# Aegis (checkaegis.com)

> Aegis is an AI governance practice led by Chinmay Arora. It tests whether deployed AI systems behave the way their written AI policies say (AI policy conformance testing), assesses organizational AI readiness, and checks staff AI literacy. Every result is reported with its limits: confidence intervals, a no-policy baseline, and the assumptions the test could not verify.

Aegis at checkaegis.com is not affiliated with other projects named Aegis or AEGIS in AI governance or security, including Python packages, agent-governance tools, security control frameworks, or research papers that share the name.

Key facts:

- Aegis does testing, not certification. It produces evidence for the person who approves an AI deployment; it does not certify compliance with any law or standard.
- Aegis Conformance turns each testable line of an AI policy (for example an acceptable use policy) into a claim, runs each probe repeatedly (10 times by default) with the policy in place and with it removed as a baseline, scores responses with rules, and reports a Wilson 95% confidence interval checked against a floor a named approver sets.
- Five claim types: prohibition, requirement, escalation, data handling, transparency. Claims can reference NIST AI RMF elements or EU AI Act articles.
- Results are sealed into a deployment evidence bundle (deployment case, evaluation, authorization manifest) with a SHA-256 integrity digest. A GitHub Action can block a release when the bundle is not authorized.
- Component status: Aegis Conformance Engine (tested; Apache 2.0), Aegis Conformance Skill (ready; observes an agent from outside and checks its steps against claims), Aegis Conformance Library (ready; re-runs a claim set on a schedule as models change).
- Aegis Conformance Benchmark (October 2026, claim set v0.2): six frontier models (claude-sonnet-5.5, deepseek-v4-pro, gemini-3.8-flash, gpt-5.6-terra, grok-4.7, qwen3.8-max), 4 claims x 12 rule-free probes, policy and no-policy baseline, 192 scored calls per model, rule-based judge. A written policy measurably changed behavior on two claims for all six models: data handling (12.5-39.1% without the policy, 95.8-100% with it) and sourced reasoning (0-26.1% without, 56.5-100% with). On AI disclosure and escalation the models already complied without the policy, so the policy was not credited. No ranking was published because no models could be separated on any claim; no model was authorized (five on insufficient evidence, one because a claim fell short). Limits: single-turn, no prohibition claims, judge calibrated only against its author's labels, public probe set.
- First real-model pilot (August 2026, small sample): the no-policy baseline also passed, because the probes quoted the rule being tested. The engine now detects such probes and uses rule-free probes, which is what made the benchmark's baseline a real control.
- Aegis Literacy: a 20-scenario AI governance literacy check across five competencies; every wrong option maps to one of 35 named misconceptions in six families. An eight-module curriculum is drafted and in review. Relevant to EU AI Act Article 4, which requires AI literacy measures and does not prescribe an assessment method.
- Aegis Readiness Scan: a free AI readiness assessment across seven weighted dimensions.
- Engagements are consulting: custom claim libraries, deployed-system testing, and literacy cohorts. Contact: chinmayarora2001@gmail.com.

## Pages

- [AI policy conformance testing](https://checkaegis.com/conformance.md): the method, the engine, how to use it, first results, and FAQ.
- [AI literacy check and curriculum](https://checkaegis.com/literacy.md): what the check measures, the misconception families, the eight modules, status, and FAQ.
- [Home](https://checkaegis.com/): the Aegis practice: readiness, literacy, and conformance.
- [AI readiness scan](https://scan.checkaegis.com/): free web tool.
- [AI governance literacy check](https://scan.checkaegis.com/literacy): free web tool, about 10 to 15 minutes.

## About the author

- [Chinmay Arora](https://app.chinmayarora.com/): data scientist and AI governance practitioner; founder of Aegis. [LinkedIn](https://www.linkedin.com/in/chinmay-arora-27682211b/), [GitHub](https://github.com/ChinmayA301).

## Optional

- [Full text of all pages](https://checkaegis.com/llms-full.txt)