Grammar Fingerprinting Research Archive

Blind topology classification, Fisher-threshold analysis, IBM cross-platform data, and transfer modelling from raw quantum readouts.

Author: Dániel Csaplár — Independent Researcher, Kazincbarcika, Hungary
ORCID: 0009-0000-7362-7232
Archive date: March 2026

---

What this repository contains

1. Sycamore (Dryad): Grammar fingerprints from raw readout sequences; blind Ward clustering vs. hardware topology; strong controls (shuffle, LSB, baselines); data-length threshold near ~8k samples for classification signal.
2. Fisher information geometry: Per-readout sweeps over data length; scalar Fisher metrics and estimated transition thresholds (N*); multi-seed runs with median + IQR reporting.
3. IBM Quantum: Raw shot collection (collect_ibm_shots.py), grammar learning, IBM-specific Fisher sweeps, optional Marrakesh vs. Torino comparisons.
4. Temporal drift (IBM): Distribution-level and grammar-matrix KL comparisons between archived and refreshed shot files; marginal / Hamming summaries for high-q regimes.
5. Transfer / universality: Ridge-style models linking IBM and Sycamore threshold estimates; clustering / purity reports; optional DOCX report builders.
6. Human EEG (eeg/): Same pipeline (unmodified parameters) applied to the PhysioNet eegmat dataset — 36 subjects, eyes-closed rest vs. first-minute mental arithmetic. Recovers the classical alpha-suppression topography from a substrate-agnostic information-theoretic metric. See eeg/README.md for details.

Not tracked in Git (see .gitignore): Dryad-scale readout_raw_data/; IBM results/ibm_raw_shots/; run logs; large *.npz caches; exploratory IBM subset analyses (10_20, no5q, raw4 filename patterns); extra Fisher phase-transition PNGs (publication figures remain under results/fisher_figures/).

---

Data sources

Platform	Role	Notes
Google Sycamore	Primary readouts for validation & Fisher	Arute et al., Nature 574, 505–510 (2019) — Dryad CC0, 10.5061/dryad.k6t1rj8. Files: readout_raw_data/*_readout_raw_data.txt or results/readout_raw_data/.
IBM Quantum	Cross-platform shots & thresholds	Collected via qiskit-ibm-runtime (not in minimal requirements.txt). Use QISKIT_IBM_TOKEN (see code/local_ibm_env.ps1.example then copy to local_ibm_env.ps1, gitignored).

---

Core pipeline (all platforms)

1. Load integer readout / shot column (output).
2. Z-normalize then SAX (default alphabet 7, quantile breakpoints; signal_processing.py).
3. Train character LSTM (grammar_learner.py: hidden, seq_len, epochs as per script).
4. Extract grammar: row-normalized transition matrix T (next SAX symbol given current).
5. Downstream: clustering (Frobenius + Ward), Fisher metric on T, KL between successive N, or comparison across runs.

Ground-truth topology labels are used only for evaluation, not for training.

---

Sycamore: blind topology validation (summary)

• Claim: Temporal structure in raw noise encodes 1D Snake / 2D Block / Bulk topology; ~84.5% mean Ward purity (multiple seeds) vs. chance and strong controls.
• Controls: Shuffled time order to chance; LSB-only bitstream to chance; static SAX + LR/RF to chance; short training / low max_pts to no signal.
• Data-length sweep: Sharp emergence of signal around ~8,000 samples (see results/parameter_sweep_sax7.csv, validation_report.txt).

Main driver: code/run_validation_pipeline.py
Criteria: validation_criteria.json

---

Fisher information analysis (Sycamore, 28 readouts)

Script: code/fisher_information_analysis.py

• Trains the grammar learner at fixed N grid: 500 to 40000 (see script for exact steps).
• Builds Fisher information matrix from T; traces Fisher trace / det / max eigenvalue, KL between successive T(N), entropy, Frobenius distance from uniform.
• Estimates per-file N* (phase-transition-style heuristic; see script and fisher_information_analysis_*.txt reports).
• Multi-seed: --tag-with-seed --seed 0|1|2 gives output suffix _seed{k}.
• Queue helper: code/run_fisher_seeds_1_2_after_seed0.ps1 waits for seed 0, then runs 1 and 2.
• Aggregate: code/aggregate_fisher_threshold_seeds.py gives median + q25/q75 of N* across seeds.
• Publication table: code/build_fisher_median_iqr_publication_table.py writes results/fisher_threshold_median_iqr_publication.csv and .md.

Typical outputs:
results/fisher_metric_vs_datalength_all_readouts_seed*.csv,
results/fisher_estimated_thresholds_per_readout_all_readouts_seed*.csv,
results/fisher_phase_transition_all_readouts_seed*.png,
results/fisher_estimated_thresholds_median_seeds012.csv.

Quick test: python code/fisher_information_analysis.py --quick

---

IBM Quantum workflows

Script	Purpose
code/collect_ibm_shots.py	SamplerV2 jobs; saves results/ibm_raw_shots/ibm_<backend>_*.txt + metadata. --resume skips complete files.
code/run_ibm_grammar_learning.py	Fingerprints for all shot files to CSV + NPZ.
code/run_ibm_fisher_threshold_sweep.py	Per-circuit Fisher-style sweep on IBM shots.
code/run_ibm_ghz_vs_hadamardlayers_threshold.py	GHZ vs Hadamard/layers threshold comparison.
code/run_ibm_sycamore_protocol.py	Protocol-style report from IBM filenames.
code/analyze_ibm_fingerprint_clustering.py	IBM fingerprint clustering analyses.

Drift / two-run comparison (same backend, two dates):

Script	Purpose
code/compare_ibm_shot_runs.py	Raw outcome histogram: TV, Jensen-Shannon, sym-KL (sparse at high q).
code/compare_ibm_shot_marginals.py	Per-qubit P(1) drift + Hamming-weight TV/JS (interpretable at 20q).
code/compare_ibm_grammar_pairwise_kl.py	Train grammar on archive vs current; KL between T matrices.

---

Cross-platform and transfer modelling

• code/cross_platform_universality_report.py — combined IBM + Sycamore fingerprint summaries (Ward purity, regime tables).
• code/fit_threshold_transfer_model.py — OLS/Ridge-style transfer of threshold estimates across sources.
• code/fisher_robustness_visualization.py — Boxplots / curves for Fisher robustness.
• code/drift_monitor_kl.py — KL vs a reference grammar matrix from NPZ + rolling windows (prototype).
• code/build_ibm_threshold_report_docx.py, code/build_transfer_threshold_docx.py — Word reports from CSVs (optional python-docx).

---

Key results (headline numbers; full tables in results/)

The README lists summary numbers only. Authoritative rows are in results/*.csv and results/*.txt.

(Magyarul: a README-ben fő számok és fájlnevek vannak; a teljes táblázatok a results/ mappában — nem másoljuk be ide mind a 28 readout sort.)

Sycamore blind topology

Result	Value / note	File
Mean Ward purity (3 seeds)	~84.5%	results/robustness_audit_sax7.txt, results/validation_report.txt
Shuffled control	50%	results/shuffled_control_sax7.txt
LSB single-bit control	50%	results/readout_lsb_bit_control_sax7.txt
Data-length sweep	Signal near 8k (e.g. 92.86% at max_pts 8000 in 60-combo sweep)	results/parameter_sweep_sax7.csv
Pre-registered criteria	PASS	validation_criteria.json, results/validation_report.txt

Fisher (28 readouts, seeds 0, 1, 2)

Result	Note	File
Median N* + IQR	Many readouts in ~6k-10k band (aligned with ~8k regime); spread by readout	results/fisher_threshold_median_iqr_publication.csv, .md
Stable example	46q Bulk: N* = 750 all three seeds	same (stable_3seeds column)
Seed-sensitive	Large IQR e.g. 28q, 30q, 32q, 36q, 53q	results/fisher_estimated_thresholds_median_seeds012.csv
Curves	Fisher vs data length	results/fisher_metric_vs_datalength_all_readouts_seed*.csv, results/fisher_phase_transition_all_readouts_seed*.png

IBM Torino (archive vs refreshed shots, example)

Result	Representative value	File
Grammar sym-KL (11 circuits)	Mean ~0.008; mean Frobenius ~0.13	results/ibm_torino_grammar_pairwise_kl.csv
20q Hamming-weight TV	Mean ~0.031	results/ibm_torino_marginals_hamming_compare.csv
Raw joint histogram	Inflated TV/KL at 20q; use marginals + grammar	results/ibm_torino_run_compare_distributions.csv

IBM Fisher and transfer

Topic	File
Backend sweeps	results/ibm_fisher_sweep_marrakesh40960.csv, results/ibm_fisher_sweep_torino40960.csv
Normalized thresholds	results/ibm_fisher_thresholds_normalized.csv, results/ibm_fisher_threshold_backend_summary.csv
Transfer model	results/threshold_transfer_model_report.txt, results/threshold_transfer_predictions.csv

---

Repository layout (high level)

code/                  # All Python tooling
readout_raw_data/      # Preferred: Sycamore *_readout_raw_data.txt
results/               # All outputs: validation, Fisher, IBM, plots, DOCX, logs
  ibm_raw_shots/       # IBM shots; archive_* folders for drift baselines
validation_criteria.json
requirements.txt
code/local_ibm_env.ps1.example
README.md

Other code/ utilities: run_validation_pipeline.py, run_readout_lsb_bit_experiment.py, build_sycamore_readout_fingerprints.py, aggregate_fisher_threshold_seeds.py, test_threshold_transfer_cv.py, ...

---

Reproduction (cheat sheet)

pip install -r requirements.txt
# IBM: pip install qiskit-ibm-runtime qiskit  (and account / token)

Sycamore validation

python code/run_validation_pipeline.py --tasks sweep,report
python code/run_validation_pipeline.py --tasks shuffled
python code/run_readout_lsb_bit_experiment.py --output-bit-index 0 --epochs 50 --max-pts 10000 --seeds 0 123 999

Fisher (full 28 files, one seed)

python code/fisher_information_analysis.py --tag-with-seed --seed 0

Fisher seeds 0-2 + median table

powershell -NoProfile -ExecutionPolicy Bypass -File code/run_fisher_seed_sweep.ps1 -Seeds 0,1,2
python code/aggregate_fisher_threshold_seeds.py results/fisher_estimated_thresholds_per_readout_all_readouts_seed0.csv results/fisher_estimated_thresholds_per_readout_all_readouts_seed1.csv results/fisher_estimated_thresholds_per_readout_all_readouts_seed2.csv --out results/fisher_estimated_thresholds_median_seeds012.csv
python code/build_fisher_median_iqr_publication_table.py

IBM shots (after local_ibm_env.ps1 or env)

. ./code/local_ibm_env.ps1   # PowerShell
python code/collect_ibm_shots.py --backend ibm_torino --resume
python code/run_ibm_grammar_learning.py

---

Negative / null results (reported)

• IBM daily calibration summaries: No above-chance grammar topology signal from pre-aggregated calibration tables; dense raw shots are needed.
• Cosmological parameter encoding (Beck hypothesis): Tested; not supported.

---

Method note

Grammar Fingerprinting adapts industrial condition monitoring (SAX + sequence learning) to quantum readout streams. Development used AI-assisted coding (e.g. Cursor); design, runs, and interpretation remain the author's responsibility.

---

License

Code: MIT
Results / documentation: CC-BY 4.0

---

Contact

Dániel Csaplár
Kazincbarcika, Hungary
ORCID: 0009-0000-7362-7232