [Blog post: How Retrodict was built](https://blog.ryanbbrown.com/p/how-i-accidentally-got-the-top-score) · [X thread: ARC-AGI-3 cost, performance, and harness specialization](https://x.com/NotRyanB/status/2086929999290937486)

How it works

Retrodict is an LLM agent that plays each game like a scientist with a lab notebook. Every frame the game returns is written into a log file, and the agent works over that file with code instead of looking at images. To learn the rules, it proposes hypotheses and tests them against its own recorded history first, writing python that replays a hypothesis over past frames, where being wrong costs nothing. Only a hypothesis that survives the log earns real actions: the agent commits a queue of moves, each carrying the exact cells it predicts the board will show, and the runner plays the queue out one action per step, returning to the model only when the plan runs out or a prediction misses, along with the diff of what went differently. What it establishes about a game is curated into a playbook memory fi