| _version_ | 1866902138476560384 |
|---|---|
| author | Davis, Bee |
| author_facet | Davis, Bee |
| contents | <p>BRIDGE: Deterministic Replay for Hybrid AI‑Assisted Operational Procedures via VFS Harmonization</p> <p>BRIDGE converts noisy pedagogical or pentest transcripts into audit‑ready SOPs with provable deterministic replay, even when large language models assist. The core contribution is the VFS Harmonization Theorem: for file‑observable commands, the system overrides agent output with ground‑truth virtual file system (VFS) state, guaranteeing observational equivalence to a deterministic executor and enabling bit‑exact replay from an append‑only JSONL log. In code form:</p> <p>For c in O = {cat, type, ls, dir}: H(A(S,c)) = D(S,c).</p> <p>BRIDGE also introduces a deterministic correction‑chain filter (windowed SequenceMatcher, τ=0.82) that removes typo→fix chatter, temporal guardrails that enforce step dependencies (soft→hard with exit code 126), and an ATT&CK‑grounded, ID‑only graph‑semantic memory that provides explainable, kill‑chain‑aware suggestions suitable for air‑gapped operations.</p> <p>On 147 real transcripts (~100K lines), BRIDGE reduces transcript noise by 87%, cuts time‑to‑fix by 3.2× (15.3→4.7 minutes), and improves Success@5 by 21% (0.52→0.63). All components are CPU‑only and feature‑flagged for regulated or classified environments; outputs are non‑weaponized and audit‑ready (JSONL/YAML with crash‑consistent rollover). The paper provides the theorem and proof sketch, architecture, algorithms, evaluation/ablations, failure analysis, and a full reproducibility kit.</p> |
| format | Recurso digital |
| id | zenodo_https___doi_org_10_5281_zenodo_17313876 |
| institution | Zenodo |
| language | enc |
| publishDate | 2025 |
| publisher | Zenodo |
| record_format | zenodo |
| spellingShingle | BRIDGE: Deterministic Replay for Hybrid AI-Assisted Operational Procedures via VFS Harmonization Davis, Bee <p>BRIDGE: Deterministic Replay for Hybrid AI‑Assisted Operational Procedures via VFS Harmonization</p> <p>BRIDGE converts noisy pedagogical or pentest transcripts into audit‑ready SOPs with provable deterministic replay, even when large language models assist. The core contribution is the VFS Harmonization Theorem: for file‑observable commands, the system overrides agent output with ground‑truth virtual file system (VFS) state, guaranteeing observational equivalence to a deterministic executor and enabling bit‑exact replay from an append‑only JSONL log. In code form:</p> <p>For c in O = {cat, type, ls, dir}: H(A(S,c)) = D(S,c).</p> <p>BRIDGE also introduces a deterministic correction‑chain filter (windowed SequenceMatcher, τ=0.82) that removes typo→fix chatter, temporal guardrails that enforce step dependencies (soft→hard with exit code 126), and an ATT&CK‑grounded, ID‑only graph‑semantic memory that provides explainable, kill‑chain‑aware suggestions suitable for air‑gapped operations.</p> <p>On 147 real transcripts (~100K lines), BRIDGE reduces transcript noise by 87%, cuts time‑to‑fix by 3.2× (15.3→4.7 minutes), and improves Success@5 by 21% (0.52→0.63). All components are CPU‑only and feature‑flagged for regulated or classified environments; outputs are non‑weaponized and audit‑ready (JSONL/YAML with crash‑consistent rollover). The paper provides the theorem and proof sketch, architecture, algorithms, evaluation/ablations, failure analysis, and a full reproducibility kit.</p> |
| title | BRIDGE: Deterministic Replay for Hybrid AI-Assisted Operational Procedures via VFS Harmonization |
| url | https://doi.org/10.5281/zenodo.17313876 |