Lebrel Wants Readers to Check Which AI Model Actually Answered Them Lebrel, led by founder and CEO José Ignacio Gisbert Valls, has launched an uncensored AI chat on encrypted, attested hardware that attaches a signed receipt, called Proof of Edition, to every answer so users can verify which model produced the text. Each receipt names the edition, base model, precision and engine, carries hashes of the request and answer, and is registered on Solana, with verification running in the user's browser rather than on Lebrel's servers. Lebrel's public watch board rates DeepSeek V4 Pro Consistent at 98.33% matching first words, Qwen3.8 Flash and GLM-5.3 Flash at 100%, and marks GLM-5.3 Divergent at 91.84%, citing a September 2026 Anthropic report on labs passing off Claude's answers as their own and resellers silently swapping models. A model name on an API is a claim. That line sits on the Lebrel website, and it is the premise behind the company's product. Lebrel offers an uncensored AI chat that runs on encrypted, attested hardware, and every answer comes with a signed receipt. A Verified seal under the answer opens that receipt and shows which model produced the text. Lebrel is led by founder and CEO José Ignacio Gisbert Valls, and it has recently added image generation alongside the chat. Gisbert Valls invites readers to bring a question their usual AI wouldn't answer, or an image they've had in mind, and try Lebrel's private, uncensored AI chat https://lebrel.ai for themselves. Once they have asked their own question, clicking Verified opens the receipt and shows which model replied, and no account is needed to try it. What the receipt shows The idea is called Proof of Edition, which Lebrel describes as an open standard for verifiable inference. Each answer is signed with the exact model that produced it, and the signature is registered on Solana. The receipt names the edition, the base model, the precision and the engine behind the answer, along with hashes of the request and the answer. The checks run in the user's browser, not on Lebrel's servers. The company says only the attested runtime can sign an answer, not Lebrel and not the cloud provider. Study Says Gemini and Grok Share a Blind Spot That Could Undermine AI Safety https://startupfortune.com/study-says-gemini-and-grok-share-a-blind-spot-that-could-undermine-ai-safety/ Artificial Epistemics says Gemini and Grok both rely on a justificationist approach to knowledge that leaves them exposed to misinformation and rogue behavior, and pitches its falsificationist Susty Code as a fix operating at the epistemological layer. - why Gemini and Grok have knowledge blind spots https://startupfortune.com/study-says-gemini-and-grok-share-a-blind-spot-that-could-undermine-ai-safety/ - AI safety vulnerabilities in large language models explained https://startupfortune.com/study-says-gemini-and-grok-share-a-blind-spot-that-could-undermine-ai-safety/ Lebrel's site shows a sample receipt to make this concrete. It records a test question, Reply with exactly: RECEIPT OK , and an answer of five tokens. The receipt was signed on 17 October 2026 by a test runtime without a GPU and registered on Solana devnet. It is a small example, but it shows what a user gets: a record they can inspect themselves, instead of a label they have to believe. Why the model name needs proof Lebrel frames the product as a response to a trust problem. On its site, the company points to a September 2026 report from Anthropic describing labs passing off Claude's answers as their own, and resellers silently swapping the model behind the name they sell. For anyone building on an AI service, that matters. If the model behind an API can change without notice, results, costs and behaviour can change with it. Lebrel's answer is to publish what each lab actually serves, and to tie every edition's output to a pinned checkpoint. A model ID maps to one checkpoint, one precision and one engine. Changes ship as a new version, never as a silent swap. A board that measures labs Alongside the receipts, Lebrel runs a public watch board for the models it routes. The company describes four steps. A fixed battery of prompts is the same for every lab and built to catch hidden prompts, refusals and swaps. Each run is compared against the previous one and against Lebrel's own run of the published weights. A model's fingerprint is how it weighs its first word, so a swapped model shows up quickly. Every verdict is published with its numbers and thresholds. The board does not show a clean sweep. DeepSeek V4 Pro is marked Consistent, with 98.33% matching first words against the lab's API. Qwen3.8 Flash and GLM-5.3 Flash both show 100%. GLM-5.3, however, is marked Divergent, at 91.84%, meaning it did not match the first words of the lab's API. Qwen4 Max is listed as not yet probed. That willingness to publish a negative verdict is what separates a measurement from a marketing page. Privacy as part of the design Verification is only half of the pitch. Lebrel encrypts requests from the user's device to a key generated inside attested hardware, using HPKE with X25519 and AES-256-GCM. The hardware is Intel TDX with a confidential GPU, and the client checks the attestation before it trusts the key. Anthropic Is Hiring a Chemical Weapons Expert and the Internet Lost Its Mind https://startupfortune.com/anthropic-is-hiring-a-chemical-weapons-expert-and-the-internet-lost-its-mind/ A LinkedIn job listing went viral this week. Anthropic, the company behind the AI platform Claude, posted an opening for a Policy Manager specialising in chemical weapons and high yield explosives. The salary range was between 245 and 285 thousand dollars a year. Hybrid role. Based in New York. The screenshot made the rounds on … - Anthropic hiring chemical weapons policy manager role https://startupfortune.com/anthropic-is-hiring-a-chemical-weapons-expert-and-the-internet-lost-its-mind/ - AI company chemical weapons expert job listing controversy https://startupfortune.com/anthropic-is-hiring-a-chemical-weapons-expert-and-the-internet-lost-its-mind/ The company says its edge relays ciphertext they cannot read, and that it never stores prompt or output contents as server-side history. Only billing metadata remains, and chat history lives on the user's device. Web search, where used, receives only the query and never the conversation. The product also adds no refusal layer of its own above the published model. Lebrel positions this for work that trips consumer filters, including security research, trust and safety, investigations and agents. Open beta Lebrel is in open beta. The uncensored edition is open for four hours a day, from 16:00 to 20:00 Madrid time. Paid usage earns points toward the LBRL airdrop. Editions, routes and pricing Lebrel serves two kinds of models. Its own uncensored editions are built on open-weight models, and each edition's weights are published with the hash of each file. At the time of writing, the DeepSeek V4.1 Flash Lebrel Uncensored edition is online, with a native context of 1M tokens, priced at $0.13 per million input tokens and $0.53 per million output tokens. The GLM-5.3 Flash and Qwen 3.8 27B Lebrel Uncensored editions are listed as paused. The second kind is routed models from other labs, served through Lebrel's attested router and anonymised. The provider sees the request but never who asked, and a signed route receipt comes with every answer. Lebrel notes that encryption applies to its own editions. The routed list includes DeepSeek, Kimi, Qwen and GLM models. For developers, the API keeps the OpenAI request format, so existing tools can connect with a base URL, a key and a model name. Lebrel lists Hermes, Kilo Code, Claude Code, Codex, OpenCode and ZCode among the coding agents that can use it. The Proof of Edition code is published on GitHub, with packages for Python and npm. What comes next Publishing the standard openly, instead of keeping verification inside its own product, is the part of Lebrel's approach that stands out. Anyone can check an answer without trusting the company that served it, and anyone can build the same check into their own tools. Whether other providers adopt the standard, or answer to the watch board, is the open question. Join the discussion Open in the community → https://startupfortune.com/community/ Almost there. Sign in and your reply posts straight away.