TABARC-Code/Chaos-Skill Claude-Code-skill

Bounded randomness for learning projects, plus a way to find out whether it helps you.
Author: TABARC-Code. Version 3.0.0. Python 3, standard library only.
## What it does
It shuffles practice order, jitters review gaps inside limits you set, suggests a difficulty level from an Elo-style rating, and hands out the occasional wildcard exercise. Draws can be sealed behind a SHA-256 hash and revealed later, so you can't peek and nudge. If you want evidence rather than vibes, it can also run a hidden control: some sessions get zero randomness, some get the real thing, and you compare.
## What it doesn't do
It doesn't make you learn faster. Spacing and interleaving have decent support. Random variation as such doesn't, as far as I could find. The experiment exists because of that gap, and with a normal amount of data it will mostly shrug and say "can't tell yet".
"Quantum" is only claimed when a live ANU quantum feed answered, passed basic health tests, and was XORed with OS randomness. I haven't been able to test that path with a real key. The unit tests mock it. Without a key you get the OS CSPRNG, which is excellent and not quantum.
## Install
Copy the folder anywhere and run scripts/chaos.py with Python 3. SKILL.md is there if you want to load it as an agent skill; the script doesn't need it. Nothing to pip install.
## First run
python3 scripts/chaos.py shuffle --items "scales,sight-reading,theory"
python3 scripts/chaos.py log --skill piano --level 4 --result 1 --p 0.8
python3 scripts/chaos.py next --skill piano
python3 scripts/chaos.py report --text
State goes in .chaos/ledger.json in the current directory. Use --dir to move it.
## The experiment
python3 scripts/chaos.py tune pick # arm stays hidden
# ...do the session, logging attempts...
python3 scripts/chaos.py tune clear
python3 scripts/chaos.py tune report # needs 4 sessions per arm
Assignment is in shuffled blocks of four, two control and two chaos. It doesn't adapt to results, because adaptive allocation biases the estimate. Rough error rates from simulation, assuming independent attempts: about 10% false alarms on pure noise, and a 0.10 real gain is spotted about half the time at 14 sessions per arm. Real sessions are messier, so expect worse. The report is descriptive. It doesn't model carryover or learning over time, and you can often guess which arm you're in.
## Commands
draw, reveal, plan, shuffle, drift, wildcard, log, next, report, tune, retro, entropy, feedback, status. Run with -h for flags.