BEAST BOX PUBLISHED RESEARCH // 26 SEP 2026
THE QUANTUM
FEEDBACK ENGINE
STAGE 015
Real experiments. Public receipts. Both the positive engineering milestone and the negative scientific result.
01 / A WORKING EXPERIMENTAL PIPELINE
We connected the parts. Then we measured.
Historical observations from independently executed Azure-hosted Rigetti QVM simulator jobs passed through the existing COSMOS SOUL and signal-fusion path into BridgePacket, CNS7 and the evolving dyn12 state. A pinned public open-weight model then forecast independent future simulator observations.
01Rigetti cloud QVM
02SOUL + signal fusion
03BridgePacket
04CNS7 → dyn12
05Qwen forecast
06Blinded scoring
The 12D state was supplied as model-facing prompt context. This experiment did not change Qwen's neural weights, produce physical quantum measurements, or establish a new physics result.
02 / RESULTS FROM SIMULATED RUNS
Here's what actually happened.
Mean absolute error against the concealed ideal probabilities for two separately evaluated cohorts. Lower means less prediction error. Eight scenarios used genuine Azure cloud simulator outputs; 24 separate scenarios used reproducible local classical simulation.
| MODEL INPUT CONDITION | AZURE QVM (8) | LOCAL CLASSICAL (24) |
|---|---|---|
| No history | 0.29211 | 0.30329 |
| Raw history | 0.07498 | 0.11606 |
| Classical statistics | 0.02116 | 0.06810 |
| Actual COSMOS CNS7 12D | 0.15821 | 0.18743 |
| Shuffled CNS7 12D | 0.19209 | 0.15863 |
| Input-drive classical control | 0.18031 | 0.20548 |
NON-LLM CLASSICAL JEFFREYS POSTERIOR REFERENCEAzure: 0.02116 · Local: 0.02603
03 / THE HONEST SCIENTIFIC FINDING
The pipeline worked.
12D did not improve this forecast.
Providing interpretable classical statistics reduced prediction error in this task. The actual COSMOS/CNS7 12D prompt-context condition did not demonstrate additional predictive value over those same statistics. That is a meaningful measured limitation, not evidence of quantum advantage or AGI.
The model was explicitly told its data came from a simulator, so source-label responses were instruction compliance, not blind simulator-provenance detection. Also, the classical matched control transformed the input drive; matching the final state norm was not independently established.
Next hypothesis: test time-varying, independently measured external signals against equally informative classical filtering and rigorously matched state controls. This would be a separate preregistered experiment, not a reinterpretation of Stage 015.
04 / TRACEABLE EVIDENCE
Don't take our word. Check the receipts.
01Sanitized cloud simulator job receipt24 job identifiers, hashes, counts and source digest02Frozen Qwen generation transcriptAll 192 generated answers, error metrics and controls03Independent offline evidence auditSource hashes, local cohort and regraded model answers04Independent GitHub Actions verificationSuccessful offline audit; no fresh provider jobs
The GitHub receipts record cloud simulator executions, not physical QPU results. Artifact auditing does not independently authenticate Azure billing or account ownership.