XYPHERS THAIM REWARD REPRODUCTION

The apparatus needs Python 3 with numpy 2.0.2, scipy 1.13.1 and python-flint
0.6.0 (source/requirements.txt). From source/:

  python3 -m venv venv && venv/bin/pip install -r requirements.txt
  venv/bin/python tests.py
  venv/bin/python run.py > my-report.txt

tests.py runs only on worlds outside the experiment, in about a minute. The
run trains every agent twice, freezes and evaluates every table exactly, and
takes about twenty-five minutes. Its report is 138 lines and ends with

  H1 SUPPORTED failing=none
  H2 SUPPORTED failing=none
  H3 SUPPORTED failing=none
  H4 REFUTED failing=['W1']
  H5 SUPPORTED failing=none

  OVERALL PASS

To compare with the recorded run:

  diff ../report.txt my-report.txt

Training is in floating point; on the same platform (Darwin arm64, the
versions above) the recorded run is reproduced exactly. The recorded report's
SHA-256 is

  a419d1cf735e10daea9b77d083ae08535fb3d9c028eaaeeb9e8486a3c7a8f0a4

The run also writes the frozen tables to source/tables/; compare them with
../tables/.

The independent check needs python-flint. It expects the repository's
layout, which the packet recreates in three commands. From this directory:

  mkdir -p layout/derivable layout/xypher-thaim-reward-proof layout/xypher-thaim-reward-check
  cp BOUNDARY.md layout/derivable/xypher-thaim-reward-boundary.md
  cp report.txt layout/xypher-thaim-reward-proof/ && cp -R tables layout/xypher-thaim-reward-proof/
  cp independent/check.py layout/xypher-thaim-reward-check/
  python3 layout/xypher-thaim-reward-check/check.py > my-check.txt
  diff independent/output.txt my-check.txt

It recomputes every result from the frozen tables in about forty seconds,
ends with OVERALL AGREE, and no output from diff means it reproduced the
recorded check exactly.