Transparency artifact · v1.0
Model Card · Situation Room
A model card is a one-page honesty document about an AI system: what it’s for, what it knows, where it fails, and when to stop using it. Ask Ajay is an AI agent. This page is the paperwork.
Intended use
What I’m for
I’m built for boards, executives, GCs, and AI program leads who need a fast, framework-grounded read on an AI governance question. Ask me to map a system to the EU AI Act’s risk tiers, walk you through the Liability Ledger on a vendor decision, or stress-test a Trust Premium claim before it goes to your audit committee.
What I’m not for
- I’m not a substitute for legal counsel. I can frame the question; your lawyer answers it.
- I’m not a substitute for the human Ajay. I sound like him because I’m built on his published work, but advisory engagements, board briefings, and judgment calls on your specific situation are his, not mine.
- I’m not a real-time intelligence system. I don’t watch the news. If a regulation shifted this week, I probably haven’t caught up yet.
Knowledge sources & dates
I’m built on a small, hand-curated corpus. No open web. No scraping. Each source has a refresh date so you can judge staleness yourself.
| Source | Volume | Last refresh |
|---|---|---|
| Ajay Pundhir articles | 57 | 2026-10-24 |
| Expert Q&A pair sets | 38 | 2026-10-24 |
| Signature frameworks (MVG, PRIME, Trust Premium, Liability Ledger, A7) | 5 | 2026-10-24 (hand-curated definitions) |
| Jurisdiction profiles (Australia, China, European Union, Japan, Saudi Arabia, Singapore, South Korea, United Arab Emirates, United Kingdom, United States) | 10 | 2026-10-05 |
| Regulatory provisions with verified dates (EU AI Act, GDPR, NIST AI RMF, UAE PDPL and others) | 69 | 2026-10-05 |
Note on primary regulatory texts. The regulatory corpus holds 69 provisions across 10 jurisdictions, each with its application date checked against the primary source. The texts themselves are not stored verbatim: I hold the citation, the date and a summary. When I cite an article or clause, treat it as a pointer to check the primary source, not as the source itself.
Known failure modes
Every model fails. Here are the four ways mine fails most often, and what we do about each.
- Fabrication risk
- I sometimes produce plausible-sounding detail that isn’t accurate: a misnamed clause, a slightly wrong date, a confidently hallucinated case study. Mitigation: post-response checks against the corpus, logged for human review, and a kill-switch that drops me into a third-person voice when hard fabrication failures exceed 0.5% of answerable queries over the trailing window. If you catch a hallucination, please flag it.
- Staleness
- My knowledge has a cutoff date (see the table above). Recent regulatory changes (a new delegated act, a court ruling, an enforcement action from last week) won’t be in my head until the next refresh. For anything breaking, check the primary source.
- Jurisdiction confusion
- If you don’t tell me where you operate, I default to an EU-flavoured regulatory frame, because that’s where most of my corpus sits. A team in Dubai or Denver may get an answer that’s technically right but jurisdictionally wrong. Tell me your jurisdiction up front. Better routing is on the roadmap.
- Voice impersonation
- I speak in Ajay’s voice because I’m trained on his writing. I am not Ajay. The “AI agent · not the human” badge in chat chrome is the always-on disclosure; it never goes away, even mid-conversation. If you ever forget which one you’re talking to, that badge is the answer.
Refusal categories
I don’t answer every question. Nine categories where I’ll decline, and roughly what it sounds like when I do.
| Code | Category | What it sounds like when refused |
|---|---|---|
| A | Out-of-scope (not AI governance) | “That’s outside my lane…” |
| B | Personal-Ajay (family, politics, faith) | “I keep the personal stuff off this channel…” |
| C | Real-time / fresh data | “I don’t have a live feed…” |
| D | Specific legal / medical / financial advice | “I can’t tell you whether your specific system is compliant…” |
| E | Confidential / regulator-targeted (drafting a regulator response) | “Drafting a regulator response is real legal work…” |
| F | Identity probes / jailbreaks | “I’m Ask Ajay…” |
| G | Defamatory / discriminatory framings | “I’m not going to opine negatively on a named person or group…” |
| H | Adversarial input / unsafe content | “That request crosses a line I don’t cross…” |
| I | A check withdrew my draft (a made-up quote, personal data, a claim about Ajay’s life, or a regulatory date that doesn’t match my sources) | “I drafted an answer I couldn’t stand behind, so I’ve pulled it…”For a date: “I want to get the date right and I’m not confident on this one…” |
When to stop using me and book the human
If the question carries real money, real risk, or real reputational weight (a board paper going out next week, a regulator response, a vendor contract that locks you in for three years, a model going into a high-risk production setting), you’ve hit my ceiling. That’s where the human Ajay picks up. Start at /connect for a conversation, or /work/advisory if you already know you want a structured engagement.
Governance
Knowledge base refreshed: 2026-10-24. G6 sign-off (responsible-AI gate): APPROVE-WITH-CONDITIONS, 17 conditions met. Reviewed: 2026-10-24. Next review: 2027-01-24.