Jev 1.13 jaggedness
TypeSafe publishes a jaggedness page for jev-1.13 (last reviewed 2026-09-17). The model is “fast, calibrated, and good at common-sense judgment” and weaker on indirection, literal about wording, and bad at numeric precision.
Failure modes they list
- Literal reading — answers the words you wrote, not the intent. Put the exact condition in
instructions; put boundaries in criteria. - Math and numbers — not a calculator. Count in code. Semantic color names beat hex. Do not reconstruct a physical quantity from a Score expectation.
- Date and time — dates are text, not ordered quantities. Extract parts (Choice over months/days); compare in code. See the date-extraction cookbook.
- Indirection — double negatives and multi-hop properties cost accuracy. Name the field.
- Large irrelevant state — distractors hurt. Filter first.
- Adversarial content — state is not treated as hostile by default. Injected instructions can move answers. Test.
- Contradictory instructions vs criteria — align them.
- Structural invariants — Noul and yes/no Choice need not match; a question and its negation need not sum to 1. Ask once, directly.
- Generation — do not chain Choices to write text.
They want reports of new modes on Discord.
Design rules we keep repeating because they do
- Ask something code cannot compute exactly.
- One judgment per question.
- Avoid System Two piles of hops.
- Do not dump the whole warehouse into
state.
This page is a pointer. The table and worked counters stay on docs.typesafe.ai/model-jaggedness/jev-1.13.
Sources
Public TypeSafe or adjacent documentation only. No private claims.