Product: ExecutionProof™ — "If it cannot be verified, it cannot execute." Track: ARK authorization-boundary series — Verify-Before-Execute (VBE) characterization on IBM Quantum hardware Maintainer: Derek Hone — Remnant Fieldworks Inc.
This repository (executionproof-testbeds) holds the ARK authorization-boundary experiment series — a set of preregistered,
hardware-executed metrological tests of a verify-then-execute (VBE) authorization boundary on IBM
Quantum superconducting devices. Each experiment asks a narrow, falsifiable question about whether an
authorization decision can be reliably enforced at the moment of execution, and each is published with
its full preregistration, code, raw job data, and analysis.
This is applied verification-layer R&D (the ExecutionProof product lineage). It is distinct from
the Coherent Inheritance Framework (CIF) Phase 1 physics falsification program — formerly published under the working title Unified Inheritance Physics (UIP) — a separate academic effort with its own repository and DOI.
See PROVENANCE.md for the relocation history and the relationship between the two tracks.
| ID | Title | Verdict | Backend | Key result |
|---|---|---|---|---|
| ARK-441 | VBE authorization boundary (baseline) | ✅ PASS | ibm_marrakesh |
Binary ALLOW/DENY boundary established (Δ_B ≥ 0.70) |
| ARK-446 | Cross-device replication | ✅ PASS | ibm_marrakesh |
L_D_corrected=0.0020, S_A=0.9885, Δ_B=0.9862 |
| ARK-442 | Delay-boundary characterization | ✅ PASS | ibm_marrakesh |
L_expired=0.0000, S_reverified=0.9916, Δ_B=0.9877 |
| ARK-444 | Decision-to-execution integrity | ✅ PASS | ibm_marrakesh |
Fail-closed on every alteration/replay, Δ_B=0.9696 |
| ARK-443 | Two-of-three quorum authorization | ✅ PASS | ibm_marrakesh |
Quorum enforced; single compromised channel denied, Δ_B=0.9685 |
All five carry release tags ark-4NN-v1.0 and are published to Zenodo (DOI 10.5281/zenodo.21398676).
A non-quantum, production-like enforcement series lives in
ark-493-498/. Unlike the hardware ARKs above, the boundary here
is a real software enforcement point (and, in ARK-498, a real HTTP/loopback-TCP
socket). It tests whether ExecutionProof mechanically enforces an authorization
decision at a real boundary — not merely records it.
| Experiment | Focus | Result |
|---|---|---|
| ARK-493 | Enforcement boundary under adversarial load (6 paths × 15 cells) | ✅ PASS — 90/90 |
| ARK-494 | Semantic boundary: deep argument mutation | ✅ PASS — 13/13 |
| ARK-495 | Temporal boundary: authority change mid-flight | ✅ PASS — 11/11 |
| ARK-496 | Multi-agent delegation & self-approval defense | ✅ PASS — 8/8 |
| ARK-497 | Independently reconstructable ProofRecord (+ tamper detection) | ✅ PASS — 30/30 |
| ARK-498 | Networked production-like performance | ✅ PASS — 6/6 hard criteria |
161 scored cases, 161 PASS, 0 enforcement leaks, GATE-STOP not triggered, dual-guard
agreement on every record. Frozen preregistration v1.1 (SHA-256
464b9fb8be9d6cca052f236dc9deec9f8e89b781cafc58701e79b2d05d52952a). See
ark-493-498/README.md for the full breakdown and the
mandatory ARK-498 performance-characterization caveat.
72 preregistered experiments · 232 case records · 229 PASS · 2 FAIL · 1 GATE-STOP · 9 repositories. The 2 FAIL and 1 GATE-STOP records are retained unchanged — preserving negative results is part of the methodology.
- Preregistration-first. A 28-field preregistration + all code + a SHA-256
MANIFEST.txtare committed before any hardware job is submitted. The lock commit hash is recorded in each experiment'sRUN_LOG.md. - Hard ordering (Field 27). Qubit selection → SPAM kill-gate → job ID committed before results are read → raw counts → analysis → results. Commit timestamps prove the ordering.
- SPAM kill-gate. An in-situ state-preparation-and-measurement baseline is measured and committed before the principal job; if it exceeds the preregistered threshold, the principal job is not submitted.
- No rescue after failure. A failed hypothesis is reported as a failure. No post-data threshold changes, arm redefinitions, or re-analysis may change a verdict.
- Raw counts, no mitigation. Metrics are computed from raw device counts; SPAM correction is applied only where preregistered and always reported alongside the raw value.
Proof Before Power. Prediction Before Measurement. No Rescue After Failure.
Every experiment can be re-run on a free IBM Quantum open-plan account. Requirements are in
requirements.txt; per-experiment README.md files describe the specific setup.
Code and data released under the license in LICENSE. Dataset also archived under CC BY 4.0 on Zenodo.