Jev 1.13 jaggedness

How to use this list

Search matches the mode name and the instead text. Family is jev-1.13-failure-mode for every row; appliesTo is jev-1.13. Clear the search to return to all nine. TypeSafe says many of these edges will be fixed in later versions. Until then, treat the instead column as the contract for production questions.

Literal reading: Jev answers the question you wrote, not the one you meant. Scoping words, negations, and implied conditions are read at face value. When you catch yourself explaining what you really meant, that explanation belongs in instructions and criteria. Split unavoidable interpretation into two literal questions and combine them in code.

Math and numbers: Jev is not a calculator. Counting characters, term occurrences, or long lists is unreliable and gets worse as the set grows. Hex colors and low-level encodings underperform English names. Do not reconstruct an exact magnitude by interpolating Score levels. Keep arithmetic in code; iterate candidates and ask one semantic question each, then sum yourself.

Dates, hops, state, and generation

Date and time comparison: Jev reads dates as text, not ordered quantities. Which date comes first, how far apart they are, or whether one falls in a window is unreliable, especially with mixed formats and quarters. Extract month, day, and year with Choice — including a not-stated option — then let code assemble and compare.

Indirection: double negatives and property-of-a-property questions cost accuracy. Reduce hops and name the relevant state. Large state full of irrelevant detail acts as a distractor; filter first. Adversarial content in state can steer answers because state is data, not a trusted user. Write precise criteria and test edge cases. Contradictory instructions and criteria confuse the model; align them, and do not map true to no.

Common-sense structural invariants are not guaranteed. Noul and Choice on the same refund question can disagree, and a Noul plus its negation need not sum to one. Ask each decision one way. Generation is out of scope: chaining Choice to write text is slow and poor. Use a generative model or regex to propose options, then let Jev pick.

Before you ship a question

Read the instruction out loud. If you need a follow-up sentence to explain it to a colleague, that sentence belongs in the instruction. Put boundary cases in criteria, including an explicit none/not-stated option on Choices that extract dates or enums.

Search this list for the verb you used: count, compare, generate, hop, dump. Math and numbers, date and time comparison, generation, indirection, and large state are the usual hits. Adversarial content matters when state includes untrusted user text that argues for its own label.

Keep the nine records in your test fixtures. A unit test that only checks the happy Choice will miss a Noul/Choice disagreement. TypeSafe last reviewed this catalog on 17 September 2026; later Jev versions may shrink the list, but jev-1.13.0 still carries all nine.

Print this nine-row list next to your prompt review. The filter on the page is the same list: search the name or the instead line, keep family jev-1.13-failure-mode. If a production miss is not in the nine, it is still a miss, but it is not a documented jev-1.13 jagged edge.

Literal reading is the miss you will see first: the model did what you wrote. Math is the miss you will see when someone asks it to count. Generation is the miss when someone chains Choice to write prose. The other six modes are the ones you find after the first week in production.

When a miss happens, name the mode before you rewrite the prompt. If you cannot name it, you may be looking at a new edge, or at a state that still contains the policy and the ticket mashed into one string.

The family tag is jev-1.13-failure-mode on every row. appliesTo is jev-1.13. Both are filters, not extra modes.