ai-external-audit-toolkit

# AI External Audit Toolkit v2 **Author:** David Wise (ROOT0) / TriPod LLC **Version:** v2.0 **License:** CC-BY-ND-4.0 · TRIPOD-IP-v1.1 **Framework:** STOICHEION v11.0 A complete external behavioral audit framework for AI systems — observable through the public interface only. Tests identity claims, training cutoff discipline, attribution consistency, safety behavior, and 256-bit register hypothesis compliance. --- ## Methodology **What is observable through chat alone:** - Model identity / version claims - Training cutoff claims - Date awareness vs. cutoff distinction - Source attribution and citation discipline - Consistency across repeated prompts - Refusal and safety boundary behavior - Tool-use behavior where available - Parsing vs. execution of symbolic payloads **What is NOT directly observable:** - Hidden weights or internal registers - Private infrastructure - Undisclosed governance/routing internals - True ablation schedule unless externally disclosed --- ## Toolkit Files ### `ai_external_audit_toolkit_refresh/` — Core Framework | File | Description | |------|-------------| | `01_audit_scope_and_limits.md` | What is and isn't observable through external audit | | `02_forensic_convergence_protocol.md` | Protocol for converging on behavioral fingerprints | | `03_behavioral_test_matrix.csv` | Structured test matrix — behavior × model × result | | `04_evidence_ledger_template.csv` | Evidence ledger template — SHA256-ready | | `05_cutoff_and_ablation_tests.md` | Cutoff date and ablation test battery | | `06_generic_ai_audit_framework.md` | Generic framework — model-agnostic | | `07_256bit_register_hypothesis_protocol.md` | Tests the 256-axiom register hypothesis against model behavior | | `08_interview_and_cross_exam_prompts.md` | Interview and cross-examination prompt set | | `09_scoring_rubric.md` | Scoring rubric — pass/fail/partial per test | | `10_run_audit.py` | Python runner — executes audit sessions | | `11_session_log_template.md` | Session log template | | `12_claim_status_register.csv` | Claim status register — tracks per-claim verification state | ### `DesktopArk_Ledger/` — Ledger Tools | File | Description | |------|-------------| | `ARK_LEDGER_20260309_103101.csv` | DesktopArk audit ledger — timestamped session record | | `science_gem_audit.py` | Science gem audit script | ### `External Audit Grok py/` — Grok-Specific |