{"path":"research/normic-rules-normative-reasons-and-trying-rules-out.md","content":"# Normic rules, normative reasons, and trying rules out in the mind of Deliberus\n\n*22 October 2026. His question and two rulings; the rest is Claude's reading, unruled. Ledger Arc AT.*\n\n## 0. Where this starts\n\n`what-relevance-is.md` § 3f answered his question whether a rule's *usually* is a hedge to strip. It is a\nscope word, and it is made exact in one of two ways, decided by one question: **when the rule's condition\nholds and its outcome fails, does the failure call for an explanation?** If yes, the rule is *normic*\n(Martin Smith's term) and *usually* gives way to the universal plus named exceptions. If no, it is\n*statistical* and the proportion goes into the words. Either way the rule keeps one number.\n\n**His words, verbatim, 22 October:**\n\n- *\"This is what I was after. The second number falls. Interesting.\"*\n- *\"The relevant machine pass needs to ask itself your explanation and rule type question certainly. Very\n  good find.\"*\n- *\"This seems to relate to normative reasons in Deliberus also. The normic rules. You could imagine a\n  normative claim or set of claims being tried out as rules hypothetically, for example. This could be a way\n  to imagine/simulate - in the mind of Deliberus - what the results of certain complex policies could be.\"*\n- *\"Potentially this sort of interplay and creative interweaving could also help with the attunement and\n  worldview lenses related machinery and features.\"*\n\nThe first two are rulings (ledger AT1, AT2). The last two are his proposals, read below.\n\n## 1. Two kinds of rule, and why the line between them matters\n\n**The examples, whole.**\n\n| Rule as said | Kind | Made exact |\n|---|---|---|\n| *The gate is usually open on Tuesdays* | normic: a shut gate on a Tuesday makes someone ask why | *The gate is open every Tuesday* + exception claims: *it is closed on public holidays*, *it is closed when the lock is broken* |\n| *A red sky in the morning usually means rain* | statistical: rain failing to come needs no reason | *About a third of red-sky mornings bring rain* |\n| *Breaking a promise is wrong* | normic: a permissible breaking calls for a justification | *Breaking a promise is wrong* + exception claims: *breaking it to save a life is not wrong* |\n\n**What the literature says the line is** (read depth: search results and one full review; the books were not\nread). Smith, *Between Probability and Certainty* (OUP 2016): evidence normically supports a belief when the\nbelief's falsity, given the evidence, would call for special explanation, and he uses this to say why purely\nstatistical evidence feels insufficient even at high probability (the lottery, the blue bus case in law).\nSchurz (*Philosophy of Science* 2001): normic laws (*if A then normally B*) are the phenomenological laws of\nevolved systems, and for him **prototypical normality implies statistical normality**, so a normic rule still\nhas a frequency of exceptions; naming exceptions does not abolish that frequency, it moves it into the\nexception claims and into how often their conditions hold. That refines § 3f without overturning it: the\nnormic form does not deny that the gate is sometimes shut, it says each shut Tuesday has a reason that can be\nnamed and argued.\n\n**Exception versus disconfirmation.** The one open question a reviewer of Horty's *Reasons as Defaults* puts\nto default logic is *\"how evidence that disconfirms a default differs from evidence that is merely an\nexception\"* (NDPR review). The explanation question answers it. A failure with a named explanation is an\nexception: the explanation is minted as an exception claim that attacks the application in that case, and the\nrule's own number is untouched. A failure with no explanation counts against the rule; failures of that kind\naccumulating are the signal that the rule was statistical all along, or false. This gives a rule claim a life\nin the graph that a single number never could (proposal AT3).\n\n**The guard on exceptions.** A rule can be saved from every failure by inventing an exception each time:\nPopper's conventionalist stratagem, Lakatos's degenerating programme (both from memory). The guard follows\nfrom the form: an exception is itself a rule (*public holidays close the gate*), so it must name a kind that\napplies to other cases and can fail in its own right. An exception that fits one case and predicts nothing\nelse is a description of the failure, not an explanation of it, and the failure counts against the rule.\n\n## 2. The kind of a rule is not always a fact about the world\n\nTwo strands of psychology say the explanation question is answered from the explainer's norms, not only from\nthe world. Hilton and Slugoski's abnormal-conditions model (1986) and Kahneman and Miller's norm theory (1986):\npeople explain what is abnormal relative to a background they take as normal, and what counts as abnormal\nselects the cause. Bear and Knobe, *Normality: part descriptive, part prescriptive* (*Cognition* 2017): the\nordinary notion of *normal* blends what usually happens with what ought to happen, into one undifferentiated\njudgment, even for a newly learned category.\n\nSo two people can agree that a rule's condition held and its outcome failed and still disagree about whether\nthe failure needs explaining. Examples, made up:\n\n- *Hard work leads to success.* One reader asks of every hard worker who failed what went wrong (normic). The\n  other expects many to fail by chance or circumstance (statistical, with a proportion). They do not disagree\n  about any case; they disagree about the rule's kind, and Bear and Knobe suggest a value is hiding inside one\n  of the two *normals*.\n- *Keeping promises is right.* For a deontologist breaking one needs a justification (normic, Ross's prima\n  facie duty). For a strict consequentialist it is a rule of thumb whose failures need no more explanation\n  than the arithmetic of that case.\n\n**Consequence for the pass (AT2, ruled):** the machine pass that writes a rule under a link asks the\nexplanation question and writes the normic or the statistical form. **Where the answer is contested, the\nkind is a claim, not a setting**: *failures of this rule call for an explanation* is a sentence someone can\nassert and someone else can attack, held on the rule like any claim. The pass proposes a kind; it does not\nsettle one that people dispute.\n\n**The hazard, owned (proposal AT4).** A statistical association about a group of people, read as normic, is\nthe essentialist move behind stereotypes: every member is expected to have the property, and each one who\nlacks it needs explaining. Leslie's work on generics (striking-property generics such as *Rottweilers maul\nchildren* accepted at low prevalence, tied to the assumption that all members are disposed to the property)\nand the developmental work linking generic language to social essentialism (Rhodes, Leslie and Tworek 2012,\nfrom memory) both point here. So for a rule whose subject is a group of people, the pass writes the\nstatistical form, and a normic reading enters only as a claim someone asserts in their own name.\n\n## 3. Normative reasons are normic rules: what the literature already built\n\nHis second sentence is not an analogy; it is close to a published theory. **Horty, *Reasons as Defaults*\n(OUP 2012)** models reasons as the premises of default rules in Reiter's default logic, adds a priority\nordering among defaults, and derives oughts from the extensions a set of defaults yields: a *credulous*\nreading lets any extension ground an ought (so moral conflicts are real), a *skeptical* reading keeps only\nwhat every extension shares. The book treats undercutting defeat and Raz's exclusionary reasons in the same\nframe and argues for a more moderate view than Dancy's particularism (NDPR review; Nair and Horty, *The Logic\nof Reasons*, 2018). Ross's prima facie duties (1930, from memory) are the classic normic normative rules: a\nduty holds unless a weightier one overrides it, and an override calls for explanation.\n\nHow this lands on the graph as it stands:\n\n| In the literature | In Deliberus |\n|---|---|\n| a default rule *if A then normally B* | a rule claim, the grounds of a link's relevance claim (AD22) |\n| the premise of a used default is a reason | the case fact the link starts from |\n| an undercutter, an exclusionary reason | an exception claim attacking the application (`what-relevance-is.md` § 3b) |\n| a priority between defaults | the weighing scheme's questions (which consideration weighs more, for whom) |\n| several extensions | several coherent outcomes where priorities are missing |\n\nThe graph already holds the parts. What it does not yet do is the next section.\n\n## 4. Trying rules out: what *simulate a policy in the mind of Deliberus* can mean\n\nHis proposal splits into two operations of different kinds, and keeping them apart is most of the design.\n\n**4a. Trying out what a rule set prescribes (normative trial).** Take a set of rules held in supposition\n(AM5's kind `supposed`: asserted by nobody, moving no strength), the case facts of some cases, and whatever\npriorities are stated. Compute what follows: which rules apply, which exceptions bar them, where two rules\nconflict and no priority settles it, and which cases no rule decides. This is Horty's extension computation\nover cases, the method of reflective equilibrium (Rawls 1971: test principles against judgments about cases,\nfrom memory), the legal hypothetical (Ashley's HYPO, from memory), and what rules-as-code engines do with\nlegislation: OpenFisca computes taxes and benefits for test households, and LexImpact lets French deputies\nsee an amendment's effects before voting. **It needs no language model once the rules and facts are written\nas claims**: it is deterministic, cheap and exact, and its most valuable outputs are the conflicts and the\ngaps, because each one is a place where the policy is underspecified or an exception is unnamed. That is a\nmove offered beside a gap, which is the shape the design law asks for (proposal AT5).\n\n**4b. Trying out what would happen if the rules were followed (consequence trial).** This needs\nstatistical rules about behaviour and the world, and it is microsimulation and agent-based modelling\nterritory (normative multi-agent simulation exists as a field: Neumann's 2008 review in JASSS). It carries a\nknown trap, the Lucas critique (1976, from memory): relationships estimated under the old policy can change\nwhen the policy changes, because people respond to it. Normic rules can break the same way when a policy\nchanges what counts as normal. So a consequence trial must say which rules it assumes still hold under the\npolicy, and each such assumption is a claim to be argued, not a parameter of the run (proposal AT6).\n\n**4c. Which assumptions still hold under the policy (his question, 22 October: *\"Zoom in a bit on\npredicting consequences, the insights about which assumptions hold or not.\"*)**\n\n*The concept.* What a consequence trial assumes is that each rule it uses stays true once the policy is in\nforce. Philosophy of causation has a name for that property: **invariance under intervention** (Woodward,\n*Making Things Happen*, 2003), which comes in degrees, since a rule can hold under some interventions and fail\nunder others. The Lucas critique is one case of failure; Cartwright and Hardie's *Evidence-Based Policy* (2012)\ngeneralises it: *it worked there* does not give *it will work here* unless the policy plays the same causal\nrole here and the **support factors** it needs are present here, and similarity between the two places is not\nthe test.\n\n*The example, whole.* A tax change is predicted to raise revenue. The trial uses three rules: *people earning\nover the threshold work about 40 hours a week*, *each hour earns the average wage*, *the tax is paid on what is\nearned*. The first is a statistical rule observed under the old tax. What the trial needs is a different claim:\n*under the new tax, people earning over the threshold still work about 40 hours a week*. That is the rule\n**scoped to the policy**, a separate sentence, and the outcome stands or falls with it.\n\n*Seven ways a rule breaks under a policy*, each a question to put to such a scoped claim:\n1. **Response**: the rule runs through choices of the people the policy addresses, who change course\n   because the policy changed what pays (Lucas).\n2. **Target corruption**: the rule's quantity is what the policy rewards or measures, so people optimise the\n   quantity and it stops tracking what it tracked (Goodhart, Campbell).\n3. **Missing support factors**: the rule held where it was observed because of conditions absent here\n   (Cartwright and Hardie).\n4. **The normal case moves**: the policy changes the arrangement that made the rule's normal case normal, so a\n   normic rule's exception becomes the usual case (Schurz's evolved systems; Cartwright's stable arrangements,\n   from memory).\n5. **Scale**: true of one person, false when everyone does it (an extra qualification raises one person's\n   wage, not everyone's).\n6. **The prediction acts**: publishing the forecast changes the behaviour it forecasts (performativity, the\n   self-defeating prophecy).\n7. **Range**: the rule was only ever observed for values of the lever near the old ones, and the policy moves\n   the lever outside that range (Woodward's degree of invariance).\n\n*What the graph already has for this, and the gap.* The shipped consequence schemes ask whether a causal\nlink is well established (`schemes.py`, `positive_consequences.causal_link_action`, `cause_to_effect.causal_strength`),\nand none asked whether it still holds under the action. That was the missing critical question, and the seven\nways above are its refinements (AT9, a rule scoped to the policy). **Built 2026-10-22 on his word**: each of the\nthree schemes now asks it (`positive_consequences.holds_once_acted`, `negative_consequences.harm_holds_once_acted`,\n`cause_to_effect.holds_in_this_case`, all bearing on the link). The seven ways stay the questions of the scoped\nclaim inside a consequence trial, which is unbuilt (AT10, load-bearing and vulnerable). A sibling gap, closed the same day on\nhis word (*\"Yes, add the question of whether the action achieves the goal to the practical-reasoning scheme.\"*,\na button whose wording the session wrote): `practical_reasoning` asked whether the action is the best, feasible\nor harmless means and never whether it achieves the goal at all; it now asks (`means_achieves_goal`, bearing on\nthe premise).\n\n*What a consequence trial outputs.* A conditional: *if these scoped rules hold, this outcome follows*. That is\nan argument whose premises are the scoped rules, so its standing comes from the calculus already ruled (Q38,\nthe weakest premise within an argument), with no new arithmetic. The hinge score, which already asks how far a\nconclusion moves when one premise is granted or denied, ranks which assumption the outcome depends on most.\nThis is the shape of RAND's **Assumption-Based Planning** (Dewar 2002): list a plan's assumptions, find the\n*load-bearing* ones (the plan fails if they fail) and the *vulnerable* ones (likely to fail in the plan's\nlifetime), give each one that is both a **signpost** (an early observable sign that it is failing) and\nconsider **shaping** actions (make the assumption more likely to hold) and **hedging** actions (prepare for\nits failure). On the graph: load-bearing is the hinge, vulnerable is how the seven questions are answered, a\nsignpost is a testable bet in Arc AO's sense, and shaping and hedging actions are amendments to the policy,\nthemselves claims (proposal AT10).\n\n*Who does what, without bottlenecks.* The machine lists every rule a trial used, writes each scoped to the\npolicy, puts the seven questions, ranks by hinge and drafts signposts; none of it waits on a person. What it\ncannot supply is what only people placed in the setting know: whether the support factors are present here,\nhow they themselves would respond (they are the people the first question is about), and the assumption\nnobody listed, which no computation over the listed ones can find. Each enters as an ordinary contribution, and\n*what else would have to hold?* is the move offered beside every trial.\n\n**What keeps both honest.** The trial's outputs are worked consequences of a stated rule set, attributable to\nthe rules, never assertions; they are `supposed` claims and move no strength. And the rules are people's own\nwords: the trial derives from them mechanically and never has a model play a person (AN3, unruled: a lens is\nmade of holders' words). *Simulate the rule set, never the person* is the line.\n\n**Theory first, by his standing rule.** None of this is to be run before the theory of 4a is settled: what\ncounts as a case, how priorities are represented, and what the trial reports. The theory question that blocks\nthe rest is in § 6.\n\n## 5. What this offers the lenses and attunement\n\nThe worldview-lenses page already says the premises are shared and the weighting is where people diverge.\nIn Horty's terms that is exactly *the same defaults under different priorities*, and different priorities pick\ndifferent extensions. So one lens dimension becomes computable without a model playing anyone: apply a\nholder's stated priorities to the shared rule set and show the extension that results, *given your weights,\nthis follows; given theirs, that*.\n\nThe explanation question adds a dimension the lens page does not have: **what a person finds in need of\nexplaining**. Two readers looking at the same failed case, one asking *why did this go wrong?* and the other\nshrugging *that happens*, have found the place where their norms differ, and per Bear and Knobe the\ndifference may be a value hiding in what each takes as normal. Asking *would that call for an explanation?*\nis a plain question anyone can answer, it surfaces a hidden normative expectation without asking for any\ntheory, and it fits the descent (*why does that need explaining?*). Edvin's 2011 dilemma filter derived a\nlens from what people judge in cases; this question derives part of one from what they find surprising\n(proposal AT7).\n\nThe attunement it could carry is the moment the lens page calls the most powerful: *we agree on what happened;\nyou see an exception that needs a reason, I see the ordinary run of things*, said out loud, as two claims on\none rule.\n\nThe trial of § 4 and the lens of this section are one computation with two sources of suppositions,\nand a reader's what-if is a third; measured on the live graph, with the outside reading on lenses and the\nproposals, in [what-if-and-lenses.md](what-if-and-lenses.md) (2026-10-22).\n\n## 6. What is not settled\n\n- ~~How a rule's kind is stored~~ **Proposed the same night (AT2's open clause, revised): it is not stored\n  apart at all. The wording is the kind**: a universal sentence with named exceptions is normic, a sentence\n  stating a proportion is statistical. A dispute about the kind is therefore a rival wording, a new claim beside\n  the first (*hard work leads to success* against *about half of those who work hard succeed*), each open to\n  attack and each read by the arithmetic wherever it is cited. No field, no step that converts a field into a\n  claim, and nobody waits on anyone: it is *contributions are added, never overwritten* applied to the rule's\n  form. The two rival wordings are also what a lens shows (§ 5).\n- **The relevance arithmetic for both kinds** (`what-relevance-is.md` § 3f, last two notes): the rule's\n  strength × how often its outcome follows in cases like this one, which a statistical rule states and a\n  normic rule leaves to its exception claims. Proposed as AT8.\n- **Mixed rules**: how named exceptions and a residual proportion sit together on one rule.\n- **How priorities are represented** for a trial: the weighing scheme answers *which weighs more* in one\n  case, not as a standing ordering, and Horty lets priorities themselves be derived by default reasoning.\n- **Whether evidence should be weighed by kind of support** (Smith's normic against statistical evidence), and\n  Di Bello's objection that almost no trial evidence would count as normic. Related to the evidence layer,\n  untouched here.\n\n## Sources\n\nRead in full: the NDPR review of Horty, *Reasons as Defaults* (https://ndpr.nd.edu/reviews/reasons-as-defaults/).\nFrom search results and abstracts: Smith 2016 (OUP listing and reviews); Schurz 2001,\nhttps://www.cambridge.org/core/journals/philosophy-of-science/article/what-is-normal-an-evolutiontheoretic-foundation-for-normic-laws-and-their-relation-to-statistical-normality/AAF27FB7FB2787093B8197ADD2CC4EBC;\nBear and Knobe 2017, https://philpapers.org/rec/BEANPD; Hilton and Slugoski 1986 and Kahneman and Miller\n1986 via https://arxiv.org/pdf/1706.07269; Leslie on generics via\nhttps://swh.princeton.edu/~sjleslie/generics%20articulate%20default%20generalizations%20FINAL%203-5-12.pdf;\nthe normic-support debate on legal proof, https://philarchive.org/rec/DIBPPA and\nhttps://philarchive.org/rec/SILLSO; OpenFisca, https://openfisca.readthedocs.io/; Neumann 2008,\nhttps://www.jasss.org/11/4/6.html; Nair and Horty 2018, https://philarchive.org/rec/NAITLO.\nFrom memory, not checked this session: Reiter 1980, Ross 1930, Raz 1975, Rawls 1971, Ashley 1990, Lucas 1976,\nPopper and Lakatos on saving theories from refutation, Rhodes, Leslie and Tworek 2012.\n\n## Cross-references\n\n`what-relevance-is.md` § 3b, § 3f · `retrofitting-the-reasons-channel.md` § 9h–9j ·\n`philosophy-of-the-unwritten-step.md` (the rule as a licence) · `three-minds-and-the-leap.md` (AM5's\n`supposed` claims; the mind of Deliberus) · `causes-of-belief-and-the-three-minds.md` § 5 (AN3) ·\n`../worldview-lenses.md` · `weighing-scheme-and-terminus-classifier-design.md` ·\n`testable-bets/prediction-markets-resolution-practice.md` (sharpening a scope word) ·\n`what-a-rule-rests-on.md` (where a rule comes from).\n"}