{"path":"research/cases-and-generalizations-from-both-ends/literature-report-2026-11-06.md","content":"# Literature on the oscillation between cases and generalizations: a helper's report, verbatim\n\n*Written by a research helper (Claude Sonnet) on 2026-11-06 for `../cases-and-generalizations-from-both-ends.md`, kept as returned. Read-status tags are the helper's; nothing here was checked by the session beyond what the note says it checked.*\n\n---\n\nREPORT: the oscillation between cases and generalizations, what the literature says\n\nRead-status tags: [F] whole text of what the source gave me; [P] part or excerpts; [A] abstract only; [S] secondary summary only. I paraphrase throughout, with locators. The SEP entries were fetched by tool and read in extract, not offline.\n\nSHORT ANSWER\nNo source treats this oscillation as one object; five literatures treat parts of it, and they agree on three things.\n1. Alternation is guaranteed to stop only when both moves raise one shared score on a bounded space (the reflective-equilibrium model, EM). Without that score it can cycle, drift, or settle in many places.\n2. What stops the loop being merely circular is always something outside it: initial commitments, an independence constraint, a top-level prior, a protected hard core, or independent detection of a new entity.\n3. Convergence inside one head does not give convergence between heads.\nOn symmetry: formally the two directions are one inference (hierarchical Bayes). In practice they differ in speed, in who gets blamed when they clash, and in how minds err.\n\n1. REFLECTIVE EQUILIBRIUM\n- Method [F, SEP Knight 2023]: mutual adjustment of judgments and principles. Wide equilibrium adds background theories and is, in pure form, \"impossibly demanding\". Rawls himself says comparison with a few other views moves us closer to the ideal without reaching it (1971: 49-50). So the ideal is asymptotic by its author's own account.\n- Anchors [F, SEP]:\n  - Daniels's independence constraint: background theories may not recycle the judgments that support the principles, or wide equilibrium gains nothing over narrow (Daniels 1996: 22).\n  - Elgin: initial commitments carry a presumption in their favour and are not given up without reason. In True Enough (2017: 64) an account is tenable if it is, or can be reconstructed as, the outcome of adjudicating initially tenable commitments into equilibrium [S, NDPR review; I did not read the book].\n- Degeneration and emptiness [F, SEP]:\n  - Rawls (1971: 50) concedes there may be no unique equilibrium even for one person.\n  - Critics add a drifting standard (Copp 1985), several simultaneous equilibria (Haslett 1987), and order and framing effects. Schwitzgebel and Cushman 2015 found professional philosophers susceptible, with expertise not reducing it.\n  - Singer (2005: 349): the method is \"close to vacuous\" because it absorbs criticism as an internal revision.\n  - Conservatism (Brandt 1979: 20): nothing shows initial credence tracks credibility. Singer's debunking argument treats footbridge-type intuitions as evolved emotion.\n- Kelly and McGrath 2010 [A, plus passages quoted in Baumgaertner]: they ask what RE can do for someone who thinks the domain has truths not of our making. They note a good method should bring rational inquirers who follow it impeccably to converge over time (p. 341, quoted second-hand), and that RE does not guarantee this. SEP reports their worry that RE cannot exclude absurd starting judgments (their example is a \"random killing\" judgment). Lewis's passage, which they quote, says two people may both reach equilibrium impeccably while one of them is wrong about the facts.\n- Formal model, Beisbart, Betz and Brun 2021, Ergo 8/15 [P]:\n  - Two alternating rules, revise theory and revise commitments, run until a fixed point. A score (faithfulness, account, systematicity) never decreases.\n  - Because the state space is finite, the process always ends at a fixed point. Every fixed point is a global optimum only for extreme weights, which amounts to ignoring one side.\n  - Built-in asymmetry: theory revision is free, while commitment revision is tethered to the initial commitments.\n  - Betz's summary says simulations \"suggest\" it usually reaches a full RE state. I did not see the simulation numbers.\n- Baumgaertner, \"Convergence and Shared Reflective Equilibrium\", Ergo (year not confirmed) [P]:\n  - Even when every agent reaches individual equilibrium, it is \"surprisingly difficult\" for a community to converge on a few equilibria.\n  - More intuitions mean more choice points and more divergence.\n  - A trade-off: either individuals may never converge (they may ignore an intuition indefinitely), or individual convergence is forced and shared convergence suffers.\n\n2. LAKATOS\n- Proofs and Refutations [P, SEP]:\n  - A counterexample can be global but not local, local but not global, or both.\n  - A proof is a thought-experiment that splits the conjecture into sub-conjectures, so a refutation targets a lemma. Incorporating the refuted lemma into the conclusion yields \"proof-generated\" definitions.\n  - The labels monster-barring, monster-adjustment and exception-barring come from course materials [S]. SEP does not use them.\n- Programmes [F, SEP, citing FMSRP: 34]:\n  - A programme is theoretically progressive if each new theory has excess empirical content over its predecessor, and empirically progressive if some of that excess is corroborated. Otherwise it is degenerating, and predicting what everyone already knows earns nothing.\n  - When a prediction fails, logic leaves open whether to blame core or auxiliaries. The negative heuristic says protect the core and adjust the auxiliary belt. The choice of what to revise is made by the programme, not by logic.\n- The real Euler record, with dates:\n  - Euler to Goldbach, 14 Nov 1750 [P, Richeson intro]. In Elementa doctrinae solidorum (E-230) he says he could not find a firm proof and checks the Platonic solids [F, Sandifer 2004]. Lhuilier's gloss on Euler's first paper: the conclusion was offered as probable, \"founded on analogy only\", from solids of several kinds [F, Gergonne's extract of Lhuilier]. That is the cases-to-generalization step, flagged as weak at the time.\n  - The E-231 proof (written 1751, printed 1758) is not rigorous by modern standards [S, Daems review of Richeson]. Legendre's proof is dated 1794 [S, Swedish thesis; Daems]. Cauchy's proof is cited by Gergonne as Jan 1811.\n  - Lhuilier 1812-13, Annales de mathématiques vol. 3, pp. 169-189 [F]: the theorem \"suffers numerous exceptions\" and holds generally only for polyhedra without re-entrant parts. He lists three kinds: an interior cavity (F+S = A+4), a through-opening (F+S = A for one opening), and annular faces. Gergonne says he himself had noticed the first two long before, and that Lhuilier was first on the third.\n  - Star polyhedra [S, Wikipedia]:\n    - Kepler recognised two around 1619. Poinsot rediscovered them in 1809 and added two more. Cauchy proved the list complete in 1813.\n    - Cayley named them in 1859 and gave a density-corrected formula, dvV − E + dfF = 2D. Two of the four have χ = −6, and two have χ = 2.\n    - Schläfli insisted χ = 2 must hold and rejected two of them as polyhedra, a view never widely held. That is real monster-barring.\n- Fidelity to Lakatos: SEP calls the dialogue a rationally reconstructed history, with the real history in the footnotes. Sandifer says Euler could not have known most of Lakatos's hard examples, and that \"Euler was wrong / had no proof / had a wrong proof\" are half-truths [F]. A historiography paper argues discovery is not immanent logic but entangled with power relations [A]. What the record shows is about a century of work by many hands. No final equilibrium was reached; the object changed (Richeson's subtitle: birth of topology).\n\n3. HIERARCHICAL BAYES, THE FORMAL TWIN\n- Kemp, Perfors and Tenenbaum 2007, Dev. Science 10(3) [P]:\n  - Cases inform an overhypothesis, and the overhypothesis informs each new case. Their marble-bags example: one draw from a new bag says little alone, but with earlier bags it supports \"all marbles in this bag are black\".\n  - Depending on task and data, abstract knowledge can be learned faster than, equally fast as, or slower than concrete knowledge.\n  - Their own caveat: \"we do not claim that overhypotheses can be generated out of thin air\". Every such model assumes how each level is generated from the one above, and a given top-level prior.\n- Goodman, Ullman and Tenenbaum 2011, Psych. Review, \"Learning a theory of causality\" [P, results excerpts]:\n  - The abstract theory was learned before most specific structures. A learner who had to learn the theory matched a learner given the correct theory at about 60 samples per system (50 systems, 691 candidate theories, 542 structures). Most runs learned the theory in under 30 samples per system.\n  - Breadth beats depth: with 5 systems learning was very slow, while with 500 systems it caught up after about three samples per system. Their phrase for this is the \"blessing of abstraction\".\n- Convergence: Bayesian induction converges in the limit; the real dependence is on the top prior and on the data being drawn from many systems.\n\n4. ALTERNATING OPTIMIZATION: EM\nWu 1983, Ann. Statist. 11(1) [P, first pages]. The likelihood never decreases across EM steps (Dempster, Laird and Rubin 1977, Theorem 1), so it converges to some value. There is no guarantee that value is the global maximum, and even convergence to a local maximum cannot be settled in general. Dempster, Laird and Rubin's proof that the parameter sequence itself converges contained a triangle-inequality error. Wu proves, under conditions, that limit points are stationary points: for a curved exponential family with compact parameter space, and, if the likelihood is unimodal and suitably differentiable, convergence to the unique maximum. The stopping rule in practice is a tolerance, a convention. The parallel is exact: monotone progress on one score gives convergence to something, and nothing guarantees it is the best thing.\n\n5. PEIRCE AND DEWEY\n- Peirce [P, SEP]: abduction proposes a testable explanation, deduction draws out expected consequences, induction is the whole process of testing. Method is self-correcting, so inquirers starting anywhere converge. But at any moment we cannot tell how far we are from the limit. The economy of research (cost against benefit of the next inquiry) paces it.\n- Dewey [A, SEP; S, Pate's excerpts of Logic 1938]: five phases (felt difficulty, problem, hypothesis, reasoning out its bearings, test). Reasoning often sends you back to an earlier phase. Closure is relative to the problem in hand: what counts as essential is whatever the inquiry needs. Earlier conclusions are reused unexamined, which is the practical stopping rule.\n\n6. KUHN AND THE HERMENEUTIC CIRCLE\n- Postscript 1969 [P]: exemplars, the concrete problem-solutions students meet first (inclined plane, conical pendulum, Keplerian orbits), come before and teach the application of the symbolic generalizations. A bare statement of the law is nearly impotent without them (p. 191, via a seminar transcript [S]). This is a developmental asymmetry: cases are learned first, as practice.\n- Hermeneutic circle [P, SEP]: whole understood through parts and parts through whole. Understanding goes deeper rather than higher. Schleiermacher calls the interpretive task infinite and the circle not vicious.\n\n7. CASES WITH GROUND TRUTH\n- Mendeleev (periodic law).\n  - Predictions: eka-aluminium, eka-boron and eka-silicon, found as gallium (1875), scandium (1879) and germanium (1886) [P, Stewart 2018]. Stewart counts only three complete predictions.\n  - Scerri 2019 [A]: at best about half of his predictions came out right.\n  - Scerri and Worrall 2001 [A+intro]: little support for the standard story that these predictive successes made the case, because accommodation, notably argon, mattered equally.\n  - Exceptions: he put tellurium before iodine against atomic weights, claiming one weight was wrong. The weights were right and the order was right; atomic number (Moseley, 1913) explained why [S, ACS and RSC pages].\n  - Argon (1894): first resisted as triatomic nitrogen, then given its own new group in 1900 [S, ACS]. The case first met a \"not really an element\" move, then forced a new column.\n- Kepler. A Mars residual of 8 arcminutes against circle-based models, measured by Tycho, led him to drop circles. Tycho's best accuracy was about 2 arcminutes [S, MacTutor]; Copernicus never hoped to do better than about 10 [S, lecture notes]. Laws 1 and 2 are in Astronomia nova (1609) and law 3 in Harmonices mundi (1619). The residual exceeded the error bar, so it could not be barred as noise; the generalization was changed (an axiom, circularity, was abandoned).\n- Neutrino. Pauli's letter is dated 4 Dec 1930, and he called the hypothesis a desperate remedy [P]. The alternative on the table was giving up energy conservation at nuclear scale. Fermi built a theory on it; Reines and Cowan cabled detection on 14 June 1956 [P]. A new entity was posited to save the law, and it was later independently detected.\n- Mercury. Le Verrier's 1859 study left an unexplained residual. The Uranus-to-Neptune precedent suggested an inner planet (Vulcan). By the 1890s that was dropped, and Einstein's paper of 25 Nov 1915 accounted for 43 arcseconds per century [P, Wells arXiv; A, Einstein Papers; sources disagree on Le Verrier's own figure, so I give none]. Lakatos judges the astronomers right not to abandon Newton earlier [P, SEP].\n- Pattern: \"posit a new entity\" won in the neutrino case and lost in the Mercury case. Nothing known at the time distinguished them. What did was independent detection, or its repeated failure over decades. Observation error bars decided Kepler's case.\n\n8. ARE THE TWO DIRECTIONS SYMMETRIC?\n- Formal: in hierarchical Bayes the two directions are conditionals of one joint distribution, so symmetric in principle. They differ in speed (abstract often learns first), and the model still needs an outside top prior.\n- Procedural asymmetries:\n  - RE: commitments tethered, theory free (Beisbart et al.).\n  - Lakatos: blame goes to the auxiliary belt first.\n  - Kuhn: cases are learned first as exemplars.\n- Psychological:\n  - Tversky and Kahneman 1971 [A+first page]: people treat a small random sample as highly representative. Their questionnaire of 84 psychologists, on a result with z = 2.23, gave a median of 0.85 for the chance that a further 10 subjects replicate it, while 0.48 would be reasonable; only 9 answered between 0.40 and 0.60.\n  - Hamill, Wilson and Nisbett 1980 [A]: in two experiments a single vivid case shifted attitudes toward the whole population (welfare recipients, prison guards) even when subjects were told the case was highly atypical.\n  - Kahneman and Tversky 1973 [A]: intuitive predictions follow representativeness and are insensitive to prior probability and to the reliability of evidence.\n  - Nisbett et al. 1987 [A]: brief formal training in inferential rules improves everyday reasoning. People already use the law of large numbers when the sampling process is clear (Nisbett et al. 1983, as described in Lehman et al. 1988 [P]).\n- My reading, not a finding of the literature: errors run in both directions, with one common cause, a salient case standing in for its class and the class for the case. I found no source claiming that one direction is uniformly worse. The corrective pattern differs by direction: abstract rules can be trained; correction of the upward error needs the rule plus the full distribution.\n\nNot read in full: Lakatos, Rawls 1951/1971, Daniels, Elgin 2017, Dempster-Laird-Rubin, Kelly and McGrath (full text), and the full text of Kemp et al. and Goodman et al. Wikipedia and course pages were used only for dates and counts.\n\nWHAT WOULD MOST CHANGE YOUR PICTURE\n1. Convergence is a property of the score, not of the oscillation. Every proof of termination (RE model, EM) rests on one shared objective on a bounded space. Lakatos's progressive/degenerating test is the one criterion in the literature that works on real, unbounded cases, and it needs a dated order (the generalization's wording fixed before the case that confirms it). So the graph would need to record when a case was filed relative to the generalization's wording.\n2. The shared version is harder than the private one. Same starting point and flawless method still diverge as choice points accumulate (Baumgaertner), and Kelly and McGrath say the same of non-convergence. A shared graph should therefore record the choice points themselves (which case was barred, which exception adopted, which entity posited), not promise convergence.\n3. Every loop that works has an outside anchor, and breadth helps more than depth. Elgin's initial commitments, Daniels's independence constraint, the Bayesian top prior, Lakatos's core, and independent detection in the neutrino and Mercury cases all play that role. Goodman et al. found evidence across many systems beats many cases from one system. The graph should therefore prefer cases from different sources over more cases from the same one.\n"}