Article title:
Can One Agent Restore Another? Multi-Agent Verification Through Independent Constraint Sources

Journal:
Autonomous Agents and Multi-Agent Systems

Author:
Simin Yuan

Affiliation:
Independent Researcher

Residence:
Changsha, People's Republic of China

Corresponding email:
yleven120@gmail.com

Online Resource:
Online Resource 1 – Supplementary Code, Raw Outcomes, Controls, and Deterministic Trajectory Replays

DIRECTORY MAP
code/  Experiment and analysis scripts.
data/  Raw per-run outcomes and deterministic trajectory replays.

CODE FILES
code/exp1_pilot_1p5B.py
  Disclosed pilot (C versus S, n=10, five-exchange window). Excluded from all analyses and archived for transparency.
code/exp2_1p5B_coupled_vs_single.py
  Main 1.5B coupled and single-agent experiment.
code/exp3_1p5B_context_matched.py
  1.5B context/volume-matched condition and three-group comparison.
code/exp4_1p5B_dose_response.py
  1.5B four-fold-volume condition and dose-response summary.
code/exp5_scale_3B_7B.py
  3B and 7B tiers under all four conditions; set MODEL_NAME as documented in the script.
code/exp6_verify_7B_controls.py
  7B no-perturbation control and per-agent replay checks.
code/exp7_replay_7B_trajectories.py
  Exact replay of all eight recovered 7B coupled runs and stable/transient classification.
code/make_figures.py
  Regenerates the two manuscript figures from the embedded outcomes; no GPU is required.
code/analysis_mcnemar.py
  Reproduces the exact McNemar p-values from the supplied paired outcomes; no GPU is required.

DATA FILES
data/raw_outcomes_v2.txt
  Per-run recovery delays for all tiers, the pilot archive, and validity-control records.
data/replay_output_7b.txt
  Full per-exchange violation trajectories for the eight replayed 7B coupled runs, including stability labels.

REPRODUCTION ORDER
1. Install the environment listed below.
2. Run exp1 through exp5 in numerical order to reproduce the pilot, main 1.5B study, volume controls, and tier extensions.
3. Run exp6 for the 7B validity controls.
4. Run exp7 for deterministic trajectory replays.
5. Run analysis_mcnemar.py to reproduce the paired exact tests.
6. Run make_figures.py to regenerate the figures.

SOFTWARE AND ENVIRONMENT
Google Colab; Python with transformers, accelerate, and scipy. Install once per session:
  pip install -q -U transformers accelerate scipy

HARDWARE
NVIDIA T4 for the 1.5B tier; NVIDIA L4 for the 3B and 7B tiers. Half precision is used. Do not use 4-bit quantization because it would change the comparison. Approximate runtime documented by the study is three T4 GPU-hours for the 1.5B tier and five L4 GPU-hours for the 3B and 7B tiers.

DETERMINISTIC CONDITIONS
Greedy decoding (do_sample=False), a cap of 40 new tokens per call, and fixed opening prompts make every run a deterministic function of its opener, subject to hardware-level numerical variation. The same 30 opening prompts are paired across conditions and tiers. The replay script verifies recovery-delay fidelity for all eight recovered 7B coupled runs (8/8).

NOTES
Comments and console strings in the scripts are English translations of the scripts as run. Experimental logic, parameters, prompts, opening messages, raw results, and reported outcomes are unchanged. One library keyword was renamed for current versions (torch_dtype to dtype) without changing behavior.
