Every headline number on this site, where it comes from, and the command that reproduces it. Figures are dated because data sources and package versions move; a rerun on a later date can legitimately differ, and the source files say how. Test counts are not listed here: they change with every commit, so the site reads them from each repository's CI instead.
Compiled 2026-10-28.
How RL agents fool you (RL-Trader)#
| Claim | Figure | Source | Reproduce |
|---|---|---|---|
| A single favourable run was not an edge | crypto agent +56.8% [+16.3%, +106.3%] across 10 seeds vs buy-and-hold +33.5%, permutation p = 0.79 | RESULTS.md §5 | python tools/real_significance.py --seeds 10 --timesteps 150000 |
| The stock agent is reliably worse than holding | −19.1% [−26.7%, −11.2%] vs +239.5%, p = 0.0021 | RESULTS.md §5 | same command |
| Memorisation, and the fix | held-out return single-asset → across tickers: stocks −25% → +31%, crypto −44% → +187% (10 seeds, all four intervals exclude zero) | RESULTS.md §1 | python tools/ablation_multiseed.py --seeds 42 43 44 45 46 47 48 49 50 51 --timesteps 60000 |
| Pre-registered replication | predictions committed (commit d344b4e) before running seeds 200–209; all four held (stock p = 0.0021, crypto p = 0.88). Not replicated: the crypto mean's interval now includes zero, +26.7% [−13.8%, +70.4%] |
docs/preregistration/ | command in the registration file; input data archived by SHA-256 |
QuantumSafe Scan#
| Claim | Figure | Source | Reproduce |
|---|---|---|---|
| Precision on the project's labeled corpus | 100% precision and recall on 26 labeled findings (18 files, 9 languages); naive line-regex 49.1% | benchmark/RESULTS.md | python benchmark/evaluate.py |
| Recall under mutation | 50/50 seeded cases found, 0 false positives on 50 negative mutations | benchmark/RESULTS-seeded.md | python benchmark/seeded.py |
| Precision on real code nobody chose | 93% (95% CI 86–97%) on 100 random findings from 37 PyPI packages, each hand-labeled with a reason; measured before the TLS rule fix it exposed | benchmark/RESULTS-precision.md | python benchmark/precision_sample.py --score |
| Scale | 5,609 findings (4,236 HIGH) across 11,275 files in 37 packages, 2026-10-28, scanner 0.3.1 | benchmark/RESULTS-realworld.md | python benchmark/realworld.py |
| Versus general security tools | quantum-relevant findings caught on the labeled corpus: QuantumSafe 26/26, Semgrep (p/r2c-security-audit) 4/26, Bandit 2/26 |
benchmark/COMPARISON.md | python benchmark/comparison.py |
AlphaGlyph#
| Claim | Figure | Source | Reproduce |
|---|---|---|---|
| The best backtest in a search is probably overfit | probability of backtest overfitting 72% over a 42-variant grid on SPY (CSCV); median out-of-sample Sharpe 0.65, ~52% of in-sample | technical report §7 | python tools/pbo_snapshot.py (pinned SPY snapshot, SHA-256 recorded) |
| The ML forecaster cannot predict direction | test AUC ≈ 0.51 | technical report §7 | model metadata, ml_model_meta.json |
| One conditional positive, and how fragile it is | Dip Buyer on AAPL/MSFT, 2021-01-01 → 2024-01-01: +21.6% vs SPY +5.6% in ranging markets; one more year shrinks the edge to +3.6 points | README, technical report §7 | backtester with those dates; regime_conditional in the result |
AlphaZero-style chess engine#
| Claim | Figure | Source | Reproduce |
|---|---|---|---|
| The plateau was a training-signal problem, not compute | Elo vs a random player: baseline −21, 10× self-play −7, material signal +28 | docs/experiment-results.md | python run_experiment.py --device cuda (9.8 h on 2× T4) |
| Strength of the classical control engine | Easy (depth 2) ≈ 1390 [1284, 1512] on Stockfish's engine scale: +14 =8 −8 vs Stockfish 19 at UCI_Elo 1320 |
docs/data/stockfish_rating_d2_e1320.json | python tools/stockfish_rating.py --stockfish <path> --depth 2 --games 30 |
Tradeski and Neural Canvas#
These two make engineering claims rather than research ones: Tradeski's indicators are hand-coded on NumPy and covered by its CI (Python and JavaScript suites); Neural Canvas runs MediaPipe's 21-landmark hand model on-device at roughly 60 frames a second. Both are live and can be checked directly.
Accessibility#
All seven sites report zero violations under axe-core 4.10 with the WCAG 2.1 A/AA and best-practice rule sets (main pages and views, 2026-10-28).