Polarity and monotonicity

No sampled transaction contained an unauthorized payment.

Someone in the room concludes there were no unauthorized payments. Before any statistical generalization is supplied, that literal inference already fails: it widens the sentence’s domain in a direction the quantifier does not license.

Look at what the same sentence does support. It entails that no sampled transaction contained an unauthorized international payment. It entails that no sampled wire transfer contained an unauthorized payment. Both of those substitutions move to something narrower, and no carries the inference through without complaint. The move to all transactions goes the other way — from a smaller set to a bigger one — and under no, that direction is simply closed.

Swap in a different quantifier and every direction flips:

Some employee leaked a document.

This does not entail that some contractor leaked a confidential document. The leaker might be a salaried manager and the document an unclassified memo. Here the safe substitutions run outward: contractor to employee, confidential document to document, so from “some contractor leaked a confidential document” you may conclude the original.

That directional behavior is monotonicity, and it is the closest thing ordinary language has to a built-in inference engine. Learn which direction each word licenses and a large class of arguments becomes checkable at a glance.

Inclusion beneath ordinary words

The engine runs on set inclusion, which sits under vocabulary you already use: every poodle is a dog, every dog is an animal, everything that sprints moves, every encrypted personnel file is a confidential document.

Poodle is narrower than dog because poodles are a subset of dogs. Animal is broader than dog. Sprint is narrower than move, since sprinting guarantees movement while movement guarantees nothing about sprinting. Monotonicity is then one question: if I replace an expression here with something broader, or with something narrower, does the whole sentence stay true?

An upward-entailing position lets you move from narrower to broader:

Some poodle sprinted. Therefore some dog moved.

Both replacements widen, and truth survives — if the specific claim holds, the general one cannot fail.

A downward-entailing position lets you move from broader to narrower:

No dog moved. Therefore no poodle sprinted.

If the broad exclusion holds, every special case inside it is excluded too.

A nonmonotone position licenses neither direction, and we will meet one shortly.

One point does most of the work in practice: upward and downward describe a position inside a sentence, not a word on its own. The same quantifier can create one direction in its noun phrase and the opposite direction in its predicate — which is exactly what every does.

The quantifiers, one at a time

Some is upward in both positions. “Some poodle barked” entails “Some dog barked”; “Some dog sprinted” entails “Some dog moved.” Run either backwards and it fails — the barking dog need not be a poodle, the mover need not have sprinted. Existential claims are cheap to make true and they collapse the moment you try to make them more specific.

Every is the interesting one, because its two positions disagree. The restrictor — the noun phrase right after every, which fixes what the claim ranges over — is downward. The predicate is upward.

• “Every dog slept” entails “Every poodle slept.” Narrowing the restrictor is safe.
• “Every dog sprinted” entails “Every dog moved.” Broadening the predicate is safe.

That asymmetry explains an error that looks like generosity and behaves like fraud. Evidence that every specialist met a standard tells you nothing about whether every employee met it — broadening the restrictor is invalid. Yet evidence that every specialist met a demanding standard does show they met every weaker standard the demanding one entails, because that is the predicate position, and it runs the other way.

No is downward in both positions, and that is the source of its unusual reach. “No employee entered” entails “No contractor entered”; “No employee moved” entails “No employee sprinted.” Excluding a broad region excludes every region inside it. It also means, as the audit sentence showed, that a negative finding never grows.

At least n is upward in both positions: “At least three poodles arrived” entails “At least three dogs arrived,” since the count survives broadening. At most n is downward in both: “At most three dogs arrived” entails “At most three poodles arrived,” since an upper bound survives narrowing.

Exactly n is nonmonotone — neither direction works. “Exactly three poodles arrived” fixes nothing about how many dogs arrived; broadening can add members. “Exactly three dogs moved” fixes nothing about how many sprinted; narrowing can remove them. Exact counts still entail bounds such as at least three and at most three; they are inert only with respect to category widening and narrowing.

Hardly any, without, and cardinal uses of few create downward-entailing environments too: “Few employees entered” supports “Few contractors entered” when few means fewer than a fixed contextual number, and “She completed the migration without using a tool” supports “without using a paid tool.” A proportional reading of few can break the first inference: ten entrants may be few employees in a company of thousands but all the contractors in a group of ten. The direction depends on which comparison the context fixes.

The table gives valid directions; it does not license ignoring scope. “Not every A is B” behaves nothing like “Every A is not B,” and a modal or attitude verb can block a substitution that would be perfectly safe in a plain extensional claim. Fix the structure first, then read the direction off the table.

Negation reverses everything

Put negation over an upward position and it becomes downward:

• “A dog moved” lets you go dog → animal and moved → changed position.
• “No dog moved,” read as it is not the case that some dog moved, lets you go dog → poodle and moved → sprinted.

This is why a broad negative claim entails narrower negative claims. If the review found no misconduct, it found no bribery, because bribery is a kind of misconduct — a genuine entailment. Reverse it and you get the mistake auditors and journalists make weekly: finding no bribery leaves fraud, coercion, and self-dealing entirely open.

Stacked operators need care. “Not every animal moved” entails “Not every animal sprinted,” since the predicate of every is upward and negation flips it downward, and sprinting is narrower than moving. It does not entail “Not every dog moved” — the single stationary animal might be a cat while every dog ran. Establish the scope first; monotonicity only tells you the direction once the structure is fixed.

Why any clusters under negation

The any in “I did not find any defect” is a negative polarity item: an expression that only survives in certain environments, and those environments are mostly the ones that reverse or suspend ordinary upward inference.

• “No reviewer found any defect.”
• “Few reviewers found any defect.”
• “I doubt that any reviewer found a defect.”
• “Did any reviewer find a defect?”
• “If any reviewer finds a defect, stop the release.”

The plain positive “A reviewer found any defect” is ungrammatical on the intended meaning; English wants a defect or some defect there. Scope matters as much as the licensor’s presence: “No reviewer found any defect” is fine, while “Any reviewer didn’t find a defect” is not, because the licensor no longer takes scope over the polarity item.

One way to see the pattern is through alternatives. Any defect activates a wide range of alternatives — defects of every kind and size. In an ordinary positive sentence that range is unhelpfully vague. Under negation the same width produces the strongest available claim: not one member of the range qualifies. The environment turns a weak expression into an informative one, which is a decent account of why the two travel together.

Downward entailment is not the whole licensing story, though. Conditional antecedents are downward-entailing on standard analyses and are also nonveridical: they do not present the embedded event as fact. Questions license any without providing the same simple entailment direction. Some licensors are weaker than others, and strict items behave differently from permissive ones. The theory extends well past a single table; the table still covers most of what you will meet.

Other negative polarity items mark the same terrain: ever (“No version has ever passed”), at all (“The patch did not help at all”), yet (“The job has not completed yet”), and idioms like lift a finger or budge an inch. Treat them as instrumentation. When any or ever sounds natural in a sentence, something above it — negation, a decreasing quantifier, a question, a conditional — is creating a reversed environment, and your substitutions there must run inward.

Free-choice any is a different word

“Any employee may apply” has no licensor and does not mean some employee. It means the choice does not matter: pick whichever employee you like, and the permission holds. “Choose any seat” is the same. Negative-polarity any widens a range under a licensor to make a strong negative claim; free-choice any asserts indifference across the range, and it typically needs a modal or generic context to appear at all. Under negation, the readings can compete: “You may not press any button” normally uses negative-polarity any as a blanket prohibition, while rarer scope and free-choice readings require strong context. In ordinary instructions, state the intended permission or prohibition explicitly.

Four claims, worked

A policy. “Every incident must be reported.” If a security breach counts as an incident, the rule already covers breaches — that is the downward restrictor doing its job. The reverse fails: from “every security breach must be reported” you cannot conclude that every incident must be. Organizations discover this the hard way when a narrow rule is cited as though it were a broad one.

An audit finding. “No sampled transaction contained an unauthorized payment.” Both positions are downward, so you may narrow freely — no sampled transaction contained an unauthorized international payment, and so on. What you may never do is widen the restrictor from sampled to all, which is the step the opening of this chapter took and the step that turns a finding into a false assurance. The sample may still license a statistical inference, but that argument needs premises about how the sample was drawn. It does not come from the sentence.

A marketing claim. “Some customers reduced operating costs.” Upward in both positions, so it supports broader redescriptions of the same outcome — those customers reduced costs, and reduced expenditure. It supports nothing narrower: not enterprise customers, not reduced infrastructure costs, and certainly not most customers or every customer can. An existential claim is the weakest useful thing a vendor can say and the easiest to make literally true.

An absence of evidence. “Investigators did not find any evidence of tampering.” This reports the outcome of a search, and monotonicity alone will never turn it into “no tampering occurred.” That step needs a separate detection premise — that had tampering occurred, this investigation would probably have found evidence of it. The negative-polarity any strengthens the description of what was not found, exhaustively; it does not supply the bridge from what was detected to what is real.

Where substitution stops being safe

Monotonicity assumes genuine inclusion under the relevant sense of the words, and four things routinely break that assumption.

Lexical sense. A bank holding deposits is not a special case of a river bank; run in “run a company” includes no sprinting. Vague categories are worse than ambiguous ones, because the subset relation itself becomes arguable — is a near-miss an incident? People will fight about the substitution while believing they are fighting about the facts.

Attitudes and modals. These environments block substitutions that would be fine in a plain extensional claim. Lois can believe Superman flies without believing Clark Kent flies, though they are one person. A team required to hire a doctor is not thereby required to hire a cardiologist: hiring a cardiologist would satisfy the obligation, but the obligation does not reach down and select that specialty. Satisfying a requirement and being required are different relations.

Presupposition and side content. The direction table tracks truth conditions and nothing else. “Even a poodle passed” ranks the poodle against alternatives, so replacing poodle with dog changes the ranking even where set inclusion is undisturbed.

Collectivity. Predicates that apply to a group as a unit do not distribute to its members, so inclusion of individuals buys you nothing. The team surrounded the building; no member did.

Monotonicity, then, is not a universal substitution license. It is one precise question asked in one place: inside this operator, under this reading, does the inference survive a move to a broader or a narrower set? Ask it and a surprising amount of everyday argument becomes checkable — including the steps that feel most obvious, which is where the errors live.