README.md

SAC-QUTAB #

A frozen six-run EuroSAT RGB study of prediction, representation, calibration and failure detection: ResNet-50 and ViT-Small/16, seeds 17, 29 and 43.

Install and reproduce headline numbers #

Use source release v1.0.1-final (or current main) and Python 3.12 in a fresh environment. Repository attributes preserve the hash-pinned file bytes on Windows, Linux and macOS:

git clone --branch v1.0.1-final https://github.com/GeminusF/sac-qutab.git
cd sac-qutab

Create the environment:

python -m venv .venv
# Linux/macOS: source .venv/bin/activate
# Windows PowerShell: .venv/Scripts/Activate.ps1
python -m pip install -r requirements-ci.txt

After setup, one command downloads the hash-pinned numeric evidence and recomputes the headline estimates:

python scripts/reproduce.py --mode headline

The evidence URL targets release v1.0-final. Until that release is published, use --evidence /path/to/all_artifacts with the original frozen extraction. A missing release fails explicitly; it never substitutes synthetic results.

Model Clean accuracy Severe-resolution accuracy Mean detector AP gain
ResNet-50 98.63% 12.94% 0.432
ViT-Small/16 96.77% 43.00% 0.402

The replay recalculates accuracy from saved probabilities and detector AP from calibrated coefficients. It does not train, run neural inference, or re-estimate bootstrap intervals. Output is written to output/headline.json. The ~134 MiB compressed evidence contains authenticated numeric outputs, not raw images or checkpoints.

Tests #

# PowerShell: $env:PYTHONPATH='src'
# Linux/macOS: export PYTHONPATH=src
python -m sac_qutab validate-config --config configs/core.yaml --campaign configs/sac-campaign-v1.yaml --world-size 1
python -m pytest -q -p no:cacheprovider
python scripts/check_submission.py

CI runs CPU checks only. GPU training and deployment are not part of CI.

Full experiment and report #

Report compilation uses retained report/figures/ assets and needs no external materials folder. For a full numeric report audit, set SAC_EVIDENCE_ROOT to your authenticated full extraction. The original prepare -> freeze -> finalize training protocol and all frozen scientific identifiers remain unchanged.

Limitations and contribution record #

There was no human semantic-preservation review. Model pretraining and training hardware differ; training throughput is not a controlled architecture comparison. Seed and domain coverage are limited. Synthetic tests establish wiring, not scientific conclusions.

This repository is a publication of an already completed study. The contribution PDF records the equal publication assignment agreed by all five members, as confirmed by the team representative. Actual GitHub activity is obtained with python scripts/audit_github.py; it is not inferred from the assignment. The local account-switching controller is intentionally outside the repository.