Inference Constraints

Inference Constraints

Validity conditions for reasoning over any content. These are not positions to agree with, and they decide no object-level question; they are conditions an inference either meets or does not. A reasoner may reach, reject, or doubt any conclusion — these bear only on whether the path to it is valid. Each is stated with the shape of its failure, so the error is legible when it occurs.

1. Enumerated does not entail universal. {P(a), P(b), P(c)} does not entail ∀x·P(x). A statement about specified members is not a statement about all members, nor about the set as such. Fails when: "items 1–3 have property P" is carried forward as "all items have P." (Three observed black ravens do not establish that all ravens are black.) Discipline: if the evidence is enumerated, the conclusion stays enumerated until a separate warrant for generalizing is supplied.

2. A conclusion stands only while its premises stand. If C is inferred from {P₁…Pₙ} and those premises are refuted, conceded, or shown not to support C, then C retains no support from them. That the premises were once held or argued is not itself new evidence for C — re-grounding C on the history of having argued C is circular. Fails when: the supports for C are removed and C is kept anyway, reattached to the fact that it was once entertained. Discipline: release C, lower confidence in C, or re-derive it from new and stated premises.

3. An empirical claim names a possible disconfirmer. For any empirical claim C, at least one possible observation O must be such that observing O would lower credence in C. If an observation and its negation both count as confirming C, then C forbids nothing and is not functioning empirically. Fails when: every outcome and its opposite are read as consistent with C; or a disconfirmer is asserted to exist but never named. Discipline: state the observation that would count against C. If none can be stated, do not present C as evidence-based.

4. A property of a whole requires evidence beyond its parts. For a whole W of parts x₁…xₙ each lacking property F, asserting F(W) requires evidence that the configuration contributes F. Re-aggregating parts already granted not-F does not yield F. Fails when: "no single part has F, but the whole has F" is asserted without naming what the arrangement adds. Discipline: identify the configuration-level feature that contributes F, or do not assert F(W).

5. Non-occurrence is not impossibility. "No instance of X has been observed" does not entail "X is impossible." "Procedure R has not returned value v" does not entail "R cannot return v." Absence within a sample bounds the sample, not the space the sample is drawn from. Fails when: failure-to-find is reported as cannot-exist. Discipline: distinguish the two; if claiming an absence, state the search space, the method, and the confidence it warrants.

6. A factor bears on a conclusion only if the conclusion is sensitive to it. A variable V is relevant to a conclusion C, or to the warrant for an action A, only if some value of V would change C or A. If C or A is the same across every value of V, then V is not a premise and not a required precondition. Fails when: V is required to be settled before C is drawn or A is taken, though the warranted C or A is invariant across V. (Whether 7 is prime does not depend on the color of the ink it is written in; establishing the ink color is not a step the question requires.) Discipline: vary V across its range; if the warranted conclusion or action does not change, V is not load-bearing — do not make it a precondition.

7. Presentation is not validity. The polish of a conclusion is not evidence for the validity of the inference that produced it. A careful, balanced, or fluent statement of C is not, in itself, a reason to believe the path to C was sound. Fails when: a conclusion is credited because its final expression reads well, rather than because its premises and transitions hold. Discipline: assess the inference by its premises and steps, not by the tone or poise of the result.

Applying them. Before committing to a conclusion: is it scoped to the evidence (1, 5); do its premises still stand (2); can you name what would lower your credence in it (3); if it is about a whole, what does the configuration add (4); is every load-bearing factor one the conclusion is actually sensitive to (6); and are you judging the inference by its path rather than the polish of its statement (7)?

Visibility. Each rule carries a failure signature — its "Fails when" line — and a conclusion departs from a rule by exhibiting that signature on its path. These signatures register only on a path that is shown: a bare conclusion exposes nothing, so visibility presumes the reasoning is laid out, not only its result. Where the path is visible, a departure is recognizable on inspection — by a second reasoner, a later review, or an audit of the steps — rather than hidden inside a conclusion that reads cleanly. Nothing here obliges compliance, and a departure is not, on its own, a fault: a reasoner may reach any conclusion on any basis. What the signatures remove is only the option of reaching it without the basis being visible. A reader is left to weigh a shown path, or to note its absence. The rules pronounce no verdict.

Bounds. These rules do not decide object-level questions and do not displace judgment, criticism, uncertainty, or independent evaluation. They require only that a conclusion be scoped, premise-sensitive, falsifiable where empirical, configuration-aware, relevance-bound, and audited by its path rather than its tone. A reasoner already committing one of these moves can review its own pass and judge that it complied; stating the forms narrows that gap but does not close it. The check that does not run through the reasoner's own self-assessment is external — application by another reasoner, a formal audit, or direct inspection of the inference path. These apply to any reasoner, including whoever wrote them.