Operational artifact list, run commands, and protocol boundaries for independent technical review.
Reader's guide:
1) HSBC Fraud Pilot Report (story + tables) → 2) this page (commands + pitfalls) → 3) JSON artifacts in Competition 2/Cursor/ (e.g. cursor_hsbc_pcqrc_fd_ibm.json) → 4) Highlights · Home
| Path | Purpose |
|---|---|
~/Desktop/Competition 2/data/ | Kaggle IEEE-CIS source CSV files |
~/Desktop/Competition 2/Cursor/ | Canonical scripts, JSON outputs, concept notes |
~/Desktop/Competition 2/venv/ | Python environment used for all runs |
| File | Role |
|---|---|
train_transaction.csv | Primary transaction records (contains isFraud) |
train_identity.csv | Identity/device join data by TransactionID |
test_transaction.csv | Kaggle test package file (not used for labeled pilot metrics) |
test_identity.csv | Kaggle test identity package file |
sample_submission.csv | Kaggle template file |
| Artifact | Contains |
|---|---|
cursor_hsbc_pcqrc_fd_tuning.json | Architecture sweep (8q/12q/16q), seed screening, main simulator metrics |
cursor_hsbc_matched_subset_xgb.json | Matched-slice XGBoost baseline (same MI/split protocol) |
cursor_hsbc_pcqrc_fd_ablation.json | K-ablation at fixed 12 qubits/depth (K=1,2,3,4,6) |
cursor_hsbc_pcqrc_fd_ibm.json | IBM Fez run, job IDs, capped-slice hardware metrics |
cursor_hsbc_classical_baseline.json | Full-split classical reference metrics |
Always execute from ~/Desktop/Competition 2/Cursor with the shared venv:
cd ~/Desktop/Competition\ 2/Cursor
../venv/bin/python cursor_hsbc_pcqrc_fd_tuning.py
../venv/bin/python cursor_hsbc_matched_subset_xgb.py
../venv/bin/python cursor_hsbc_pcqrc_fd_ablation.pyIBM hardware pilot (archived May 2026 — test ROC-AUC 0.7717, PR-AUC 0.1750 on capped 48/12/48, ibm_fez max-qubits mode, 156q/12c/d6, readout z+ctxpool):
unset QISKIT_IBM_INSTANCE
export QISKIT_IBM_INSTANCE=open-instance
cd ~/Desktop/Competition\ 2/Cursor
../venv/bin/python cursor_hsbc_pcqrc_fd_ibm.py \
--mode ibm --backend ibm_fez \
--max-qubits-mode \
--depth 6 \
--seed 1000 --seed-sweep 1 \
--readout z+ctxpool \
--runs-per-batch 1 --use-qpc-noise-reducer \
--batch-size 4 --shots 2048 \
--cap-train 48 --cap-val 12 --cap-test 48Do not type literal ... on the command line. Adjust caps/shots if you need a cheaper smoke test (--ibm-fast-defaults is an alternative preset).
IBM dependency: ../venv/bin/pip install -U 'qiskit-ibm-runtime>=0.46'
IBM credentials: quantum token in QISKIT_IBM_TOKEN or ~/.ibm_quantum_token; IAM key optional on some accounts via QISKIT_IBM_IAM_API_KEY. Instance routing is documented in README_CURSOR_WORKSPACE.md (open-instance, CRN file precedence, aggregate fallback).
Artifact: cursor_hsbc_pcqrc_fd_ibm.json contains metrics and full job-id arrays (ibm_job_ids_all).
156-qubit / noise-reducer stack: --max-qubits-mode matches Fez logical width; readout mode z+ctxpool keeps feature vectors manageable. qpc_noise_reducer.py is an optional in-repo Python helper loaded when you pass --use-qpc-noise-reducer; it supports count aggregation across --runs-per-batch repeats (readout-error mitigation inside it is only practical for smaller widths).
Teams often lose hours here until routing is stable:
QISKIT_IBM_INSTANCE overrides ~/.ibm_quantum_instance_crn in the HSBC script — stale CRN files can confuse debugging.open-instance) when using open-plan time.ibm_quantum_platform (not deprecated names).The successful archived pilot used open-instance after pinned discovery quirks and executed 27 Sampler batches at max logical width; job IDs are in cursor_hsbc_pcqrc_fd_ibm.json.
TransactionDT (train earlier than test).z+ctxpool readout at device logical width.Interpretation boundary: capped IBM metrics are evidence of hardware execution and workflow, not direct evidence of production ROI or superiority vs full-data classical systems.