BENCHMARKS.md

KODEX — BENCHMARKS

Honest numbers only. Each figure is computed live (fixed seed, CPU, no network) or sourced from a named on-disk file. Pre-registered misses are kept, not hidden. Nothing here is published until go-live.

#	Code	Status	Headline benchmark
01	KAIROS	BUILT	MPC tracking 0.0461 (<5%); Exp-B 0 escapes / 150,000; 77.8 µs/step
02	KFLOW	BUILT	GNN RMSE@k8 0.0844 vs naive 0.1721 (beats naive; AC-18 0.05 bar MISSED, kept)
03	KGATE	BUILT	model-free clamp: 0 escapes / 150,000 steps
04	KHALO	BUILT	sourced ECE 0.0327 +/- 0.0136 / cov90 0.888 +/- 0.034; live GP ECE 0.0771
05	KMAT	BUILT	DFT-validated CHGNet screen, 8 candidates
06	KOIL	BUILT	live strain surrogate rel-L2 0.0038; sourced 0.24% @0.15 ms, quench AUC 0.9998
07	KORE	BUILT	learned NN equilibrium: rel-L2 0.0035 vs analytic, 0.318 ms/field; sourced FNO 2.5% vs 1% bar
08	KQUBIT	BUILT	REAL VQE (PennyLane): recovers H2 ground state to 0.0002 mHa on default; runs on real QC hardware — no advantage yet
09	KWARD	BUILT	real-device AUC 0.98 (591 MAST shots), ECE 0.0349; independent-precursor AUC 0.975 (n=1 Mirnov + P_rad ONLY, label-independent); labels heuristic (Ip-quench)
10	KYRO	BUILT	complete 16/16 map (12 turbulent / 4 quiet (16/16 complete)); turbulent/quiet 16/16 correct, R²=0.858 / cov90=0.875; 0.257 ms vs 2.89 GPU-h/pt (μ=400)
11	KBREED	BUILT	TBR surrogate R²=1.0 vs neutronics engine; nominal net TBR 0.742; OpenMC = auto fidelity upgrade
12	KBURN	BUILT	dispatch surrogate R²=1.0 (fuel mix → captured power); +22% dynamic
13	KEYE	BUILT	virtual-sensor fault detection 0.767 at FAR 0.0
14	KFLUX	BUILT	coil-fluence surrogate R²=1.0 vs neutronics engine; OpenMC = auto fidelity upgrade
15	KFUSE	BUILT	multi-fidelity: RMSE 1.481→1.05 (R²=0.787); 3rd fidelity = real-mass gold (deferred)
16	KHEAT	BUILT	heating/current-drive surrogate: driven-current I_cd R²=0.949 over 3888 configs (reduced CD; RF/NBI ray-tracing = upgrade)
17	KISO	BUILT	servo R²=0.723 + medical-isotope yield from real FENDL-3.2 σ(E) (6 isotopes, Mo-99/Tc-99m flagship)
18	KMIT	BUILT	REAL error mitigation: ZNE cuts H2 energy error 44.38→5.029 mHa (0.887 removed); readout TV 0.1472→0.0 — real NISQ tool today, role post-~2036
19	KPATH	BUILT	MPC-imitation controller R²=0.879 over 6 axes
20	KQERN	BUILT	REAL quantum kernel: MAST-disruption AUC 0.919 vs classical 0.929 (ties — honest null); runnable on real QC hardware
21	KQROSS	BUILT	REAL FT resource estimator: classical↔quantum crossover ~N=50 needs ~6e+04 physical qubits, roadmap ~2032 — no FT advantage this decade
22	KSENSE	BUILT	REAL quantum metrology (GHZ, PennyLane): ideal Heisenberg 4.0× vs dephased 2.05× ≈ SQL 2.01× — dephasing erases the advantage (honest null)
23	KTENSOR	BUILT	SVD/MPS ROM of the real CGYRO flux DB: effective rank ~2.93 of 4 — modestly compressible, NOT strongly low-rank (honest characterization; classical, no quantum advantage)
24	KBENCH	BUILT	open fusion-ML benchmark suite: 3 citable tasks (CGYRO turbulence, MAST disruption, flux ROM) with real data + KODEX baselines
25	KDRIVE	BUILT	RL policy (cross-entropy): tracking 0.0551 vs naive 0.1225, twin-in-the-loop
26	KDYN	BUILT	REAL Trotter quantum dynamics: 1st-order error 1.34637→0.03381 over steps (~1/n); runs on real QC hardware — honest cost curve, no advantage yet
27	KECON	BUILT	generic LCOE calculator + breakdown — FINANCIAL FIREWALL (no Kronos numbers)
28	KEDGE	BUILT	divertor edge surrogate: heat-flux→target-temp R²=1.0, CuCrZr limit q~13.5 MW/m² (reduced 0-D; SOLPS/EIRENE = upgrade)
29	KFORGE	BUILT	inverse-design chained KYRO+KORE → found a quiet operating point (a/L_T=2.228, shear=1.511, Q_tot≈0.0)
30	KFUEL	BUILT	fuel-cycle balance: self-sufficient=True, doubling 5.17 yr (illustrative TBR)
31	KGEN	BUILT	generative (PCA latent, dim 8); samples plausible equilibrium fields (diffusion = roadmap)
32	KLAW	BUILT	equation discovery: 7 terms, R²=0.903 — rediscovered the critical-gradient onset from the CGYRO map
33	KLINQ	BUILT	REAL quantum linear solver (VQLS): 4×4 reduced-MHD solve fidelity 1.0 vs classical (rel-resid 0.0); FT-HHL 4e+04 phys qubits — no advantage, role post-2036
34	KMIND	BUILT	foundation model: one shared trunk across 4 domains (66958 params); transfer vs matched single-task — pos ['equil'], neg none, neutral ['state', 'disrupt'] (transport data-starved) — honest multi-task result
35	KPILOT	BUILT	fleet router: plain-language query → the right code (LLM agent = roadmap upgrade)
36	KQOPT	BUILT	REAL QAOA optimizer: 1.0 of optimum on a design QUBO (6 qubits); plug in your QUBO, run on real QC hardware — no speedup yet
37	KRAD	BUILT	impurity-seeding radiation evaluator: 3 impurities (Ar, N, Ne); least core dilution = Ar
38	KSCOUT	BUILT	multi-fidelity campaign: tight 15.0 GPU-h over 16 hotspots → mean confidence ρ 0.638 vs 0.281 all-μ=400 (2.27×); triage-then-escalate
39	KSEEK	BUILT	active-learning: proposes next CGYRO run (a/L_T=2.225, shear=0.58); LOO std-vs-error corr -0.283 (honest: weak on 16 pts)

Regenerate: PYTHONPATH=<kronos-ml>:<kronos-toolkit> python benchmarks/run_all.py.