QUANTUM ENGINES Reproducible RFQ decision laboratory | 28 October 2026 RUN THE SAVED APPLICATION Extract source.zip. From its top-level directory run: python3 -m http.server 4791 --bind 127.0.0.1 --directory app Open http://127.0.0.1:4791 The saved application needs no installation, account, API or network feed. REPRODUCE THE EXPERIMENT Requires Node.js 20+ and Python 3 for the optional preview server. Measured reference runtime: Node v24.18.0 on Apple M1 Max, macOS arm64, 32 GB. No third-party packages are required. node --test work/engine.test.mjs node work/build.mjs This regenerates app/evidence.json, app/replay.json and work/synthetic-data.json. Timing values and generatedAt naturally change. Seeded inputs and policy results should agree, allowing normal floating-point variation between JS engines. The original run is preserved in outputs/evidence.json in this bundle. FILES app/engine.mjs Generator, filter, integration and decisions app/ui.mjs Browser interface and rendering app/benchmark-worker.mjs Fresh dedicated browser benchmark worker work/build.mjs Chronological experiment and evidence generation work/cold.mjs Fresh-process benchmark entry point work/engine.test.mjs Numerical, causal and simulation checks work/synthetic-data.json All 6,000 reproducible observations and evaluator truth outputs/evidence.json Saved run with seed, input SHA-256, metrics and timings PHONE DEMONSTRATION Open /#demo for a continuous interactive walkthrough without changing sections. Phones default to this view when no section is specified. The live quote controls, outcome, evidence, benchmark and original proposal are available on one page. The full laboratory also has a dedicated Research proposal view at /#research. FIVE-MINUTE DEMONSTRATION 1. Choose Buying pressure. Explain the distinction between observations and the uncertain state estimate. Nothing after the RFQ is visible until revelation. 2. Move the quote towards zero. More fills can mean negative expected value. Use model quote to balance spread capture, conditional markout and costs. 3. Change inventory. A sale reduces a long position but deepens a short one. Reveal one outcome. A single random result differs from an expectation. 4. Open Evidence. The information comparison keeps the optimiser fixed. Then run the actual CPU benchmark. Precomputation changes runtime. 5. Under the hood gives the equations, limitations and two next experiments. MODEL This is an ask-only, one-step decision model with known, specified parameters. Hidden directional state z is -1, 0 or +1, with a stationary Markov chain. The policy receives two Gaussian observations and the current inventory. It never receives realised z, private shocks, fills or future markouts. The filter deliberately ignores additional information in past midpoint moves. The Bayesian filter is an assumed-correct reference, not an empirically trained model. Historical context from Finn's research was read only. No existing research code, private RFQs, CV, portal or manuscript is included or reproduced here. Fill probability is sigmoid(1.7 - 1.15*quote + 0.9*z + 0.9*e). The five-second midpoint change is 1.5*z + 0.65*e + 0.65*epsilon. e and epsilon are independent standard normal shocks before fill selection. The shared private shock e makes conditional markouts adverse. Execution and hedge cost is 0.16 + 0.04*abs(q-1) basis points per fill. Inventory adjustment is 0.12*((q-1)^2-q^2) basis points per fill. Candidate sell quotes are 0.0 to 4.0 basis points in steps of 0.1, plus decline. The objective is expected fill times net marked value minus inventory adjustment. These are relative price basis points per unit notional, with no currency-P&L claim. PREVENTING LEAKAGE Seed 271828 generates 6,000 RFQs in order. Training IDs 0-2999 choose the fixed quote using realised training outcomes. Validation IDs 3000-3999 choose the first Simpson resolution in [16,32,64,128,256] meeting 1e-7 bp error and zero decision differences. Test IDs 4000-5999 are held out from both choices. Filters carry observations forward causally across partitions. There is no future normalisation, smoothing or fitted scaler. The precomputed table depends only on stated model parameters and the selected numerical resolution. RFQs are 10 seconds apart. Each markout arrives 5 seconds after its RFQ. Examples in the UI are selected by pre-decision posterior, not outcome value. Poisoned truth fields and future perturbation tests verify the policy boundary. EVALUATION State blind uses stationary state probabilities with the same inventory-aware optimiser. The fixed-quote baseline is selected solely on training outcomes. Expected values integrate future shocks conditional on held-out hidden states, and therefore use truth only in evaluation. Realised counterfactuals share the same fill uniform and price shocks across policies. This is possible because the synthetic environment assumes no policy feedback on the market. Inventories are independent snapshots sampled from -3 to +3. Consequently, averages are one-step decision metrics, not a self-financing strategy P&L. The paired 95% interval uses 1,000 block bootstrap resamples of 50 consecutive RFQs with seed 90210. It does not include model uncertainty. Fill calibration compares both predictors at the same training-selected quote. Per-policy Brier fields in the JSON are diagnostic at different selected quotes and should not be used as an apples-to-apples predictor ranking. NUMERICS AND TIMING Offline moments are A=E[fill] and B=E[fill*markout] for each of 3 states and 41 quotes. Online value is sum_z posterior[z]*(A*(quote-cost-risk)-B). Composite Simpson integration uses private shocks from -9 to +9. Normal tail probability excluded is below 2.3e-19. Validation selected 64 panels. The independent test reference recomputes the integrals with 1,024 panels. Tests also compare independent midpoint quadrature and seeded Monte Carlo. State uncertainty is unrelated to the small observed numerical discrepancy. Each saved timing method runs in a separate fresh Node process. Each browser method runs in a separate fresh worker. The first measured calculation is cold. It includes table creation for cached computation, then filtering, scanning 41 candidate quotes and selecting the best action for 128 consecutive RFQs. One untimed warmup precedes nine warm timed batches. Median and descriptive p95 are reported. The speed comparison uses direct and cached 64-panel computation at the same validated accuracy. The 1,024-panel result is an accuracy reference. Data generation, input decoding, module loading, rendering and network are excluded from calculation timing. The Node launch-to-exit duration is retained separately and includes the full multi-batch benchmark process. Batch time divided by 128 is average computation per RFQ, not tail latency or an individual request latency guarantee. This is CPU precomputation, not a GPU measurement, quantum speedup or fastest-possible classical implementation claim. PROPOSED EXTENSIONS No quantum circuits, physical QPU access or NVIDIA GPU execution are implemented. Quantum: compare small deterministic simulated-circuit features with strong classical/nonlinear/random features on identical splits, including total cost. GPU: test batches of 128, 1,024 and 16,384 scenarios on authorised hardware. Include compilation, allocation, transfer, synchronisation, precision, memory, decision error, amortisation and cold/warm total timing. Report a slowdown if seen. PRIMARY SOURCES https://arxiv.org/html/2509.17715v1 https://arxiv.org/html/2508.20225v6 https://arxiv.org/html/2608.02002v2 https://arxiv.org/html/2407.04510 https://arxiv.org/abs/2511.02136 https://arxiv.org/html/2412.19027v3 https://developer.nvidia.com/cuquantum-sdk https://developer.nvidia.com/cuda-q The first paper was read in full, including architecture, 4.2, 4.3 and appendix. Its hardware-noise and processing-order concerns are not solved by this toy. The demonstration does not reproduce those papers or establish algorithmic novelty. DEPLOYMENT Only the app directory is deployed to an isolated static Cloudflare Pages project. There is no server function, database, credential, analytics SDK or recurring job. The simulator and benchmark run on the visitor's CPU. Existing private sites, root site, email records and access boundaries are separate and unchanged. EXISTING GPU RESEARCH PROPOSAL research/GPU_Research_Proposal.pdf is Finn Smith's exact three-page research proposal and discussion draft of 27 October 2026, published with his authorisation. The PNG pages are renderings for convenient reading inside the application. The source manuscript has not been edited. It proposes adaptive precision, iterative refinement and factorisation refresh across sparse/KKT system sequences. The current HMM, quote grid and scalar Simpson moments do not implement or validate that method. Future constrained quote/hedge problems might supply such system sequences, subject to an explicit formulation and fair accuracy checks. The proposal's first study remains control or engineering constrained optimisation.