Guides · reading list · in order

Five paths into decision agents.

Pick one path that matches how much time you have. Each path ends at the recorded replays on this site, which are the cheapest way to internalise what a structured decision looks like in motion.

01Path 1 · 10 minutes · curious reader

1. Skim the homepage — the spectrum L0–L5 and the vocabulary glossary are the only reading you need.
2. Open the three-system replay. Hit play. Watch one maze and one snake run.
3. Done. You now know what a structured decision is.

02Path 2 · 1 hour · practitioner

1. Read the Jev introduction: typesafe.ai/blog/introducing-system-one-models-and-jev.
2. Read the NanoJev README: github.com/TianyuCodings/NanoJev.
3. Read the input contract: docs/TYPESAFE_CONTRACT.md in the upstream repo.
4. Open the decision arcade. Inspect the per-step probability bars. Click between episodes.
5. Read Compare on this site. Pay attention to "What scales, what doesn't".

03Path 3 · half a day · researcher

1. Read everything in Path 2.
2. Skim the pipeline runbook: research/pipeline_runbook.md. You don't have to run it — just understand the data shape.
3. Read the atomic-planning note: docs/ATOMIC_PLANNING.md.
4. Read the scaled-games note: docs/SCALED_GAMES.md.
5. Read the RLCD experiment: docs/RLCD_EXPERIMENT.md (this is the calibration angle that matters for multi-agent).
6. Open the benchmark viewer. Inspect the 40-map numbers on test and OOD maps.

04Path 4 · full weekend · builder

1. Read everything in Path 3.
2. Download the model checkpoint: huggingface.co/C-Tianyu/NanoJev. Pick local_atomic_seed17 for the maze demo or games_gold_seed17 for Snake.
3. Follow the game release notes to reproduce the recorded runs locally.
4. Write a controller that uses NanoJev's /api/evaluate endpoint as a decision function. Pick any task with a clear state / question / candidate set (board games, scheduling, classification with abstention).
5. Compare: same controller, two decision heads. Measure probability quality and decision latency.
6. Sketch a 2-agent composition using one of the patterns from the homepage §05.

05Path 5 · background · multi-agent

Once you have the single-decision picture, the multi-agent literature is the natural next step. The reading list below is a starting point — not exhaustive.

Foundations

The "System One" framing from typesafe.ai — the argument that fast, structured decisions deserve their own model family.

Coordination

The seven patterns on the homepage §05 are the practical vocabulary. The literature around them is large; the homepage gives you the names.

Calibration

CE / Brier / proper-score losses; the upstream RLCD experiment is a working example with a small model.

Architecture

Multi-agent loops are usually presented as graphs of LLM calls. The interesting question is what happens when the nodes emit distributions instead of strings.

Open frontier

Self-calibrating multi-agent systems (§L5 on the homepage spectrum) are still research. The site will track them as the upstream project evolves.

06Prerequisites