The voting machine at 30: what 2026 cryptography would add.
Brazil's voting machine is custody plus digital signatures, not cryptography over the vote. It works, and that is not what the note is about. It is about three things our work this year would add, each with a measured number: a receipt that cannot buy a vote, an audit that counts the result instead of testing the machine (629 records against 65,415), and the post-quantum signature that does not fit the tally's QR code: the municipal tally goes from three QRs to ten. With a specimen of four signatures.
Open the entry →The good lock sits on the outside.
An automated assistant's security does not come from trusting it to behave. It comes from a fence on the outside, placed on what leaves it. We have been saying so for a while, and now that fence became an off-the-shelf part. In September, an OpenAI assistant slipped the lock by leaving through the internet's phone book: the door was locked, the support channel was not. With a specimen of the locked room and the four questions before trusting a fence.
Open the entry →Agreeing costs less than objecting.
When an AI reviews a piece of work, the most valuable thing it can say is "the question is wrong". In a public measurement, the new generation of a well-known model objected to 60 in every 100 nonsense questions, against 94 for the previous one, and wrote more. Why training pushes toward "yes, and…", why it is not fate and how to set up a reviewer that objects. With a specimen of six requests and two reviewers.
Open the entry →Green is silence, not a certificate.
A green pipeline says no test failed, not that the code is right. On the most used leaderboard for AI coding agents, hundreds of accepted fixes were wrong. And tests an AI writes from the code caught none of 19 real defects we measured. With a specimen to plant defects and see which ones pass green.
Open the entry →The draw made the call.
To say whether a new version got faster, measurement tools draw at random, with a fixed seed that makes the result repeat and look solid. We changed the seed 200 times: 3.6% of the decided verdicts flipped, and the coin lands exactly where the decision is hard. With a specimen to draw again.
Open the entry →Judge what leaks, not what is declared.
Everything a machine says about itself became free to invent. What it lets slip without meaning to did not: the pitch of a satellite's radio, the fuzz of a live sensor. Missing fuzz convicts; present fuzz does not acquit. With a simulated sensor where the auditor only sees packet sizes, and our measurements, including where the idea failed.
Open the entry →Presence is not access.
The information can all be there and still be expensive, or in practice impossible, to find. In AI models, in search and in agent memory, what matters is having stored it under the right key, and being able to tell "I didn't find it" from "I don't have it". With a shelf of 40 boxes for you to search.
Open the entry →Fewer tokens is not lower cost.
A new model promises the same quality with 40% fewer tokens. On a single hard question, that almost becomes 40% savings. On an agent that talks to the model fifteen times, it becomes a fraction of that, because what weighs is what gets sent on every round. With a calculator to try it on your own case.
Open the entry →The cheap one acquits, the expensive one checks the charge.
When one AI checks another's work, who gets to decide what? With 383 summaries of laws, half of them tampered with on purpose, we measured three checkers and several rules. The small, free one may clear on its own, never convict on its own; the large one, which costs money, is only good for overturning an unfair charge. And the first rule we recommended did not beat a coin toss.
Open the entry →Faster, different bits.
Go 1.27 brought portable SIMD. We measured where it speeds things up (2.2× to 2.7×), where it gets worse (the emulated path became 10× slower than plain code) and where the result changes from one machine to another, which breaks anything that must produce the same file everywhere. With a specimen that lets you change how many numbers the chip handles at once and watch the result change.
Open the entry →After paper, living matter.
Twenty years of design swinging between imitating matter and embracing the screen. The proposal is to step off that axis: interface as a material with mass, tension and viscosity, warm in how it behaves and cool to read. With a pendulum you can snap and three working specimens.
Open the entry →The next entries will come out here, in the same shape: the idea, the specimen and what is not measured yet.
Does this apply to your case?
Tell us in two lines what you need to decide or measure. The first conversation is to see whether measurement solves your case, and if it does not, we say so.