README.md

1# Quantum Flywheel: exact normalized-MIP prototype 2 3**Applicant:** Joseph W. Anady / ThatAIGuy Research 4**Version:** QN-MIP-20261012-v1 5**Prepared:** 12 October 2026 6**Status:** completed local numerical experiment; four-page report; no submission or cloud/QPU execution. 7 8## Main deliverable 9 10`report/Quantum_Flywheel_MIP_Prototype_4_Pages.pdf` is the four-page report for the form's PDF upload field. It includes the completed experiment, numerical results, limitations, references and a proposed three-month extension. `report/prototype.tex` is its editable LaTeX source. The original 17-page manuscript is background only and is not included in this packet. 11 12## What was actually executed 13 1460 randomly generated **linear stochastic recurrent network models**, not trained language models or biological data. Four sizes (8, 10, 12, 14 nodes), three families (dense, two-module, sparse plus ring), five network seeds per size/family. Stationary Gaussian population covariances are computed analytically. Every nonempty proper bipartition is enumerated, modulo complement symmetry: 163,140 total partition evaluations. 15 16The objective is explicitly 17 18`Phi_SI(S) = H(X_next,S | X_now,S) + H(X_next,B | X_now,B) - H(X_next | X_now)` 19 20`K(S) = min(H(X_now,S), H(X_now,B))` 21 22`J(S) = Phi_SI(S) / K(S)` 23 24where B is the complement. Natural logs; unit innovation covariance; fixed variable units. The minimized score J is distinct from the raw information loss Phi_SI at its minimizing partition. This is a specified observational stochastic-interaction bipartition problem, not a calculation of full IIT 3.0/4.0 or a consciousness assay. 25 26### Observed exact-MIP recovery 27 28| Method | Exact recoveries | Rate | 29|---|---:|---:| 30| Queyranne on normalized J | 11/60 | 18.3% | 31| Simulated annealing | 515/600 | 85.8% | 32| Simulated QAOA, p=1 | 410/600 | 68.3% | 33| Simulated QAOA, p=2 | 457/600 | 76.2% | 34| Uniform, 256 draws | 214/600 | 35.7% | 35 36Queyranne has one deterministic run per network. Other methods have ten repetitions per network. QAOA repetitions are readout batches from one trained circuit per depth/network, not independent optimizer runs; confidence intervals resample networks within fixed size/family strata. Annealing repetitions are independent searches, each including four restarts. All denominators and budgets are preserved in the report. 37 38QAOA uses an exhaustive diagonal cost truth table and exact expectation evaluation during classical parameter optimization. It does not constitute an efficient hardware encoding. Its table construction already solves the reference optimization. No quantum advantage, scalable QUBO reduction, noise robustness, or QPU result is claimed. 39 40## Files 41 42- `protocol.json`: fixed primary protocol, written before the experiment; root seed 20261012. 43- `environment.json`: execution platform and pre-execution SHA-256 hashes of `benchmark.py` and `protocol.json`. 44- `execution_summary.json`, `run.log`: actual elapsed-time and progress records. 45- `code/benchmark.py`: network generation, independent Gaussian oracles, exhaustive search, Queyranne, annealing, statevector QAOA, sampling and initial correctness checks. 46- `code/analyze.py`: raw-result aggregation, stratified network-bootstrap intervals, complete saved-outcome audits and post-run submodularity checks. 47- `code/diagnostic_unnormalized.py`: explicitly post-run ablation; no alteration of primary outcomes. 48- `code/reproduce.py`: rerun in a new workspace without overwriting the evidence. 49- `code/verify_bundle.py`: standard-library-only checksum verification. 50- `results/instances.csv`: one row per network, full-precision values and timings. 51- `results/trials.csv`: 2,400 stochastic trial records (600 each for SA, uniform and each QAOA depth). 52- `results/summary.json`: report-level statistics, intervals, size strata and paired contrasts. 53- `results/n*.json`: 60 detailed network result records, including optimizer termination statuses. 54- `data/n*.npz`: 60 numerical instance archives (all arrays use standard NumPy storage; no pickle objects). 55- `validation/`: initial tests, saved-outcome audit, all submodularity witnesses and the unnormalized diagnostic. 56- `CHECKSUMS.sha256`: hashes of every included payload except the checksum file itself. 57 58### NPZ schema and bit convention 59 60`W` is the generated recurrent matrix, `Sigma` its stationary covariance, and `node_permutation` the applied relabeling. `J`, `Phi_SI`, and `normalizer` have length `2**(n-1)-1`. Array index `z` encodes the canonical partition mask `(z << 1) | 1`: bit zero (paper node 1) is always on side S. Node labels in the PDF are one-based; mask bits and witness indices are zero-based. Bits set to one are in S, other nodes are in B. The all-ones full-system mask is invalid and excluded from the reference tables. 61 62`optimal_z` lists all optima within `1e-10 + 1e-8*abs(Jstar)` (one per instance here). `qaoa_cost` has one extra entry for the invalid all-in-one state, assigned cost 2; other entries equal `J/max(J)`. `qaoa_p1_probabilities` and `qaoa_p2_probabilities` include the invalid state. Shot arrays have shape `(10,256)`. Invalid QAOA shots consume the budget and are not replaced. `qaoa_p*_theta` stores alternating gamma/beta parameters; `qaoa_p*_optimizer_history` stores rows `[restart, penalized_energy, gamma0, beta0, ...]` for every evaluated parameter vector. Optimizer selection minimizes the penalized expectation, never the optimum's probability. 63 64## Verify and reproduce 65 66From this folder, verify the delivered bytes without installing dependencies: 67 68```sh 69python code/verify_bundle.py 70``` 71 72The execution used Python 3.13.5 with the versions pinned in `requirements.txt`. Create an isolated environment and install them, then reproduce in a **new, nonexistent** directory: 73 74```sh 75python -m venv .venv 76. .venv/bin/activate 77python -m pip install -r requirements.txt 78python code/reproduce.py --output ../mip_reproduction 79``` 80 81The wrapper copies the implementation and fixed protocol to the new workspace, runs the full benchmark and analysis, and leaves this delivery unchanged. To run only the implementation correctness checks: 82 83```sh 84python code/benchmark.py --validate-only 85``` 86 87`benchmark.py` resumes existing instance results rather than overwriting them. Running it directly in the delivered folder is therefore **not** a fresh replication. Re-running analysis can change generated timestamps or floating-point output on a different stack; do not mistake a new checksum file for authentication of original evidence. Seeded outcomes should be repeatable with the same numerical stack, but floating-point optimizer trajectories and wall times may differ across platforms. 88 89Optional report compilation (requires a LaTeX installation with the packages in `prototype.tex`): 90 91```sh 92cd report 93pdflatex -interaction=nonstopmode -halt-on-error prototype.tex 94pdflatex -interaction=nonstopmode -halt-on-error prototype.tex 95``` 96 97## Validation and provenance boundaries 98 99Initial tests: 70 weighted cuts and 35 Gaussian mutual-information examples matched exact Queyranne minima; QAOA matched independent dense matrix exponentials; zero-interaction and conditional/joint log-determinant identities passed. The output audit checked all stored trial scores/hits against their saved objective tables and all QAOA readout batches against their stored shots. 100 101Post-run diagnostics (not part of the frozen primary protocol): Queyranne found 60/60 unnormalized stochastic-interaction minima; normalization changed the exact MIP in 37/60 cases. All 60 normalized tables violate submodularity, with explicit witnesses retained after scanning 6,804,480 elementary inequalities. These findings are not presented as preregistered hypotheses or as a general hardness proof. 102 103The sums of per-instance times are 87.24 seconds on the recorded environment; setup for initial tests and JIT compilation is separate. The report's timing table does not substitute QAOA readout time for full preparation/training time. CPU model naming does not imply all 80 cores were used: execution was serial with a single BLAS thread. 104 105This archive contains newly generated synthetic benchmark evidence, not recovered Pantheon runtime data, original manuscript measurements, external quantum hardware logs or private account exports. No API credential is required, embedded, or used. No files were uploaded to the RFP and no submission was made. 106