CHECKS OF PAPERCLIP WITH PYTHON
krueng.ai/paperclip_python.html

The page says how long Paperclip takes to wake an agent, that only one agent can
claim a task, that one of two budget requests stops an agent and the other does
not, and that a task with the title of an open task is not made twice. None of
that was typed in from memory. Each claim was printed by one of the page's own
files or by a script in this folder, and what each printed is saved beside it as
a *_output.txt file. The page shows those files. Run them again and compare.


WHAT YOU NEED

1. Paperclip running on 127.0.0.1:3100, started as in step 1 of the page:

       npx paperclipai@2026.1005.0 onboard --yes --no-install-service --data-dir paperclip-data

2. The page's files in one folder (every code block on the page has a Download
   button), and these scripts in a folder named "checks" inside it:

       your folder/
           hello.py  company.py  cashier.py  hire.py  order.py  correct.py
           runs.py  race.py  budget.py  spend.py  report.py
           checks/
               README.txt  scratch.py  wake_times.py  race_many.py  ...

   Each script here looks for cashier.py one folder up.

3. A Python with the requests library, as in step 1 of the page.


THE CHECKS

Run each from inside the checks folder, for example:

    ..\venv\Scripts\python wake_times.py        (Windows)
    ../venv/bin/python wake_times.py            (Mac, not run by us)

Each check makes a company of its own, named "Check ...", and never touches
Three Geckos Noodles. When it has finished it tries to delete that company. On
Paperclip 2026.1005.0 the delete often failed with an error 500, and then the
check archives the company and says so in its last line.

    what_it_gets.py   The versions, and the PAPERCLIP_ environment variables that
                      Paperclip hands to a program it starts.
    wake_times.py     Twenty orders, one after another. The time from creating
                      each task to reading "done".
    race_many.py      Twenty agents ask for one task at the same moment, from
                      twenty threads. Five rounds. How many got it each time?
    same_title.py     A task is created again with the same title: for the same
                      agent, for another agent, for nobody, and after the first
                      was cancelled.
    two_budgets.py    A budget of 100 cents set two ways on two agents, then 120
                      cents of cost reported for each.
    long_text.py      A long text kept in a task: how much of it comes back in a
                      list of tasks, and what happens to JSON kept there. It is
                      why paper_agent.py (Tutorial 5) sends its jobs as plain text.
    install_size.py   The size of what npx downloaded, and of the data folder.
                      Give it the data folder:  install_size.py C:/pclab/data
    scratch.py        Not a check. The other scripts import it.
    page_config.py    Not a check. It tells the site's tools which file printed
                      which output, and which number on the page comes from where.


THE PAGE'S OWN OUTPUTS

The output boxes on the page are the *_output.txt files one folder up. They
were made by running the page's files in the order listed in page_config.py
(STEPS), against a Paperclip that had no company in it. The order matters: the
steps expect the tasks to be numbered THR-1 to THR-4. To repeat the run, start
Paperclip with a new --data-dir and follow the page from step 2.

start_output.txt is the summary Paperclip printed when it started, with the
colour codes taken out. timings.txt has the time each step took.

TUTORIAL 5, THE PAPER TEAM

paper_output.txt, paper_options_output.txt and paper_big_output.txt are three
runs of paper.py, made on 10 October 2026 in that order against a Paperclip
with no company in it. They need more than the other steps do: Ollama running
with the models qwen2.5:3b and qwen2.5:14b, and the ddgs library
(pip install ddgs). They make a company of their own, Research Desk. The paper
they work on is paper.txt, a short research paper for a software team, and the
example paper of the second and third runs is model_paper.txt.

Run them again and you will not get our words back. The researcher searches
the web, the web answers differently on another day, and the model writes from
what it is given. What should repeat is the shape: a quality score from 1
to 100 for each round, a stop at the target or after three rounds, and the best
paper at the end.

Tutorials 1 to 4 and their checks were run on 9 October 2026, and long_text.py
and Tutorial 5 on 10 October 2026, all on one Windows 11 laptop. Nothing was
run on a Mac.