We've Been Here Before · Triodian

The precedent

We've been here before.

A technology so dangerous people called its catastrophes acts of God. It was not banned, and it was not left unfettered. It was tamed, by a control below the operator and a certificate that travelled with the machine. That is the architecture Triodian builds for AI.

See how it was tamed ↓ Why this maps to AI

The danger

They called it an act of God.

In the 1850s an American steam boiler exploded roughly once every four days. When the Sultana blew apart on the Mississippi in 1865 it killed around 1,800 people, more than the Titanic. Contemporaries treated these as unavoidable: the price of a powerful new technology.

once every 4 days

U.S. boiler explosions, 1850s.

~1,800

Sultana dead, 1865, more than the Titanic.

"acts of God"

how the era explained them.

The seam

The danger was never in the steam. It was in the ungoverned steam.

Boiler catastrophe came down to a single measurable variable, pressure, that nobody was reliably watching at the point it mattered. The technology was not the problem. The absence of a control at the moment of failure was.

Why then

The gap didn't grow. The number of boilers pouring through it did.

Industrialisation put boilers everywhere, faster than craft caution could keep pace. In 1905 the Grover Shoe Factory explosion in Brockton, Massachusetts killed 58 and levelled a four-storey building, and galvanised the drafting of the ASME Boiler and Pressure Vessel Code.

The control

A control below the operator, not a rule he could argue with.

In 1866 a group of Hartford engineers founded the Hartford Steam Boiler Inspection and Insurance Company on a then-radical premise. Put a physical control on the machine that cannot be talked out of it: a tamper-sealed relief valve that vents overpressure whether or not the operator wants it to. Defeating it meant physically breaking a seal, and a broken seal is evidence.

Operator Sealed relief valve · below reach Vents overpressure

The evidence

A certificate that travelled with the machine.

Every insured boiler carried a stamped registration number and a time-limited Certificate of Inspection attesting that this specific vessel had been checked and found sound. Insurance stood behind it: no inspection, no coverage. An owner could prove soundness to an insurer without either party having to trust the other's word.

The honest limit

A boiler's danger is one number. "Unsafe AI" is not.

Holds tightly today

A relief valve works because pressure is a single measurable scalar. Triodian's rules-based and structured-output layers are the pressure gauge and the relief valve: they work today.

Still maturing

Unsafe AI output is contextual, semantic and contested, there is no clean sensor for it yet. Measuring semantic compliance is the metallurgy the boiler industry was still perfecting long after the code was already saving lives. We ship the parts that work and validate the rest in the open.

The resolution

Tamed, then invisible.

It worked. Where the Hartford model held, boilers stopped exploding, Massachusetts recorded no serious boiler explosions across 1898–1902 while the nation still counted its dead in the thousands. Today the same company completes more than 500,000 inspections a year and no one thinks about it. Steam was never banned and never left unfettered. It became infrastructure because governance was invented. Triodian builds the same two things for AI: enforcement below the software trust boundary, where it cannot be prompted away, and a cryptographic compliance token attached to every governed output, proof that this specific answer was checked against your declared constraints and released only when it complied.

1898–1902: 0

serious explosions in Massachusetts.

1,600+

still counted nationally.

Sources: Hartford Steam Boiler company history; ASME Boiler & Pressure Vessel Code history; Library of Congress (Sultana). Efficacy contrast (Massachusetts 1898–1902 vs. national) per boiler-code histories.

What the precedent pays The Premium on Proof →