PROJECT · 0003 · ARENA GUIDE

The ARENA curriculum, in route order

Stack ARENA 3.0, submodule-pinned · Output 13 route entries · 7 visuals · Audience doers, with plain-words entries throughout

What does the ARENA curriculum actually teach, in the order a part-timer should learn it, and what makes each idea click?

This is the study guide for TARA — Technical Alignment Research Accelerator, cohort 2: one entry per ARENA part, in the route order the program's learning plan prescribes. Every entry has the same anatomy: the mental model in plain words, a visual built for that idea, ARENA's real exercise list with its own timed budgets for whoever wants hands, what the part unlocks, and the sources to go deeper. Start at the first row below and simply work down.

The entries — study them top to bottom

PART 0.0 · W36 · START HERE

Einsum is a contract between dimensions

the picture: contraction schematic · exercises: einops operations, ~45 min

PART 0.2 · W37 · CORE

You build ResNet34 by hand before anyone hands you a model

the picture: interactive convolution, step through the nine windows · exercises: eleven, ~3–3.5 h

PART 0.3 §1 · W38 · LIGHT

Momentum buys the crossing; plain descent buys the zigzag

the picture: two real descent paths on a canyon · exercises: four optimizers, ~70–100 min

PART 1.1 · W38–39 · CORE

Ten exercises build a transformer around one stream

the picture: the exercise map · exercises: ten plus one optional, ~2–3 h

PART 1.2 · W39–40 · CORE

Mech interp reads the activations, not the outputs

the arc: hooks, attention patterns, induction heads, ablations · no picture yet, lands with receipts

PART 3.1 · W40 · CORE

An eval is a threat model with a scoreboard

the picture: the eval anatomy · exercises: five, ~2.5–3.5 h

PART 3.2 · W41 · CORE

Eval datasets are generated, not found

the build: prompts, variance, concurrency, scorer-first · exercises: seven, ~1.8–2.7 h

PART 3.3 · W42 · CORE

Inspect turns an eval into a record you can argue with

the build: datasets, solvers, scorers, logs · exercises: six, ~1–2 h

PART 3.4 · W43 · CORE

An agent is a loop, not a model

the picture: the agent loop · exercises: six, ~1.5–2 h

PART 3.5 · W44 · CORE · GATE WEEK

Control assumes the model is trying to fool you

the picture: backdoors vs monitors and protocols · exercises: the heaviest set, W44 track-choice gate

PART 4.2 · W45 · CORE · API-ONLY

Misalignment is measured, not assumed

the work: shutdown resistance and alignment faking on real APIs · exercises: six, ~1.5–2.5 h

W46 · LIGHT · PICK ONE

Week 12 is a fork, and either road is one session long

the fork: SAE literacy or intro to RL · decided at the W44 gate

PART 4.3 §1–2 · W47 · LIGHT · GATE WEEK

Reasoning models say more than they do, so measure the gap

the work: chain-of-thought faithfulness measurements · exercises: six, ~1–1.5 h · W47 proposal gate

The route, in order

W36	0.0 prerequisites `[light]` — verify, don't redo	kickoff
W37	0.2 CNNs & ResNets `[core]` + 0.0 §2 recap	Session 1 — full coverage
W38	0.3 §1 optimizers `[light]` + 1.1 §1–2 `[core]`	transformer begins
W39	1.1 §3–4 `[light]` + 1.2 §1–2 `[core]`	foundations milestone
W40	1.2 §3 hooks `[core]` + 3.1 intro to evals `[core]`	threat model
W41	3.2 dataset generation `[core]`	capstone env spec v0
W42	3.3 running evals with Inspect `[core]`	—
W43	3.4 LLM agents `[core]` (§4 `[light]`)	check-in prep
W44	3.5 AI control `[core]`	GATE — track choice
W45	4.2 science of misalignment `[core]`, API-only	intent amendments
W46	1.3.3 §1 SAEs `[light]` or 2.1 intro to RL `[light]`	pick by track
W47	4.3 §1–2 interpreting reasoning models `[light]`	GATE — proposal
W48–49	ARENA slot → build	capstone + demo day

Everything not on this table — ray tracing, hand-built backprop, VAEs and GANs, five optional interpretability sets, all of RL past 2.1 — is a written-down skip with a reason. Nine of the curriculum's 32 parts carry the whole route.

The first picture

The flagship entry visual: convolution as a window that walks — user-paced, one output cell per window position.

Artifacts

data/raw/README.md — provenance: ARENA 3.0 pages (pinned `81533b2`) + Learning Plan v2
charts/*.html — seven single-file visuals: einsum contraction, convolution step-through, optimizer descent paths, transformer exercise map, eval anatomy, agent loop, control protocol
posts/*.md — thirteen guide entries covering the whole core route, each closing with its own sources

The sibling project 0002-tara is the conceptual companion: one standalone interactive page per program week, written for non-technical readers.