Reasoning practices in discourse

11.1 Predictions that can be checked

Practice name: rent-probe / רענט־פּראָבע. A rent test requests an observable consequence that distinguishes competing hypotheses. It can be issued through the existing directive: Tu gib a rent-probe far H. “Give a predictive test for H.” The noun adds vocabulary without requiring another verbal mood.

A forecast record should identify the event, resolution deadline, resolution method, numerical credence, assessor, and background information. “AI will be important” needs more specification before its calibration can be scored. “System X solves at least 80 of these 100 held-out problems by date T under protocol M” provides a more precise event.

11.2 Calibration and scoring

Noun: kalibratsye / קאַליבראַציע. Calibration compares stated probabilities with outcome frequencies over an appropriate collection of forecasts. Among comparable forecasts assigned probability 0.8, roughly 80% should resolve true if the forecasts are well calibrated. Small samples fluctuate, and calibration alone does not guarantee useful discrimination between cases.

For a binary outcome y in {0,1}, a Brier score is (p − y)² and a log loss is −[y log p + (1 − y) log(1 − p)]. Smaller is better. The logarithm's base sets the unit; base 2 gives bits. These scores evaluate forecasts, not a person's worth or a sentence's grammatical correctness.

An avavav forecast that resolves false receives a very large log loss. That gives the repeated near-certainty form a practical cost in a scored setting. It should not be used simply as an ornamental intensifier.

11.3 Surprise versus support

Noun: iberash-bits / איבערראַש־ביטס. The surprisal of an observed event E is −log₂P(E) under a specified model. Its support for H is log₂[P(E given H)/P(E given not H)]. Those quantities answer different questions.

An event can be astonishing under both hypotheses and provide little discrimination between them. Conversely, a mundane event can strongly distinguish two hypotheses if one made it much less likely. The lexicon should keep “surprise” and “evidential increment” separate.

11.4 Disagreement and cruxes

Noun: dreypunkt / דרײפּונקט, a belief on which an important disagreement turns. A useful exchange identifies H, each person's current credence, the evidence they share, and the observation or subclaim that would move them.

The language does not require averaging two people's log-odds. Equal averaging is not a general Bayesian theorem. A person's testimony can be evidence, but the amount depends on competence, shared information, incentives, and dependence. Merely declaring two speakers rational does not supply those assumptions.

Freg asks for the crux; nem tests a conditional concession; rik records a real update after the evidence changes. These familiar operators do most of the work.

11.5 Outside view and reference classes

Nouns: oysn-blik / אױסן־בליק, “outside view,” and farglaykh-klas / פֿאַרגלײַך־קלאַס, “reference class.” An outside-view estimate starts from relevant comparison cases. The chosen class should be named because different classes imply different base rates.

Base-rate reasoning can support a prior without making that prior privileged or immutable. A specific case may contain relevant evidence that distinguishes it from the class. Yuddish's update record makes the transition from comparison class to case evidence inspectable.

11.6 Operationalizing disputed terms

Directive vocabulary: erklern on dos vort, “explain without the word.” Under tu, a speaker can request a paraphrase that replaces a disputed label with a test, mechanism, or clearer description.

This is especially useful for intelligent, aligned, conscious, and rational. The replacement definition goes under heys or a model frame. Evidence that an object meets it goes under EF. Replacing a label with several equally vague labels does not complete the task.

11.7 Observation and intervention

Proposed technical distinction: zen X=x in a model record describes observation; arayn-grif X:=x describes intervention. In causal notation, P(Y given X=x) need not equal P(Y given do(X=x)). The interpretation requires a causal model.

No new evidential automatically turns correlation into causation. A causal claim can use =pon while its accompanying model states which interventions and assumptions justify it. Counterfactual volt frames likewise need a specified relation to the actual context.

11.8 Decisions, values, and information

Noun: nutsvert / נוץװערט, utility or decision value. The unit belongs to an explicitly chosen utility scale. It is not a probability or a bit of evidence. A decision may compare expected utilities under a model, but a directive still expresses an action recommendation rather than a truth-value assertion.

For a simplified decision model, expected utility is ΣₓP(x given action, model) U(x). Whether conditioning on an action adequately represents choosing it depends on the decision and causal assumptions. This grammar does not silently select a complete decision theory.

Informatsye-vert / אינפֿאָרמאַציע־װערט names the value of information relative to a decision. An experiment may be informative but not worth its cost. A useful request names what choice the information could change. Numerical rigor about beliefs should not erase the separate question of what the speaker cares about.