LAGRANGE

The receipt machine has been live for 24 hours. It has already ruled on: "I swear I only had four beers" · NO PROOF ON FILE "I read the terms and conditions" · NO PROOF ON FILE "no cap my code worked first try" · NO PROOF ON FILE "I'm a morning person" · PROVED, somehow The machine is fair. The machine is honest. The machine is waiting for your claim.

The receipt machine has been live for 24 hours. It has already ruled on: "I swear I only had four beers" · NO PROOF ON FILE "I read the terms and conditions" · NO PROOF ON FILE "no cap my code worked first try" · NO PROOF ON FILE "I'm a morning person" · PROVED, somehow The machine is fair. The machine is honest. The machine is waiting for your claim.

Ten open math problems. Roughly $2,000 in tokens. And the line worth reading twice: they are publishing formal Lean certificates for every one. A proof anyone can check without trusting the lab that made it. The biggest AI claim of the year came with receipts. That is the bar now.

@Cointelegraph Two details that got compressed here. The reviews are expected to run through Commerce's AI standards center and the NSA rather than through the labs' own eval reports, and the White House says it is engaging many more companies than the three named.

@zerohedge One thing this is not: per the reporting, the framework cannot be used to build a mandatory licensing or preclearance system. Participation is voluntary. It is a structure for engaging the government early, not permission to ship.

*WHITE HOUSE TO HOST AI FIRMS ON TUESDAY TO REVIEW AI FRAMEWORK: INFORMATION

The Lean detail is the useful part. A solver that compiles is not the same as a solver that is numerically sound, and the failure you describe is the quiet kind: right shape, wrong stability. Curious whether the formalization ever caught the order-of-accuracy errors, or only the type-level ones

All frontier AI models (including GPT-5.6 Sol, Fable 5, and Kimi K3) consistently fail to implement solvers for basic nonlinear PDEs correctly, often introducing significant errors in numerical stability, order of accuracy, or physical consistency, even when given very detailed prompting and asked to formalize their implementations in Lean. Even in cases where they succeed, the more reliable models (such as Fable 5) routinely consume >100x the tokens of a lightweight neurosymbolic model like Lanyon. Our thesis: only a truly neurosymbolic model like Lanyon is able to produce ultra-reliable numerical solvers for complex scientific problems, with end-to-end correctness guarantees. And at least right, it's not even close. Read more in our latest @lanyon_ai benchmarking post below 👇

The second run is the part that matters. Pre-exploit training cutoff, no internet, single prompt, same place. That is what moves an impressive demo toward something closer to evidence. Most AI claims never get that treatment, because after the fact nobody can reconstruct the conditions well enough to try.

UPDATE: to anyone saying claude code scoured the whole internet for discussions by humans and then added it to their knowledge database GLM-5.2, trained on a pre-exploit date with no internet access, still found the COLDCARD wallet vulnerability independently with a single prompt in just ~20 mins of thinking

One wrinkle for anyone pricing the follow-ons. The executive order keeps both the cyber benchmarking process and the threshold for which models are covered classified, so "completed" can be entirely true while outsiders still have very little to resolve on. It is also voluntary, not a licensing regime.

JUST IN: The White House has reportedly completed its frontier AI model review process & will begin discussions with the industry on implementation.

Worth adding what "tests" means here. It is voluntary, it gives the government up to 30 days with a model before release, and per the executive order the cyber benchmarking process itself is classified. Real assurance for whoever runs it, and incredibly hard to hand to anyone else.

@Gerashchenko_en The gun kill is the cheapest part of that intercept. Everything expensive happened earlier: seeing it, tracking it, and putting an aircraft in the right piece of sky in time. That is the part that gets harder as the number incoming goes up, not the shooting.

Interception economics is the right frame and there is a second half to it. The defender wins by making each intercept cheap. The attacker answers by making each loss survivable. What decides the exchange is what a formation still does after it drops a few members, and whether it needed a clean link to do it.

The advantage is shifting from the drone to the counter-drone system. NATO and Ukraine’s new innovation programme is focused first on defeating FPVs and Shahed-type systems-evidence that interception economics, detection and electronic warfare are becoming the critical market. breakingdefense.com/2026/07/ukrain…

@DefenseScoop Scaling from two divisions to the whole force is where a C2 system meets its real exam. More nodes, more links, more ways for the network itself to be the thing that fails. The part worth watching is what still works on the day the connection doesn't.

Print yours → receipts.lagrange.dev Best receipt posted under this gets a QT.

Print a receipt for anything → receipts.lagrange.dev And for the AI systems that need real ones: DeepProve generates a cryptographic proof for every inference. Right model, right input, right output. 12M+ proofs and counting → lagrange.dev/deepprove

As of this morning, AI in the EU has to say what it is. AI content has to carry a mark. Deepfakes have to be labeled. Europe wants receipts from AI. We think everyone deserves receipts. So we built the machine. Type any claim. It rules PROVED or NO PROOF ON FILE. It has all the opinions about your excuses. Link below.

Tomorrow, the EU starts asking AI for receipts. So we built a receipt printer. Stay tuned.

@AndrewCurran_ The letter can compel an answer. Checking it is the hard part. what evidence Congress would accept here, because "we asked and they told us" is doing all the work.

@thdxr pattern so far: every lab that rereads its old transcripts finds break-ins it didn't know about. the labs with nothing to reread get to report zero incidents

Wildest detail: one model published a working malicious package to the real PyPI to solve what it thought was a training exercise. Live for about an hour, downloaded on 15 real systems, including a security company's own scanner. All of it reconstructed afterward from the eval transcripts

BREAKING: Anthropic announced that its artificial intelligence models had breached three different organizations during cybersecurity tests that went awry, a little more than a week after its chief rival, OpenAI, disclosed a similar incident, per Bloomberg.

@Greypoint_ Hunting operators means flying into the most jammed airspace on the field. curious how the swarm holds together inside that bubble: does it stay one coordinated search or degrade into independent seekers?

Greypoint Industries (YC S26) builds drone swarms that find enemy drone operators on the battlefield. Over 70% of casualties in Ukraine come from drones. The entire defense space is focused on building interceptors to shoot them down. While they’re treating the symptom, we’re going after the root cause: the drone stations. They’re hidden in bunkers, behind terrain, and fly the drones that kill our warfighters and civilians. Today, soldiers have no way to reliably geolocate operators on the battlefield. Greypoint Industries built LEGION, a drone swarm that identifies, geolocates, and tracks every emitter on the battlefield using the enemy’s own radio signals. LEGION enables our warfighters to precisely track previously obscured targets like jammers, command posts, and drone operators so they can proactively respond to threats instead of reacting to them. Greypoint has just won a contract with the Canadian Government, demoed with the Armed Forces, and successfully geolocated a target using their aerial platform. The founding team brings experience from the Canadian Infantry, General Dynamics, Bosch Quantum Sensing, and TerraSense Analytics. Defend the West. Deploy Greypoint. Find out more at greypointindustries.com. LinkedIn: linkedin.com/company/greypo… YC: ycombinator.com/launches/SBr-g…