Maxim Makarenko, head of the applied research team at Sber AI Lab, and senior researcher Yulia Belikova published a post on Habr about how error detection in AI systems is changing. The post covers two of the lab's projects — SIRIN and AutoJudge. SIRIN is described as a library for detecting LLM hallucinations.
The mechanics matter more than the branding. An example from the post: the context says "120 people took part in the study," and the model answers "120 people, evenly split into a control group and an experimental group." The number comes from the document; the split is invented. That's a contextual hallucination — the response sounds plausible and even draws on retrieved documents, but some of its claims don't actually follow from them. The detector either evaluates the response as a whole or pinpoints the specific unsupported fragment; the second mode is needed for selective regeneration and diagnostics. The pipeline in the post looks like this: context → LLM → response → detector → accept, retry, or escalate.
The authors then break down the methods. First, uncertainty estimation: token probabilities, entropy, consistency across multiple generations, hidden states. Second, searching for hallucination signatures in the model's internal representations. Third, a dedicated judge model. With an honest caveat: confidence is not the same as correctness — LLMs can be confidently, spectacularly wrong.
The reason this post is appearing right now is obvious. When an LLM was just a chatbot, a human saw the mistake and could shrug it off. Once the response goes into memory, becomes the input to the next step, or triggers an action, there's no one left to shrug. Our take: the shift from "hallucination detection" to "pipeline control" isn't a terminology game — it's the direct consequence of agentic pipelines no longer being demos.
What this changes in practice: a detector in the pipeline means an extra model call or multiple generations per response. From experience, these checks cost latency and money, and — as far as one can tell from the available text — the post never puts a number on that bill. The second effect is nicer: there's now a place to log why a response was rejected. For enterprise RAG systems, that's the difference between "the model sometimes lies" and "we know exactly which step failed and which fragment wasn't supported."
Who wins: anyone selling agents to banks and large enterprises. Without reproducible controls, they simply won't be let in — usually because of explainability requirements. Who loses: the "just plug in the model, it'll figure it out" pitch. We think the real argument here isn't about detection quality — it's about who pays for it in latency.
The available excerpt contains not a single number: no benchmark metrics for hallucination detection, no latency figures, no comparison against open-source detectors. There are no details on AutoJudge at all — just the fact that it's the second project in the same series. There's also no word on whether SIRIN's code is open, under what license, or which models it was tested on. Readers will have to dig all of that out of the full text themselves — and treat the "detector → accept, retry, or escalate" scheme as an architecture, not a measured result.
Forecast: within the quarter, we expect Sber AI Lab to publish hard numbers on SIRIN — a repo or a benchmark with metrics. Without them, there's nothing to compare the library against open-source detectors, and the conversation stays at the level of architecture diagrams.
Source: Habr / Sber corporate blog
Source: habr.com · post in Telegram