Taxonomy

Classification

❓ Is Jev a PTC, an LTC, or a slogan sitting on an old LTC?

Using the OOM taxonomy (concrete product vs abstract category vs life-altering capability):

• PTC (product-level): Jev — a named, callable, priced model from TypeSafe AI, early access 15 September 2026.
• LTC (category): TypeSafe wants “System One Model” to be a new category. The conservative classification is: Jev is a PTC of structured decision inference, which is a specialization of the existing LTC Large Language Model (or, if the unpublished architecture is not an LM at all, of discriminative sequence models). This research does not promote “System One Model” to a new LTC until there is a second independent product in the class or a paper that forces a new mechanism node.
• Capability: typed calibrated decision — given state S and a closed question Q, return a distribution over Q’s answers plus a usable uncertainty signal, fast enough to sit inside ordinary request/response software.
• Not a LAC. A life-altering capability would be unattended semantic automation at economy scale. Jev is an early PTC that gestures at that LAC. The manifesto’s TFP 3% is a bet, not an unlocked milestone.[3]
• Party: TypeSafe AI; Diogo Almeida; DCVC as investor.

Convergence. Strong with: structured outputs, tool-calling, agent judges, cheap inference, “code owns the loop.” Weak with: frontier chat, reasoning models, multimodal agents. Jev gives those up on purpose.[1]

Trend-analysis rule

❓ Is there an exponential trend here, or a stepwise product launch?

Apply the trend rule honestly:

• What looks like a trend: inference price for short, closed judgments falling toward classifier economics, while chat/reasoning tokens stay expensive. That is a measured trend in the industry (small classifiers and embedding models have been cheap for years). Jev is a stepwise product that tries to put frontier-ish language understanding on that cheap curve.
• Classification: stepwise product launch, not an exponential capability curve with a published doubling time. There is one public model, one price point, one day of independent tests.
• Jevons Paradox is the name, not a measured consumption series. Naming a model after rebound demand does not show rebound demand. Flag as naming bet.
• If latency stays under ~100 ms and calibration holds on real workflows and price holds without subsidy, then a new usage regime (judge-every-turn, map-reduce over corpora, in-loop agent gates) becomes plausible. Those are milestones still needed, not observations.

Global problems (light touch)

❓ Which listed global problems does cheaper judgment actually touch?

• Automation / labor / UBI debates: a cheap, typed judge is how you would try to take humans out of routine classification. It does not, by itself, move unemployment or TFP. The manifesto’s TFP footnote is the company’s preferred scoreboard, not evidence.[3]
• Education / evaluation: Every’s writing-check demo is a teacher’s-aide pattern (flag AI tells, flag planted defects) — a small instance of “verify everything.”[17]
• Not in scope: biotech, climate hardware, payments rails. Do not stretch.

Unresolved after this pass

❓ What would a second research round still have to get?

Architecture paper; parameter count; a public reliability diagram; residency/DPA; whether list prices survive twelve months; a second independent accuracy study larger than Every’s; any All-In / Moonshots discussion of this SKU (transcripts through 11 September 2026 predate 15 September); a Community Note if one is later attached to the launch thread.