Cromwell's Rule¶
Core Idea¶
Cromwell's rule is the structural injunction never to assign a prior probability of exactly 0 or exactly 1 to a contingent proposition, because Bayesian updating cannot move a probability away from those endpoints. If P(H) = 0, then for any evidence E the posterior P(H|E) = 0 as well, so the proposition is permanently unfalsifiable from below; if P(H) = 1, the proposition is permanently unrevisable. Closed beliefs are evidence-sterile: no observation, however striking, can disturb them. The rule is named for Cromwell's 1650 plea — "think it possible that you may be mistaken" — but the underlying fact is a property of any learning system that updates by combining new information with prior commitments.
The structural content rests on four commitments. There is a belief system that updates by combining evidence with priors; the update mechanism is multiplicative, since Bayes' rule multiplies prior by likelihood and zero kills the product; any closed commitment — full certainty for or against — short-circuits all subsequent learning about that proposition; and the structural prescription is to maintain a small but nonzero credence in any proposition that could in principle be true or false, so that evidence can act. The descriptive core (a multiplicative update has an absorbing boundary at 0 and 1) is pure structure; the prescriptive face ("never assign 0 or 1") is a maxim layered on top. The rule has a natural dual — never treat any single piece of evidence as carrying infinite likelihood ratio, since that pins the posterior at the boundary just as a boundary prior does — and both halves flow from the same fact that zero is absorbing for multiplication.
How would you explain it like I'm…
You Might Be Wrong
Keep A Tiny Maybe
Never Zero, Never Certain
Structural Signature¶
the belief state over a contingent proposition — the multiplicative update rule (prior × likelihood) — the absorbing boundary at 0 and 1 — the closed commitment that sterilizes evidence — the infinite-likelihood-ratio dual — the nonzero-credence prescription
The pattern is present when each of the following holds:
- A graded belief state. An agent holds a credence — a probability between 0 and 1 — over a proposition that could in principle be true or false (contingent, not logically settled).
- A multiplicative update rule. Beliefs revise by combining new evidence with the prior through multiplication (Bayes: posterior ∝ prior × likelihood), the operation whose algebra carries the whole structure.
- An absorbing boundary. Because zero kills any product, a credence of exactly 0 stays 0 and a credence of exactly 1 stays 1 under every possible observation; the endpoints are fixed points the multiplicative machinery cannot leave.
- A closed-commitment failure mode. Any proposition pinned at 0 or 1 becomes evidence-sterile — no observation, however striking, can move it — so the defect lives in where credence sits, not in the sincerity or effort of the reasoner.
- A dual on the evidence side. Treating any single piece of evidence as carrying infinite likelihood ratio pins the posterior at the boundary just as a boundary prior does; both flow from the same absorbing-zero fact.
- A nonzero-credence prescription. The structural fix: keep a small but strictly positive credence on every genuinely contingent possibility (and bound likelihood ratios), so evidence retains the power to act.
Composed, these make "stay off the boundary" a mechanical precondition for learning, separable from the moral exhortation to doubt.
What It Is Not¶
- Not
bayesian_updating. Bayesian updating is the general mechanism of revising credence by prior × likelihood; Cromwell's rule is the specific boundary constraint on that mechanism — keep priors off 0 and 1 or the update is inert. It is a corollary of the algebra, not the algebra. - Not
falsifiability. Falsifiability is a property of theories (do they forbid observations); Cromwell's rule is a property of credences (are they pinned at the multiplicative boundary). A falsifiable theory held at probability 1 is still Cromwell-violating. - Not
epistemic_humility. Humility is an attitude of acknowledging fallibility; Cromwell's rule is a mechanical fact — boundary credence is multiplicatively inert regardless of the believer's sincerity. The defect is in where credence sits, not in temperament. - Not
belief_formation. Belief formation is how credences arise; Cromwell's rule constrains where they may sit if they are to remain revisable. One is genesis, the other a no-go region. - Not
modal_reasoning. Modal reasoning concerns possibility and necessity as logical categories; Cromwell's rule concerns graded probability and the absorbing endpoints of multiplicative update — a probabilistic, not a modal, claim. - Common misclassification. Applying the rule to settled propositions. Logical truths and physically impossible claims should sit at the boundary; Cromwell's rule governs only contingent propositions, and keeping nonzero mass on refuted absurdities is paralysis, not open-mindedness.
Broad Use¶
- Bayesian statistics — the origin substrate: a prior with zero mass on a parameter value can never be updated to nonzero mass, so proper priors should be everywhere positive on the support.
- Machine learning and language modeling — zero-count events destroy probability estimates and yield undefined perplexity; smoothing methods (Laplace, Good–Turing, Kneser–Ney, backoff) exist precisely to keep unseen events at small but nonzero probability.
- Legal reasoning — the presumption of innocence is a deliberate nonzero prior on innocence, and appeals and habeas corpus exist because no finding should be infinitely irrevisable.
- Scientific practice — falsifiability requires that no theory be treated as certain, and pre-registration and adversarial collaboration assume nonzero credences in alternative hypotheses.
- Intelligence analysis — doctrines that prohibit treating any assessment as 100% certain (confidence scales rather than certainty) implement the rule institutionally; major intelligence failures illustrate the cost of closure.
- Organizational decision-making — "strong opinions, weakly held," pre-mortems, and red teams each keep organizational beliefs off the 0/1 boundary so disconfirming information can act.
- Reinforcement learning — epsilon-greedy and Thompson sampling keep nonzero exploration probability on apparently dominated options, since pure-greedy zero-exploration is Cromwell-violating.
- Ideology — theological certainty, conspiracy theory, and fundamentalism violate the rule by pinning a proposition at 1 (or its negation at 0), with predictable evidence-sterility.
Across these the substrate ranges from formal priors to legal verdicts to RL policies, while the structural fact — a boundary commitment is multiplicatively inert — is invariant.
Clarity¶
The rule sharpens the difference between low probability and zero probability, and between high probability and certainty. Informal reasoning conflates these — "essentially impossible," "basically certain" — but the structural consequences are categorically different: one permits learning, the other forbids it. Naming the rule converts a loose verbal distinction into a sharp operational one, since a credence at 0.001 can still be moved by evidence while a credence at exactly 0 cannot.
It also clarifies why dogma is not merely an epistemic vice but a structural failure mode. The dogmatic agent has placed mass on the boundary and is mechanically unable to learn, regardless of intellectual honesty; the defect is in the prior, not in the reasoning process. This relocates the diagnosis of "this person responds to disconfirming evidence by intensifying their commitment" from a moral judgment to a structural one — the prior is at the boundary, so the only available update is a re-explanation of the evidence. The clarity is to identify evidence-sterility as a mechanical property of where credence sits, separable from the sincerity or effort of the believer.
Manages Complexity¶
A single principle covers a wide family of phenomena that otherwise look unrelated: zero-frequency problems in language modeling, evidentiary closure in law, intelligence failure, religious dogma, delayed scientific revolution, and exploration failure in reinforcement learning. The unifying explanation is the same in each — a commitment of 0 or 1 prior probability makes the agent multiplicatively inert — so a long catalogue of failures collapses to one diagnosis.
The corresponding intervention vocabulary is equally unified, even as it appears under different names across domains: floor the prior (add small mass to every possibility), smooth (redistribute counts), audit for closed beliefs (find propositions held at 0 or 1 and ask whether they should be), and institutionalize revisability (appeals, pre-registration, red teams, exploration noise). Recognizing that flooring, smoothing, revisability institutions, and exploration noise are the same structural move lets a practitioner facing a novel closure problem reach for a known correction rather than invent one, managing the complexity of "why won't this system learn?" through a single mechanism with a single family of fixes.
Abstract Reasoning¶
Cromwell's rule supports several characteristic inferences. "This system's beliefs cannot move past evidence X" points to a closed commitment somewhere upstream of X. "This community responds to disconfirming evidence by intensifying its commitment" diagnoses a Cromwell violation: the prior is at the boundary, so the only update available is motivated re-explanation. "This estimator returns undefined values on unseen events" calls for a smoothing correction, the failure being structurally Cromwell. "This decision rule treats one option as never worth trying" reveals an implicit zero on that option's payoff probability, so expected return can never exceed it and the rule cannot discover otherwise — allocate nonzero exploration.
It also supports a dual reading that explains why disproving certainty is hard: if a community holds P(H) = 1, no Bayesian argument from within their framework can dislodge it, so the remedy is not more evidence but a structural intervention on the prior itself — reframing, conversion, generational turnover. The reasoning habit the prime installs is to treat any failure-to-update as a question about where credence sits rather than about the quality of the evidence, and to recognize that the boundaries 0 and 1 are not extreme confidences but absorbing states from which the multiplicative machinery cannot return.
Knowledge Transfer¶
The rule transports cleanly across substrates because the underlying mechanism — multiplicative update by evidence — is shared. From statistics into law: the prior should be positive (presumption of innocence as a nonzero prior), and evidentiary standards should be revisable (appeals as posterior updates against an ostensibly final prior). From language modeling into policy: smoothing of unseen events transfers to contingency planning for unseen scenarios, since a plan that treats some scenarios as impossible cannot adapt when they occur. From reinforcement learning into strategy: exploration noise transfers to funding long-shot projects at small but nonzero levels. And from Bayesian inference into personal epistemic hygiene: the meta-rule "stay off the boundary" transfers to the practice of holding even strong beliefs as revisable.
The transfer holds because the object underneath — a belief state, a multiplicative update rule, and an absorbing boundary at 0 and 1 — is identical whether the believer is a statistical model, a court, an intelligence agency, an RL agent, or a person. An engineer adding Laplace smoothing, a legal system providing appeals, and a planner maintaining nonzero credence in tail-risk scenarios are doing the same structural work: keep credence off the multiplicative boundary so that evidence can act. The prime is packaged as a named, prescriptive injunction with a faint evaluative charge, and that maxim framing colors its presentation; but the load-bearing content is a substrate-neutral mathematical fact — zero is absorbing for multiplication — and it is that fact, not the moral exhortation to doubt, that licenses the same correction (flooring, smoothing, institutionalized revisability) across statistics, law, science, intelligence, organizational decision-making, and machine learning.
Examples¶
Formal/abstract¶
A naive Bayes language classifier exhibits the absorbing boundary mechanically. The model estimates \(P(\text{word} \mid \text{class})\) by counting word frequencies in a training corpus and scores a document by multiplying these per-word likelihoods together with the class prior — the multiplicative update rule. Suppose the word "thaumaturgy" never appeared in the spam training set: the maximum-likelihood estimate is \(P(\text{"thaumaturgy"} \mid \text{spam}) = 0\), a closed commitment at the boundary. Now any email containing that word gets a spam score of \(0 \times (\text{everything else}) = 0\) — the absorbing zero — so no matter how many other overwhelmingly spam-like features the message carries, the classifier can never assign it to spam: the single zero count has sterilized all other evidence. This is Cromwell's rule as an engineering bug, and its standard fix is the prime's floor-the-prior / smoothing intervention: Laplace (add-one) smoothing replaces every count \(c\) with \(c+1\) over a denominator inflated by the vocabulary size, guaranteeing every word retains a small but strictly positive probability so unseen events can no longer kill the product. More refined estimators — Good–Turing, Kneser–Ney, backoff — are the same structural move with better-calibrated mass redistribution. The infinite-likelihood-ratio dual appears symmetrically: a feature treated as carrying certainty (likelihood ratio \(\infty\)) pins the posterior at the boundary just as a zero prior does, which is why robust estimators also cap likelihood ratios.
Mapped back: The naive Bayes classifier instantiates every role — per-word probabilities as the belief state, the product of likelihoods as the multiplicative update, the zero count as the closed commitment, its destruction of all other evidence as the absorbing boundary, and Laplace smoothing as the nonzero-credence prescription that restores learning.
Applied/industry¶
Legal procedure and reinforcement-learning exploration apply the same fix in human-institutional and machine-decision substrates. A criminal-justice system that treated a guilty verdict as certain (posterior pinned at 1) would be multiplicatively inert to exculpatory evidence — newly discovered DNA, a recanted confession — because no observation could move a credence already at the boundary; this is exactly the structural failure that appeals, habeas corpus, and post-conviction review exist to prevent, institutionalizing the rule that no finding should be infinitely irrevisable, while the presumption of innocence implements a deliberate nonzero prior on the defense side so prosecution evidence retains the power to act rather than meeting a zero it cannot lift. In reinforcement learning, a pure-greedy policy that always selects the currently highest-estimated action places an implicit zero on the exploration probability of every other action; since it never tries them, it can never gather the evidence that one is better, so its value estimate for an undertried-but-superior option can never rise — a Cromwell violation that traps the agent in a local optimum. The fix is the prime's exact prescription: \(\epsilon\)-greedy and Thompson sampling keep a small but nonzero exploration probability on apparently dominated options, ensuring evidence can still flow. Intelligence analysis completes a third domain — doctrines forbidding "100% certain" assessments and mandating confidence scales are the rule made institutional, since assessments pinned at certainty cannot update on contradicting collection.
Mapped back: Appeals procedures and exploration noise realize the prime end-to-end — a verdict or action-value as the belief state, evidentiary or reward updating as the multiplicative rule, a verdict-at-1 or zero-exploration as the closed commitment that sterilizes evidence, and appeals, the presumption of innocence, and \(\epsilon\)-greedy exploration as the institutionalized nonzero-credence interventions.
Structural Tensions¶
T1 — Nonzero credence versus genuine impossibility (boundary). The rule forbids 0 and 1 for contingent propositions, but logically settled or physically impossible propositions should sit at the boundary — and over-applying the rule means refusing to ever close a question, leaving credence on flat-earth or perpetual-motion claims out of misplaced humility. The failure mode is treating the maxim as universal and squandering attention keeping nonzero mass on the genuinely impossible. Diagnostic: ask whether the proposition is contingent or settled; Cromwell's rule applies only to the former, and applying it to logical truths or refuted absurdities is paralysis dressed as open-mindedness.
T2 — Small credence versus effectively-zero in practice (scalar). "Strictly positive" is the mathematical fix, but a credence of \(10^{-40}\) is operationally indistinguishable from zero and will never be moved by any evidence a real agent can gather, so the rule's guarantee that "evidence can act" is hollow at extreme-but-nonzero priors. The failure mode is satisfying the letter of the rule (nonzero) while violating its spirit (revisable in practice), so a belief is technically open but practically sterile. Diagnostic: ask how much evidence it would take to move the credence to a decision-relevant level; if no attainable evidence suffices, the prior is functionally on the boundary despite being formally off it.
T3 — Flooring the prior versus distorting calibration (measurement). Smoothing keeps unseen events nonzero, but every flooring scheme redistributes mass and thereby biases the estimates it touches — add-one smoothing notoriously over-weights rare events. The failure mode is curing the absorbing-zero problem while introducing a systematic miscalibration that degrades the very inferences the floor was meant to protect. Diagnostic: ask what the smoothing does to the well-estimated probabilities, not just the zero ones; a floor that fixes the boundary at the cost of distorting the bulk has traded a sterility failure for a calibration failure, and the redistribution scheme must be judged on both.
T4 — Prior-side boundary versus likelihood-side boundary (scopal). The rule's familiar face guards the prior, but its dual — never assign infinite likelihood ratio to single evidence — guards a different point, and an agent scrupulous about priors can still pin the posterior by over-trusting one observation. The failure mode is keeping an open prior while treating a single test result, witness, or signal as conclusive, pinning the posterior at the boundary through the evidence channel the prior-side rule never inspects. Diagnostic: ask whether any single piece of evidence is being treated as certain; boundary-avoidance must cover both the prior and the likelihood, and guarding only the prior leaves the dual route to sterility open.
T5 — Revisability versus decision commitment (temporal/sign). Holding every belief revisable is correct for learning but corrosive for action — at some point an agent must act as if a proposition is true, and perpetual openness becomes indecision. The failure mode is "strong opinions, weakly held" decaying into weak opinions never acted on, where the refusal to commit (to a verdict, a diagnosis, a strategy) is rationalized as Cromwellian humility. Diagnostic: distinguish the credence used for updating from the threshold used for acting; the rule governs the former and says nothing against acting decisively on a high-but-sub-unity credence, so invoking it to defer all commitment confuses epistemic openness with practical paralysis.
T6 — Individual update versus structural closure (scalar). The rule is stated for a single agent's belief state, but boundary commitments often live at the level of a community or institution whose collective prior is pinned even when individuals waver. The failure mode is prescribing the individual fix (floor your credence) for a closed system — a dogmatic community, a captured institution — where no member's private update can move the shared boundary commitment, because the closure is enforced socially. Diagnostic: ask whether the boundary credence is held by a person or by a structure; the prime's dual reading already notes that dislodging \(P(H)=1\) at the collective level needs structural intervention (reframing, turnover), not more evidence, so applying the individual smoothing fix to an institutional closure misses where the boundary is actually anchored.
Structural–Framed Character¶
Cromwell's rule sits at the middle of the structural–framed spectrum — a balanced hybrid in which a substrate-neutral mathematical fact is delivered as a named, prescriptive injunction. Its frontmatter grade (label framed, aggregate 0.5) records the even split: all five criteria sit at 0.5, none reaching either pole.
The tension runs through every diagnostic and has a single source: the prime has two faces. Its descriptive core — zero is absorbing for multiplication, so a credence at 0 or 1 is multiplicatively inert — is pure formal structure; its prescriptive face ("never assign 0 or 1," "think it possible that you may be mistaken") is a maxim layered on top. Vocabulary travels partly (0.5): the absorbing-boundary fact restates cleanly as Laplace smoothing in NLP, \(\epsilon\)-greedy exploration in RL, and the presumption of innocence in law, yet the named-maxim packaging and its doubt-vocabulary ride along. Evaluative weight is mixed (0.5): the underlying algebra is value-neutral, but the injunction carries a faint moral charge against dogma and toward humility. Institutional origin is mixed (0.5): the multiplicative fact is a property of any Bayesian updater, but the prime as named is the eponymous Cromwell injunction with a 1650 provenance. Human-practice-boundedness is mixed (0.5): the fix applies to a naive Bayes classifier and an RL agent — pure machine substrates — yet the prime is most often deployed as a prescription for human reasoners and institutions. And import-vs-recognize is mixed (0.5): invoking the rule does recognize a real absorbing-boundary fact, but it also imports the prescriptive "stay off the boundary" frame rather than merely spotting a multiplicative regularity.
The entry is explicit that the load-bearing content is "a substrate-neutral mathematical fact — zero is absorbing for multiplication" while the maxim framing "colors its presentation." That two-faced character — formal core, prescriptive packaging — is exactly why the prime balances at the spectrum's center, consistent with the assigned 0.5.
Substrate Independence¶
Cromwell's rule is substantially substrate-independent — composite 4 / 5 on the substrate-independence scale. Its descriptive core is a substrate-neutral mathematical fact — zero is absorbing for multiplication, so a credence of exactly 0 or 1 is multiplicatively inert under Bayes' rule — and its domain breadth (4) is wide: Bayesian statistics, language-model smoothing, the presumption of innocence and appeals in law, falsifiability and pre-registration in science, confidence-scale doctrine in intelligence analysis, pre-mortems and red teams in organizational decision-making, exploration policies in reinforcement learning, and the evidence-sterility of dogma. Structural abstraction sits at 4 because the absorbing-boundary fact is pure formal structure that restates as Laplace smoothing, \(\epsilon\)-greedy exploration, or institutionalized revisability, even though the named-maxim packaging and its doubt-vocabulary ride along. The component that lifts the prime off the human-institutional band is that the fix applies to fully mechanical substrates — a naive Bayes classifier whose zero count sterilizes all other evidence, an RL agent whose pure-greedy policy pins exploration at zero — where no human reasoner is in the loop. Transfer evidence is a strong 4: the same correction (floor the prior, smooth the estimates, bound likelihood ratios, institutionalize revisability) recurs concretely and is documented across statistics, law, science, intelligence, organizational decisions, and machine learning. The mathematical fact travels broadly into formal tooling; only the prescriptive named-injunction framing holds the composite at a solid 4.
- Composite substrate independence — 4 / 5
- Domain breadth — 4 / 5
- Structural abstraction — 4 / 5
- Transfer evidence — 4 / 5
Relationships to Other Primes¶
Parents (1) — more general patterns this builds on
-
Cromwell's Rule is a kind of Bayesian Updating
Cromwell's rule is a specific boundary CONSTRAINT on bayesian_updating: because zero is absorbing for multiplication, contingent priors must stay off 0 and 1 or the update is inert. A corollary of the algebra, a specialization of the general updating mechanism.
Path to root: Cromwell's Rule → Bayesian Updating → Inductive Reasoning
Neighborhood in Abstraction Space¶
Cromwell's Rule sits in a moderately populated region (44th percentile for distinctiveness): it has near-neighbors but no dense thicket of synonyms.
Family — Inference & Evidence (26 primes)
Nearest neighbors
- Paradox of Unanimity — 0.73
- Texas Sharpshooter Fallacy — 0.72
- Eventual Consistency — 0.71
- Belief Formation — 0.71
- Evidence — 0.71
Computed from structural-signature embeddings · 2026-06-14
Not to Be Confused With¶
The most fundamental confusion is with bayesian_updating, of which Cromwell's rule is a corollary rather than a synonym. Bayesian updating is the general mechanism — posterior proportional to prior times likelihood — by which any belief state revises on evidence. Cromwell's rule is a specific structural constraint that this mechanism imposes: because zero is absorbing for multiplication, a credence of exactly 0 or 1 cannot be moved by any evidence, so contingent priors must stay strictly off those endpoints. The distinction matters because one can fully understand and correctly apply Bayes' rule and still violate Cromwell — by placing a boundary prior, or (the dual) by treating one observation as carrying infinite likelihood ratio. Bayesian updating is the engine; Cromwell's rule is the warning that the engine has two fixed points it can never leave. An analyst who equates them misses that Cromwell is a named failure mode of Bayesian updating, with its own diagnostic ("where does credence sit?") and its own family of fixes (flooring, smoothing, bounding likelihood ratios).
A second genuine confusion is with falsifiability. Both concern beliefs that resist disconfirmation, but they operate at different levels. Falsifiability is a property of a theory's content — whether it forbids some possible observation, so that the observation, if made, would refute it. Cromwell's rule is a property of an agent's credence — whether the probability assigned to a proposition is pinned at the multiplicative boundary. A theory can be perfectly falsifiable in principle yet held by a community at probability 1, in which case no observation will move them even though the theory itself makes refutable predictions: that is a Cromwell violation, not a falsifiability failure. Conversely, an unfalsifiable theory is a defect in the theory's structure regardless of any credence. Confusing the two leads to the error of thinking a falsifiable theory is automatically revisable, when revisability also requires that the believer's prior stay off the boundary.
A third confusion worth marking is with epistemic_humility, which the prime's maxim framing ("think it possible that you may be mistaken") strongly invites. Epistemic humility is a disposition — an attitude of acknowledging one's fallibility. Cromwell's rule, stripped of its exhortative packaging, is a mechanical fact: a credence at 0 or 1 is multiplicatively inert no matter how humble or arrogant the believer. The prime's clarity section makes exactly this point — dogma is "not merely an epistemic vice but a structural failure mode," and the defect lives in the prior, not in the reasoner's character. The distinction is practically important because the fix for a Cromwell violation is structural (move credence off the boundary, institutionalize revisability through appeals or exploration noise), not attitudinal (resolve to be more humble). Treating Cromwell as a synonym for humility relocates a mechanical correction into the register of virtue, where it loses its operational teeth.
For a practitioner the distinctions converge on one diagnostic discipline. When a system "won't learn," do not ask whether its theory is falsifiable, whether the reasoner is humble, or even whether Bayes' rule is being applied — ask the Cromwell question: where does the credence sit? If it is pinned at 0 or 1 (or a single piece of evidence is treated as certain), the failure is the absorbing boundary, and the remedy is to floor the prior, smooth the estimates, bound the likelihood ratios, or — at the collective level — intervene structurally on the shared commitment, none of which the neighboring concepts prescribe.
Solution Archetypes¶
No catalogued solution archetypes reference this prime yet.