Deterministic, Zero-Custody Machine-Audit Reproducibility of Bitcoin-Anchored Digital Audit Trails: A Double-Blind, HMAC- Randomized Validation Study Across a 7,239-File Multi-Entity Evidence Corpus

Deterministic, Zero-Custody Machine-Audit Reproducibility of Bitcoin-Anchored Digital Audit Trails: A Double-Blind, HMACRandomized Validation Study Across a 7,239-File Multi-Entity Evidence Corpus Ordinal 14 Regulatory Submission Software (Audit Target): AuditLog.AI Date: December 02, 2025 Inventor and Primary Contact Dr. Fernando Telles, BMedSc(Adv), MD(Dist)¹² Position: CEO & Founder, CDA AI Email: [email protected] Phone: Provided on Request Address: 21 Shields St, Flemington VIC 3031, AU Web: www.aihumansynergy.org Engineers Lead Software Engineer: Dr. Jacob Yang, BEng, MEng, PhD¹ Software Engineer: Benjamin Hookey, BEng (Mechatronics & Robotics), FSEng¹ Independent Investigator Operator 2: Dr. Sam Francis MBBS, MTrauma, PhD, DPM, FFPM (Principal Investigator, Pharmaceutical Physician) Affiliations ¹ Cardiovascular Diagnostic Audit & AI Pty Ltd (ACN 638 019 431) – Registered Australian company conducting AuditLog.AI software development, audit and research services ² Telles Investments Pty Ltd (ACN 638 017 384) – Private IP holder IP Rights US Provisional #63/826,381 · AU Provisional #2025902482 · AU Trade Mark #2535745 & #2549093 IP Priority Date: 17 June 2025 (Global) All reproducibility methods and QMS scripts are proprietary components of Sentinel Protocol v4. Submitted as part of the AuditLog.AI Global Regulatory Submission Package (FDA/EMA/TGA/PCAOB/ISA Alignment) Sentinel QMS v4 | Reproducibility Study (Ordinal 14) | AuditLog.AI © Cardiovascular Diagnostic Audit & AI Pty Ltd | 2025-12-02 Page 1 of 12 QMSv40001ORD14 Deterministic, Zero-Custody Machine-Audit Reproducibility of Bitcoin-Anchored Digital Audit Trails: A Double-Blind, HMACRandomized Validation Study Across a 7,239-File Multi-Entity Evidence Corpus Ordinal 14 Abstract Importance Digital audit trails are vulnerable to content alteration, selective omission, metadata manipula‐ tion, and unverifiable reproduction. Most GxP systems rely on vendor‑controlled logs rather than cryptographic proof. A zero‑custody, Bitcoin‑anchored, deterministic machine‑audit framework that demonstrates proof‑of‑unchanged for regulatory‑grade evidence has not previously been shown. Objective To evaluate the reproducibility integrity of AuditLog.AI under independent, deterministic machine‑audit re‑execution using Sentinel QMS v4 by: 1) Reproducing all dual‑hash outputs (SHA‑256 and RIPEMD‑160(SHA‑256)) for 16 historical regulatory runs, and 2) Detecting both expert‑designed and HMAC‑randomized negative controls across single‑entity and multi‑entity blinded evidence streams. Design, Setting, and Participants Single‑site, dual‑operator deterministic machine audits were conducted under full‑screen capture with pre‑audit evidence freezing and Bitcoin anchoring of pre‑ and post‑audit states. All audits were executed on an independent QMS environment with newly installed tools using macOS at CDA‑AI Headquarters (Melbourne, Australia) between 18 and 01 December 2025. The primary corpus comprised 1,087 original files across 16 AuditLog.AI sessions (REGULAT‐ ORY‑RUN001–016), including FDA/EMA/TGA reference sets, session logs, and Ordinals 11–13 evidence bundles. For Stage IIIB, this was extended with a blinded multi‑entity corpus of 6,152 files (SENTINFRA, METAVAL, and corporate/trust evidence), yielding an aggregate of 21,717 files across 6 QMS sessions. Exposures (Interventions) Stage I — Amplify stress test (QMS RUN001–002). Deterministic re‑execution of frozen posit‐ ive‑control sets (Ordinals 11–13; regulatory references; session logs v4001–v4071) using Sentinel QMS utilities. All inputs were frozen and anchored as pre‑audit evidence (RUN001), then re‑processed and re‑anchored as audit outputs (RUN002). Sentinel QMS v4 | Reproducibility Study (Ordinal 14) | AuditLog.AI © Cardiovascular Diagnostic Audit & AI Pty Ltd | 2025-12-02 Page 2 of 12 QMSv40002ORD14 Stage II — Deterministic re‑audit of the complete AuditLog.AI corpus (QMS RUN003–004). QMS scripts ( qms_prepare_manifests_v4.py , qms_prepare_evidence_v4.py , qms_prefix_evidence_v4.py ) cloned manifests and frozen evidence into QMS run roots, pre‐ serving bytes and directory structure. The cloned corpus was re‑hashed with batch_dualhasher_multi_v3.3.py , compared to original manifests using qms_compare_manifests_v4.py , and both pre‑audit state (RUN003) and audit results (RUN004) were anchored to Bitcoin. Stage IIIA — Human‑designed adversarial tamper challenge (QMS RUN005–009). A quali‐ fied medical investigator (Operator 1) introduced 34 randomly selected negative controls across 16 regulatory runs, using pre‑specified challenges: micro‑edits to numerical results, document‑level changes (PDF/DOCX), file deletions, spurious insertions, file swaps, and metadata‑only edits; three filename‑only changes served as controls. All tampering was per‐ formed on an unprotected working copy under continuous full‑screen capture, with the original frozen corpus preserved. The QMS pipeline re‑hashed and compared tampered versus baseline manifests, anchoring the plan (RUN005), baseline reconstruction (RUN006), tampered corpus (RUN007), positive‑control Unprotected Storage Simulation (USS) re‑export (RUN008), and final QMS comparison (RUN009). Stage IIIB — Dual‑arm multi‑entity blind HMAC deletion challenge (QMS RUN010–016). Stage IIIB extended testing to a dual‑arm, dual‑operator, double‑blind cryptographic audit: Arm 1 (Multi‑Entity): Baseline AMP_REPRODUCIBILITY-RUN010_MULTI_BASELINE, HMAC‑deleted working corpus RUN012_MULTI_HMAC_WORKING_TAMPER, and posit‐ ive‑control USS re‑export RUN014_MULTI_PC_ReExport_USS. Arm 2 (AuditLog.AI): Baseline AMP_REPRODUCIBILITY-RUN011_AUDITLOGAI_BASELINE, HMAC‑deleted working corpus RUN013_AUDITLOGAI_HMAC_WORKING_TAMPER, and pos‐ itive‑control USS re‑export RUN015_AUDITLOGAI_PC_ReExport_USS. Evidence directories were first anonymized using a randomized QMS Bates‑renaming procedure (qms_bates_renamer_v4.1 ), producing content‑agnostic blinded identifiers to prevent targeted deletions and obscure evidence identities. A HMAC‑based randomization script (qms_hmac_salt_blinding_v4.py ; k_global = 20 , max_frac_per_run = 0.10 ) then generated a secret SALT and applied deletions to “WORKING_TAMPER” runs only, moving selected files into _HMAC_DELETED/ while leaving baseline QMS runs intact. Operator 1 and a system‑naive Operator 2 alternated execution across baselines, HMAC tamper runs, and positive controls. Both operators were blinded to the HMAC selection pat‐ tern and to file identities beyond Bates‑renamed labels. All six runs were summarized in the six‑way QMS comparison aggregator RUN016_AMPLIFY_QMS_AUDIT. All steps were per‐ formed under full‑screen recording with audio narration and anchored session logs on an independent QMS environment. Main Outcomes and Measures Primary outcomes were: 1. Dual‑hash parity, defined as exact SHA‑256 and RIPEMD‑160(SHA‑256) concordance Sentinel QMS v4 | Reproducibility Study (Ordinal 14) | AuditLog.AI © Cardiovascular Diagnostic Audit & AI Pty Ltd | 2025-12-02 Page 3 of 12 QMSv40003ORD14 between frozen originals and re‑executed outputs. 2. Negative‑control detection, defined as correct identification of all deliberately altered or deleted files with no false positives on untouched files. 3. OP_RETURN payload and Bitcoin anchoring parity, defined by concord‐ ant SENTINEL|TYPE|ripemd160|sha8 payloads, TXIDs, and block heights between user‑side and AMPLIFY ledgers. 4. Human verification audit time, constrained by pre‑registered checklists and Bitcoin‑anchored metadata. Results Stage I (Amplify stress test). A total of 12,945 files (59 GB) were processed, with a PRE→POST dual‑hash execution interval of 58 seconds and 100.0% dual‑hash parity between original and re‑executed outputs. RUN‐ 001 and RUN002 were successfully anchored on Bitcoin (TXIDs 3e9a4db08c6badf62eff‐ b497b6ada36f5b27b30c5a3f954fd96073d7714792c4 and 9c2eb239046535b65ff14f0fea7dde5b‐ c7899d913ae9e24ad33069961305d156 ; blocks 923818 and 923939). Stage II (Deterministic re‑audit of the complete corpus). Across 16 regulatory runs, 1,087/1,087 files (100.0%) were reproduced with matching SHA‑256 and RIPEMD‑160(SHA‑256) hashes ( rows_unmatched_total = 0 ). RUN003 and RUN004 were anchored with OP_RETURN payloads conformant to SENTINEL v4 schema (TXIDs 9721d68449ce57b9ba0d25ff86bc50d57eeac5c132156ff922026e47b66af8f7 and 4873e651228ed6374801b52d1c5723b7bf95b7ef66c330e6d5e0772b7c793e9e ; blocks 924058 and 924061). Stage IIIA (Human adversarial tamper challlenge). Across all tamper categories (content edits, deletions, insertions, swaps, metadata tampering), 34/34 negatively control files (100.0%) were detected as mismatches. All 1,053/1,053 un‐ touched files and three filename‑only controls retained full dual‑hash parity (0 false positives). RUN005–RUN009 were successfully anchored (TXIDs d1d78c40ff7a8907b022cbcc1d5005bf0b‐ d889622f9620169438326edd007682 – 9de8e1d1ac4313e1dbb43afaff205d99beaf6a38e20c0bb83371ccf9a3111d73 ; blocks 924781 – 924793). Stage IIIB (dual‑arm, HMAC‑randomized machine‑audit challenge). In Stage IIIB we extended the Sentinel QMS v4 pipeline to a double‑blind, dual‑arm HMAC dele‐ tion challenge spanning six QMS runs (RUN010–RUN015). A frozen Multi‑Entity corpus (6,152 evidence files) and the AuditLog.AI regulatory corpus (1,087 files) were each (i) re‑hashed as blinded baselines, (ii) cloned to external User Simulation Storage (USS) as positive controls, and (iii) cloned again to working copies in which a secret HMAC SALT randomly deleted 20 Bates‑labelled files per arm (k=20; max‑fraction‑per‑run 0.10). Across all six Stage IIIB runs this produced 21,717 evidence‑file instances (7,239 unique files × 3 passes). Sentinel QMS v4 de‐ tected 40/40 (100.0%) HMAC‑selected deletions at the dual‑hash level (SHA‑256 and RIPEMD‑160(SHA‑256)) with 0/21,677 false‑positive mismatches. Human verification, con‐ strained by pre‑registered checklists and Bitcoin‑anchored metadata, required a total of 5,258 seconds (≈87.6 minutes) to exhaustively verify all Stage IIIB outputs, corresponding to a mean 0.24 seconds of human time per evidence file (0.62 s/file in tamper runs, 0.05 s/file in Sentinel QMS v4 | Reproducibility Study (Ordinal 14) | AuditLog.AI © Cardiovascular Diagnostic Audit & AI Pty Ltd | 2025-12-02 Page 4 of 12 QMSv40004ORD14 baseline + positive‑control runs). The six‑way QMS aggregator (RUN016) reproducibly summar‐ ized this pattern on an independent QMS environment, and all Stage IIIB runs were anchored on Bitcoin (TXIDs 5b0fdf19a0c2a9813079a1b73c30a4fda4c0fda85844a7de6d6398fda254dce0 – 0915251fd7130ebd4f568574b9f6c4824e530dc74e9413c90a20a96e15313262 ; blocks 925776 – 925812). Conclusions and Relevance Across four Stages (I–IIIB), AuditLog.AI and Sentinel QMS v4 reproducibly re‑executed frozen regulatory evidence under adversarial and cross‑platform constraints, with 100% dual‑hash parity for all PASS baselines and positive controls, 100% sensitivity to HMAC‑randomized deletions, and no false‑positive digest mismatches across 21,717 Stage IIIB evidence com‐ parisons. The combination of (i) deterministic dual‑hash manifests, (ii) blinded Bates identifiers, (iii) explicit HMAC‑driven non‑compliance challenges, and (iv) tightly logged Human Verification Time (HVT) per evidence file demonstrates that modern digital audit trails can support CRO‑grade reproducibility and machine‑auditability without sacrificing human interpretability. These findings suggest a practical template for regulators and sponsors seeking verifiable evidence integrity, reproducible AI‑assisted auditing, and auditable human‑in‑the‑loop QA at scale. Sentinel QMS v4 | Reproducibility Study (Ordinal 14) | AuditLog.AI © Cardiovascular Diagnostic Audit & AI Pty Ltd | 2025-12-02 Page 5 of 12 QMSv40005ORD14 Stage IIIB — Human Verification Time (HVT) Per-run HVT is expressed as seconds of human verification per evidence file (Δ / N). Baseline and positive-control runs are expected to pass; tamper runs are expected to fail with exactly 20 HMAC-selected deletions per arm. RUN ID Auditor HV Start (UTC) HV End (UTC) Δ (s) N (files) HVT (s/file) RUN010 (MULTI BASELINE) Dr. Fernando Telles 2025-11-30T02:30:1 3Z 2025-11-30T02:33:0 9Z 175 6,152 0.03 RUN011 (AUDIT‐ LOGAI BASELINE) Dr. Sam Francis 2025-11-30T02:35:0 9Z 2025-11-30T02:38:4 5Z 216 1,087 0.20 RUN014 (MULTI PC USS) Dr. Sam Francis 2025-11-30T05:14:3 8Z 2025-11-30T05:17:2 8Z 170 6,152 0.03 RUN015 (AUDIT‐ LOGAI PC USS) Dr. Fernando Telles 2025-11-30T04:14:2 6Z 2025-11-30T04:18:0 6Z 220 1,087 0.20 Baseline + PC Total 781 14,478 0.05 RUN012 (MULTI HMAC TAMPER) Dr. Fernando Telles 2025-11-30T02:43:4 0Z 2025-11-30T03:29:1 6Z 2,736 6,152 0.44 RUN013 (AUDIT‐ LOGAI HMAC TAMPER) Dr. Sam Francis 2025-11-30T05:20:0 0Z 2025-11-30T05:49:0 1Z 1,741 1,087 1.60 Tamper Chal‐ lenge Total 4,477 7,239 0.62 Combined Total 5,258 21,717 0.24 Sentinel QMS v4 | Reproducibility Study (Ordinal 14) | AuditLog.AI © Cardiovascular Diagnostic Audit & AI Pty Ltd | 2025-12-02 Page 6 of 12 QMSv40006ORD14 Page 7 of 12 QMSv40007ORD14 Cental Figure. Stage IIIB Double-Blind, HMAC-Randomized Machine-Audit and Human-Verification Workflow. This figure summarizes the Sentinel QMS v4 Stage IIIB protocol evaluating the reproducibility of the AuditLog.AI and Multi-Entity evidence corpora under a dual-arm, dual-operator, double-blind HMAC deletion challenge. Panel 1 shows preparation of a clean QMS environment: frozen, Bitcoin-anchored regulatory evidence is cloned into an isolated workspace and blinded using randomized Bates identifiers. Panel 2 depicts branching of each baseline run into a USS positive-control copy and an HMACdeleted working copy, with operators alternating arms while remaining blinded to identities and deletion patterns. Panel 3 shows the deterministic machine-audit aggregator (RUN016), which compares all six Stage IIIB runs using identical QMS comparison logic, yielding dual-hash parity for baselines and positive controls, and detecting HMAC-selected deletions only in working-tamper runs. A downstream Human Verification Layer (HVT) confirms each QMS-identified mismatch using timestamped START/END audit metadata; this layer validates but does not modify machineaudit results. Sentinel QMS v4 | Reproducibility Study (Ordinal 14) | AuditLog.AI © Cardiovascular Diagnostic Audit & AI Pty Ltd | 2025-12-02 Page 8 of 12 QMSv40008ORD14 Sentinel QMS v4 | Reproducibility Study (Ordinal 14) | AuditLog.AI © Cardiovascular Diagnostic Audit & AI Pty Ltd | 2025-12-02 Stage IIIB Human-Verification Artifacts and Dual-Hash Digests This table lists the human-verified QMSv4 human-verification (HV) PDFs for each Stage IIIB run and their associated dual-hash digests. For each QMSv4 RUN ID, the table records: • the auditor who performed the HV step, and • the SHA-256 and RIPEMD-160(SHA-256) digests of the final, human-annotated HV PDF. These PDFs contain the signed pre_HV_*.md tables, initials for each verified mismatch, and final signatures/time-stamps for HV START and HV END. The same digests appear in the corresponding HV metadata JSON files (HV_METADATA_QMS_COMPARE_*.json) and in the QMSv4 session logs, allowing any third party to: 1. recompute the digests from the archived PDFs and confirm bit-level integrity; and 2. re-derive the Human Verification Time (HVT) per run from the recorded START/END times. This table therefore provides the cryptographic evidence for the HVT metrics reported in Stage IIIB (0.62 s/file for tamper runs, 0.05 s/file for baselines and positive controls, 0.24 s/file overall). QMSv4 RUN ID Auditor Human-Verified PDF SHA-256 Human-Verified PDF RIPEMD-160 RUN010_MULTI_BASELINE Dr Fernando Telles eb67cafa27277b45d1e84692dbe2dfcc3a977ab223113e10baae95c8d815baa5 1de0e8bb30ad3bfe8139e46edf20b5336637bd1c RUN011_AUDITLOGAI_BASELINE Dr Sam Francis 7dc86eba2e283c56f5ed1a5a06039918fcdb1b8759132e1743bd89ea4ea0c0f8 5e72d5d8228afab4cbd8582d18286cab2b0cd55a RUN012_MULTI_HMAC_WORKING_TAMPER Dr Fernando Telles 5ce6f2fad142ce844413702171458d3745d158e4c9274d5867af07bb1bb6cc2f 964db5e7151a540a7811c1c825863ffa3d947734 RUN013_AUDITLOGAI_HMAC_WORKING_TAMPER Dr Sam Francis 8bab2e7ccfdb4ff7c86e6aba4fbfddba5c4714357d4971b93f8f3183afd209fc e99ec1af9464a5175fd44b88d8ffafdcced092f7 RUN014_MULTI_PC_ReExport_USS Dr Sam Francis cf60696770e5f9d774f48c2b007281adf43f6535cdc5e8b56c2daae38b338df6 4a88f82e5089e344bc78260740f18163bf943028 RUN015_AUDITLOGAI_PC_ReExport_USS Dr Fernando Telles 4d0d30e19c674fc1684f7a2f1be8e151183d91f18302802d7b56de549cd4f7e0 39a4c5446764ab7eb9d4a86ba1493a670ab9f1a2 Page 9 of 12 QMSv40009ORD14