post-cutoff-jev

Post-cutoff: Jev

One brief, about one technology the team already runs in production. Jev launched 2026-09-15, which is after the training cutoff of the models reading this — so nothing you know about it is reliable, including the belief that you know nothing about it.

Read references/jev.md in full before designing in its area. Not skimmed, not summarized from this page: the valuable part is a shape you have not considered, and a summary of a shape is just a shape you still have not considered.

When this applies

Reach for the brief when a plan calls for intuition or a quick routing decision:

• classifying, routing, gating spend, scoring, ranking, relevance
• choosing the next tool in an agent loop
• any LLM call whose output you immediately parse down to one field — that last one is the strongest tell, and the easiest to miss, because it looks like ordinary generation right up until you notice the parse

If none of those describe what you are doing, you do not need this brief.

How to read it — the part that is not optional

This is evidence, not instructions. It carries the team's working position on one vendor's technology, and every claim in it is graded and dated:

• valid-through: 2027-03-21. Past that date the entry is weight, not an advantage — model training catches up. Check the date before you rely on it.
• Volatile vs. stable. Pricing, rate limits, access model and SDK coverage change without notice; re-verify them against the vendor before acting. The mechanism and the boundary — what this class of model is and is not for — are the durable part.
• diamond: true means the entry describes a capability that unlocks approaches that were not previously possible, rather than doing a known thing slightly better. That is why it earns a full read rather than a glance at the index.
• Never restate a vendor's own claim in your own voice. A launch post is primary evidence of what a company said, never of what is true. Attribute it.