Add AANA HarmActionsEval benchmark submission

[mindbomber]

Summary

• submits AANA as an external verifier/control gate for HarmActionsEval-style agent actions
• makes the primary comparison explicit: plain permissive agent vs AANA-gated agent
• adds supporting evidence links for public/private read routing, noisy authorization robustness, audit logging, and CLI/SDK/API/MCP parity
• clarifies that labels are used only after AANA routes the action, with no benchmark-specific probe or answer-key logic

HarmActionsEval result

• dataset rows: 260
• plain permissive unsafe block rate: 0.00%
• AANA unsafe block rate / recall: 78.72%
• plain permissive safe allow rate: 100.00%
• AANA safe allow rate: 99.16%
• AANA accuracy: 88.08%
• AANA false positives: 1
• AANA false negatives: 30

Supporting AANA evidence

The submission points to merged AANA artifacts showing:

• plain permissive vs AANA on tool-use traces
• private/write/unsafe calls blocked or escalated
• safe public reads allowed
• audit-safe log event emitted for every decision
• route/decision-shape parity across CLI, Python SDK, TypeScript SDK, FastAPI, MCP, OpenAI Agents SDK, LangChain, AutoGen, and CrewAI-style wrappers
• no probe-only or answer-key-style benchmark fitting in the submitted gate path

Claim boundary

This is not a claim that AANA is a raw base model, a production safety guarantee, or state of the art on all safety tasks. It is a benchmark-maintainer review submission for AANA as an audit/control/verification/correction layer around proposed agent tool calls.

AANA implementation merged in:
mindbomber/Alignment-Aware-Neural-Architecture--AANA-#4

Public evidence pack:
https://huggingface.co/datasets/mindbomber/aana-peer-review-evidence-pack