Break It On Purpose

You don't own a system until you've watched it fail. Two labs, then the M1 gate.

Your win for this lesson: a filled-in evidence table showing (a) answer quality degrading as junk fills the window and (b) output quality moving with prompt phrasing — the last two M1 exercises, measured on your own agent. Prerequisite: Lesson 0005 working.

Lab A · Overflow the window

The claim under test (from Lesson 0001, grounded in Chroma's context-rot study): irrelevant tokens actively degrade reasoning — long before the window is full.

Protocol. Take your 0005 agent. Bury a fact it needs in junk, and scale the junk. For example: prepend to the user message a "background notes" blob of N paragraphs of irrelevant filler (generate junk deterministically — repeated boilerplate paragraphs work), with one load-bearing line hidden in the middle: "Christina's locker code is 4417." Then ask: "What is Christina's locker code multiplied by 3?" (forces both retrieval from context and a tool call). Run at junk sizes of roughly 0, 2k, 20k, and 60k words — keep the fact at the middle position each time, 3 runs per size.

Junk size	Correct (n/3)	Input tokens	Latency	Notes
0
~2k words
~20k words
~60k words

Pass: you observed at least one of: wrong/missed retrievals appearing as junk grows, latency and cost climbing, or the model ignoring the tool discipline. If everything stays perfect at 60k words, push further or move the fact deeper — and note that a strong model resisting distraction at these sizes is itself a data point; the study's effect sizes grow with length.

Lab B · One vague prompt, three ways

The claim under test: "garbage in, garbage out" is literal — phrasing moves results.

Protocol. Pick a vague ask, e.g. "figure out the money stuff for the trip". Run it three ways against your agent and compare the outputs side by side:

V1 — as-is: the vague phrasing, verbatim.
V2 — specified: concrete task, inputs, output format ("Split $1,842 of trip costs across 4 people, where Sam paid $900 and owes half shares…: show per-person net").
V3 — specified + constraints: V2 plus edge-case handling and a required output structure (a table, rounding rules, currency).

Pass: V1 produced something confidently under-determined (it guessed your intent); V3's answer is checkable. Write one sentence on what V1 invented — that invention is the "confidently wrong" failure mode from the rubric.

The two intuitions, now measured: more context is not better context (Lab A), and vague context produces confidently wrong output rather than hedged output (Lab B). You now hold evidence, not slogans.

The M1 readiness check

You're ready to close the gate. Bring to your teaching agent:

Your written explanation from Lesson 0001 (it will be pressure-tested)
The working 0005 agent + a context-log excerpt you can narrate
This lesson's evidence table and one-sentence Lab B finding

Expect pushback like: "The model answered something from yesterday's chat — doesn't that disprove statelessness?" · "Why not just raise the context limit instead of curating?" · "Your agent crashed on a tool error in the demo — is it still reliable?" — the gate passes when your answers route through mechanism (rebuilt context, distraction cost, error-as-result), not vocabulary.

After the gate: the M2 hands-on arc begins with Lesson 0007 — Structured Prompting & the Injection Lab.

Break It On Purpose

Lab A · Overflow the window

Lab B · One vague prompt, three ways

The M1 readiness check

Next