Part 5 of 6 · 2 min read
Exercises
Easy
Easy 1: Run it (30 minutes)
Run pytest -q and python run_all.py in code/mini-hitl.
Done when: you report requests per task and harm leaked for all and irreversible, fatigued and not.
Hint
the fatigue flag in Params turns the attention model off.
Easy 2: Classify tools (30 minutes)
Take ten tools of an agent you know and classify each as low, reversible, irreversible or costly.
Done when: each has a gate decision and a one-line reason.
Hint
reversibility and cost are separate questions.
Easy 3: Write a request (30 minutes)
Write the approval request for a 400-dollar refund.
Done when: it shows action, arguments, reason, evidence, undo path and the choices.
Hint
include the undo path ("refund can be reversed within 24 hours").
Medium
Medium 1: Find the break-even load (2 hours)
Sweep tasks_per_hour and find where irreversible beats all.
Done when: you plot leakage for both policies and report the crossover load.
Hint
replace(Params(), tasks_per_hour=r).
Medium 2: Edit outcome (3 hours)
Add an "approve with edit" outcome that fixes a harmful action with some probability.
Done when: you report how leakage changes and what you assumed.
Hint
after a catch, with some chance the harm is fixed instead of blocked.
Medium 3: Bad risk scores (2 hours)
Make the agent's risk score less informative (means 0.45 and 0.55).
Done when: you show when the risk gate stops beating the irreversible gate.
Hint
change risk_mean_harm and risk_mean_ok.
Hard
Hard 1: Real gate (8 hours)
Build a pause-and-resume approval on a real agent with a checkpointer and run 20 tasks. (I did not run this, so your numbers are the ground truth.)
Done when: you report wait times, worker utilisation, any repeated effects and the library version.
Hint
store the approval request in the checkpoint and resume by thread id.
Hard 2: Measure reviewers (6 hours)
Run the section 7 experiment with at least five people. (I did not run this.)
Done when: you report catch rate, time and approval rate for both conditions, and compare the result with the attention model.
Hint
randomise the order of the actions.
Hard 3: Argument tampering (5 hours)
Test that an approved action cannot run with changed arguments. (I did not run this.)
Done when: a test fails when arguments differ from the approved hash and passes otherwise.
Hint
hash the tool name plus normalised arguments at request time and again at execution.