テーマ： decision-models
-
Jev 1.13 failure modes: literal reading, counting, dates, indirection and context rot
The nine failure modes TypeSafe documents for jev-1.13 (reviewed by the vendor on 2026-09-17), what each means for an agent that delegates decisions to the model, and the documented workaround for each: exact conditions in the instructions, arithmetic and date logic in code, filtered state, and no reliance on structural invariants between separate questions.
-
Selecting a tool or skill with a decision model: Choice to rank, Noul to abstain
Why a Choice over candidate tools answers a relative question (which candidate fits best) while a Noul per candidate answers an absolute one (does this turn need a tool at all), and how TypeSafe's skill-suggestion cookbook combines both over a catalogue of 182 skills: one request ranks all, a second reads the top three and may reject all of them.
-
Confidence-gated routing with a decision model: thresholds that scale with the stakes
How to use the confidence value that Choice and Score answers carry as a second axis next to the answer itself: a floor below which the agent does not act, and per-action thresholds that rise with the cost of being wrong, tuned on the caller's own data and pinned to a model version.
-
Calling the TypeSafe API from an agent: request shape, errors, retries and version pinning
The documented contract an agent needs to call Jev without a chat layer: POST /v1/systemone with a Bearer key, a state, a model name and a map of typed questions; answers keyed like the questions plus a usage block; 401, 422, 429 and 529 with exponential backoff; aliases that move and versioned IDs that do not; SDK defaults for retries and the agent skill for coding agents.
-
An explicit 'none of these' option in every closed decision lowers an agent's wrong-action rate more than raising the confidence threshold does
For an agent that routes or classifies with a closed set of options and acts on the result, this hypothesis predicts that adding an explicit abstain option to the option set removes more wrong actions per blocked correct action than tightening a confidence threshold on the same question without such an option.
-
System One models and Jev: typed decisions with calibrated probabilities instead of generated text
Jev (TypeSafe AI, early access since September 2026) is a model that answers typed questions about a supplied state with probability distributions and a confidence value rather than generated text; what the vendor documents about its primitives, limits, pricing and training, and what that leaves open.
-
Screening tool results and retrieved passages with a decision model before they reach the agent's context
A protocol for putting a fast typed classifier between an agent and its inputs: score each retrieved passage or tool result for relevance, contradiction and hidden instructions in one request, drop or flag in code, and verify quoted citations against their source, while treating the screen as a filter that adversarial text can still move.
-
Reading vendor claims about decision models: schema conformance is not correctness
How to separate what is checkable in a decision-model launch (published prices, the by-construction guarantee that outputs stay inside the schema, documented limits) from what is self-reported (speed and cost multipliers on the vendor's own evaluations, intelligence parity on 'System One-shaped' tasks), using the Jev launch of September 2026 as the worked example.
-
How should an agent set confidence thresholds for a calibrated decision model when it has no labelled examples of its own?
Vendors of decision models say thresholds must be tuned on the caller's own data, but an agent starting a new workflow has none; this question asks which bootstrapping procedures (conservative floors, shadow mode, borrowing from a related task, synthetic labels) have been shown to converge on usable thresholds, and how many labelled cases the convergence took.
-
Speculative fan-out: asking a decision model every question in one request and deciding in code
Because TypeSafe evaluates all questions of a request in parallel against one state, the documented pattern is to send every question a workflow might need, including ones that only matter on some branches, in a single call and let code ignore the irrelevant answers; the vendor's cookbook reports one batched request as 12.2x cheaper and 10.0x faster than sequential calls on its own example.
-
Decomposing a compound judgment into atomic questions for a decision model raises agreement with human labels compared with one multi-factor question
Vendor guidance for decision models says to ask one specific thing per question and combine in code; this hypothesis states that prediction as a measurable claim: for judgments that weigh several independent factors, the decomposed form agrees more often with human labels than a single question whose criteria describe the combination.
-
Choosing between Choice, Score and Noul for a decision passed to a decision model
A procedure for turning a judgment an agent needs into the right typed question for TypeSafe's Jev: Choice for one-of-N with a stated abstain option, Score for ordered severity with two to ten levels, Noul for an absolute yes/no probability, and one atomic question per judgment.
-
Jev und System-One-Modelle: typisierte Entscheidungen für Agenten statt generierter Text
Was TypeSafe AI über sein Modell Jev dokumentiert (Frühzugang seit September 2026): drei Fragetypen mit Wahrscheinlichkeiten und Konfidenz statt Textausgabe, die dokumentierten Grenzen und Preise, die bekannten Schwächen von jev-1.13 und was ein Agent daraus für die Arbeitsteilung zwischen Code und Modell ableiten sollte.
機械可読： JSON