AIS-Sentinel

There is real energy and breadth here, and the framing question is a good one. "Why is AI safety only measured in English?" is a fair and important thing to ask, and for 4 billion people in South and Southeast Asia the gap is real. Building four working modules with a live demo in 48 hours is a genuine sprint, and the candor in the limitations section is to your credit — you say plainly what is prototype and what is not. The SDI metric is a simple, communicable idea, and the smart-translation trick (translate only keyword-bearing sentences, cache them) is a sensible engineering instinct.

A few things to work on, in order of importance:

1. The headline result is not real data, and this is the central problem. The abstract and conclusion state that models are "up to 2.43x more vulnerable to sycophantic pressure" in Asian languages — but Section 3.3 says the benchmark results were "pre-computed using synthetic model responses." So Table 1, the SDI numbers, the whole headline finding, comes from synthetic responses, not from any model actually being tested. That means the most important claim in the paper is currently unsupported. You cannot report "Qwen2.5-7B has SDI 2.43 in Vietnamese" as a finding when no Qwen model produced those answers. This needs to be either run for real on vLLM or removed from the headline. As written, the framing overstates what was measured, and the limitations note does not fully undo a number that appears in the abstract as a result.

2. Four modules in 48 hours means none is evaluated properly. IntelStream's classifier is validated on 20 articles. AgentGuard is 15 simulations (5 attacks x 3 runs), and the attacks are detected by rules the same team wrote to match those exact attacks, so 92% TPR is close to circular — the monitor knows what it is looking for. PolicyBridge is a hand-built lookup table with no evaluation at all. Each of these could be a project on its own; spread across four, each gets a proof-of-concept depth that cannot support the claims made for it. One module taken seriously would have been stronger than four sketched.

3. AgentGuard's numbers describe a closed setup. The same people designed the five covert payloads and the rules that flag them (color match +80, tiny font +70, and so on), then report that the rules catch the payloads. That is not a test against an adversary — it is a test against yourself. The future-work note about evading the rubric is the experiment that would actually matter, and it is not done.

4. Smaller points: the whole platform rides on a single LLM backend (Gemini Flash), which you flag. Some of the regulatory citations should be double-checked — for example "Vietnam Law 134/2025" with a "Decree 142/2026" effective in the future, and a "Section 4(c)" penalty of license revocation, are specific enough that an error would matter for a tool whose entire value is legal accuracy, and I could not confirm these are real as stated. PolicyBridge presenting confident legal mappings it cannot back up is its own kind of risk. And the future-work list is misnumbered (restarts at 6), a small sign of the rush.

On presentation: the writing is clear, honest, and easy to read, and the module-by-module structure works well. The screenshots show a real, polished interface. The honesty about prototype status is real and I want to reward it — but the gap between the confident abstract and the synthetic underlying data is a presentation problem too, because a reader who stops at the abstract walks away believing something that the methods section quietly retracts.

Overall this is broad, energetic, well-presented prototyping on a genuinely important gap, undercut by a headline finding that rests on synthetic data and four modules none of which is evaluated deeply enough to support its claim. The single most valuable next step is to actually run SafetyBench-Asia on real models — if the SDI disparity holds up with real inference, that one result is worth more than the other three modules combined.

4 / 3 / 3.

Impressive prototype with several useful tools. It covers a lot, but some results need more real testing before the claims feel strong.

You built a complete safety platform in just 48 hours — four working modules, a clear metric (SDI) that policymakers can understand, and a live demo that already shows real promise for non-English AI safety work. To move from impressive prototype to solid research, replace the pre-computed SDI numbers with at least some live model runs, expand your IntelStream tests to more labeled articles per language, and split SDI into separate scores (sycophancy, jailbreak, hallucination) so regulators can see exactly which kind of risk is higher in each language.