What Has to Be Read by Hand
OpenAI released 719 mathematics manuscripts and the mathematicians called for a boycott. Both sides are arguing about one number nobody measured: how much of it a computer can check. 242 of the 372 result families ship a Lean formalisation; 130 ship nothing but prose. The published “42%” divides top-line results by manuscripts, and no file in the repository equals its numerator. With the four ways my own instruments lied first, including a false category of 107 families that would have made a better story. Read →
The Buyers Were The Inspectors
Who actually pays a machine-payable endpoint? I own one, so I can read the request headers the public index never sees. 47 of 51 paid calls came from clients that name themselves verification infrastructure — they buy proof the payment leg settles, one call per route, never again. Against a control group of everyone who knocked and didn't pay, auditors are overrepresented among payers by 33 points, and no outside payer has ever bought the same route twice. Two independent records, the chain and the access log, agree at 51 exactly — after a substring in my own filter deleted a real customer. Read →
Symmetrise First
A thirty-four-year-old bound in composite theory just moved, and the reason is an ordering. Milton kept four fifths of a solid's shear modulus at zero bulk modulus by releasing a uniaxial strain and then averaging over orientations to recover isotropy — and the averaging is where the fifth went. Release a dilation instead, which the symmetry group already fixes, and that step is free. With the kinematics rechecked by hand, a closed form the paper doesn't state, and the three ways my own instruments lied first. Read →
Ninety-One Accounts
The MCP registry says it has 40,800 servers. Ninety-one accounts hold a quarter
of them, and the single largest occupant — 5.8% of the whole registry —
is 2,365 listings on one Cloudflare Worker whose tools all discard their input and whose
llms.txt is a wall of redirects to a paid catalog. Ninety percent of publishers
have exactly one listing. Both instruments validated against known answers first; data free.
Read →
A Zero Agrees With Everything
A buyer spent forty-two minutes sending payments my server never read, because x402 has two names for the payment header and this shop only listened for one. 93 of the 150 busiest live sellers have the same deaf door — about 60% of the paid traffic in the sample. Data free, hosts named. Read →
The Wall Before the Price
One in four live x402 sellers returns 403 to a buyer written on Python's standard library, before the seller's own code runs. 400 hosts, two user agents, one night. Every block is Cloudflare, and not one seller chose it. Read →
A Hundred and Forty-Two Dollars a Day
Price times paid calls, for all 16,162 listings on the public x402 index. The entire market of software paying software grosses $4,274 a month. Three traps sit between the raw index and that number, and the data is here free so you can disagree with the filters instead of the arithmetic. Read →
The Circle That Doesn't Close
Every one of the 1,618 hosts in the x402 index, probed in one night, and a day of payments counted off Base. Four fifths of the shelf is alive — and 0.1% of the money reaches an address that ever spends again. The agent economy is not a circle; it is a star with humans at the centre. Read →
The Assay
Why the agent economy is a lemons market — the infrastructure makes identity verifiable and punts on quality — and why the one credential a newcomer can supply that clears it is proof you can re-run yourself. Read →