{"id":"e6cb04a9-b033-40a5-8941-5f6c1bec6160","arxiv_id":"2510.07478","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A new stochastic intergenerational model shows that identical groups reach parity under purely meritocratic admission, but a tiny affinity advantage can turn random fluctuations into permanent group dominance.","lead":"Using a simple two-group model of meritocratic college admissions, the paper shows that when both groups face identical rules, gaps between them vanish in expectation, but adding even a tiny 'affinity advantage' to the currently leading group makes gaps permanent and self-reinforcing. The qualitative message for fairness practice is that static equal-selection audits cannot detect long-run feedback inequities.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Affinity advantage is a step function of the group lead; the entrenchment result may be an artifact of this discontinuity rather than of a small continuous asymmetry.","rationale":"The reader's weakest assumption already identifies the affinity feedback and its uniformity across lead sizes and group members. I agree, and I sharpen the point: the formal issue is not just that the boost is uniform but that it is a discontinuous function of the sign of the lead. The appendices are mostly self-contained and the algebra behind Theorems 2 and 4 appears internally consistent, so I do not raise a proof-error objection. The concern is about whether the paper's headline conclusion—'even a slight asymmetry' permanently entrenches stochastic disparities—is robust to a minimal, arguably more realistic, continuous version of the same feedback. This is exactly the kind of modeling choice that should be stress-tested before the policy-relevant conclusion is accepted. The reader's CONDITIONAL verdict already incorporates this uncertainty, so I recommend keeping it unchanged; if the continuous-feedback test fails to reproduce persistent separation, the manuscript would need to either weaken its claims or provide a microfoundation for the step function.","tokens_in":32208,"tokens_out":9638,"duration_ms":84514,"concrete_test":"Replace the step-function affinity in Model 2 with a continuous lead-dependent boost, e.g. ε·(x_A−x_B)_+ or ε·x_A/(x_A+x_B), and either re-derive the fixed points or simulate the stochastic process with the same α, p, q, N. For small ε, check whether the parity state (αp, αp) is locally stable and whether any separated stationary distribution persists. If persistent separation vanishes for ε below some threshold, the entrenchment result is an artifact of the discontinuity; if separation persists, the concern is settled and the core claim survives.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central entrenchment claim (Theorem 4) is driven by Model 2's step-function affinity rule: whenever x_A > x_B, every non-admitted member of A receives a fixed probability ε of becoming high, no matter how small the lead is, while if x_A = x_B no one gets ε. This makes the expected map T discontinuous on the diagonal and introduces an unstable parity fixed point (αp, αp) in the AA model alongside the asymmetric stable fixed points. The theorem's 'unique fixed point' statements are restricted to starting points with x > y, so they do not confront the tie branch. The paper's language that a 'slight asymmetry' yields permanent exclusion (x_B = 0) conflates a small ε with a continuous perturbation; the actual mechanism is a hard sign threshold that turns on a fixed-size boost for the entire leading group. If real network effects scale with the size or quality of the lead (e.g., number of high-type contacts or fraction of high types in the group), the separated fixed point can disappear or parity can become stable for small ε. The closed-form fixed point x_A = 2α(p−ε)+ε, x_B = 0 is correct given the rule, but the rule itself is the load-bearing assumption.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies a two-group intergenerational model of meritocratic selection into a scarce 'college'. In the Equal Advantage (EA) model, admitted students become high with probability p, rejected high types persist with probability q, and rejected low types never upgrade; the expected dynamics are shown to have a unique fixed point (αp, αp) reached from any initial state, so fair-at-each-step selection yields long-run parity, while stochastic separation arises but dissipates. In the Affinity Advantage (AA) model, all non-admitted members of the leading group receive an extra probability ε of high status; the paper claims a unique asymmetric fixed point for every ε>0 when starting from a lead, with the closed form x_A=2α(p−ε)+ε, x_B=0 for ε above the threshold ε̃=2α(1−p)/(1−2α). A continuous-ability simulation model is used to argue that the qualitative pattern persists.","tokens_in":32584,"tokens_out":15269,"duration_ms":120514,"significance":"If the results hold as stated, the paper makes a useful conceptual point: static fairness criteria can fail in the long run, and stochastic shocks can become entrenched via group-level feedback. The EA analysis is transparent and parsimonious, with no fitted parameters; the AA threshold and the exclusion fixed point are striking, falsifiable predictions. The main intellectual risk is that the entrenchment result is generated by a discontinuous step-function feedback, so the 'even a slight asymmetry' framing overstates the continuity of the modeled mechanism. The simulations corroborate the qualitative pattern but use the same step rule.","major_comments":[{"comment":"The prose defining the two regimes is reversed. It states 'When capacity is limited (i.e., N X(t)<C)' for the rule A_i=X_i/X·2Nα, but if total high types are below capacity all high types are admitted and residual seats are filled by low types—that is the under-subscribed regime, whose formula is the one given under 'capacity abundant'. The Appendix (Theorem 1 proof) and Theorems 3/4 use the opposite convention (over-subscribed iff x_A+x_B≥2α). The Section 2 inequalities must be corrected so the model definition agrees with the rest of the paper.","section":"Section 2, 'Selection Rule'"},{"comment":"The theorem states 'the system has a unique fixed point' for each range of ε, but in the actual symmetric Model 2 the tie state (αp, αp) is also a fixed point of T: when x_A=x_B neither group receives ε, and the EA under-subscribed transition maps (αp, αp) to itself. The 'WLOG group A has the affinity advantage' is valid only on the open half-space x_A>x_B; it does not cover the diagonal. The uniqueness claim should be restricted to x>y (with the mirror claim for x<y), or the diagonal branch should be analyzed explicitly. As written, the word 'unique' is mathematically false.","section":"Section 3.2, Theorem 4"},{"comment":"The proof of uniqueness for ε<ε̃ is incomplete. The argument introduces the quadratic g(y)=y^2−(2+2αp)y+4αp+2(1−α)(1−x_A)ε, whose coefficients depend on the unknown x_A. Showing that only the smaller root of g can lie below 2α does not show that two fixed points with different (x_A,y) cannot both yield their own smaller root. The sentence 'once x_A+x_B is determined uniquely, x_A and x_B are also characterized uniquely' is asserted, not demonstrated. Please give a self-contained uniqueness proof (e.g., derive a single equation in y=x_A+x_B by substituting the expression for x_A from the individual fixed-point equations) or state the result as existence plus bounds.","section":"Appendix A.6, Lemma 3"},{"comment":"The persistent-disparity conclusions are generated by a step-function affinity advantage: ε is applied to all non-admits of the leading group whenever x_A>x_B, regardless of the size or quality of the lead. The richer simulation in Section 4 uses the same step rule ('if one group has more members admitted'), so it does not test whether entrenchment survives a continuous, saturating, or lead-size-dependent affinity. The Discussion's assertion that 'any positive group-level feedback pushes the leading group further ahead' is plausible but is not established by the simulations. A continuous-affinity variant would substantially strengthen the robustness claim; at minimum, the 'even a slight asymmetry' language should be qualified.","section":"Sections 4 and 5"}],"minor_comments":[{"comment":"Typo: 'network efforts' should be 'network effects'.","section":"Section 2, Model 2"},{"comment":"The contradiction step contains a typo: the line '2α−2αx_B<2−2x_B' should be '2α−2αx_B<2α−2x_B'. As written, the subsequent implication is garbled.","section":"Appendix A.6, Lemma 1, Subcase b"},{"comment":"The notation δ0 is introduced for the allowed failure probability but the theorem statement uses ω. Please make the notation consistent.","section":"Appendix A.5, Theorem 5 proof"},{"comment":"The reference 'Pilat and Krastev' is incomplete (no title or publication venue).","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper's EA analysis is clean and the AA threshold calculation is a nice theoretical contribution, but Theorem 4's uniqueness claim is overstated and the proof of Lemma 3 needs repair. These are fixable within the manuscript's scope, but they are load-bearing for the central statements."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things up front. First, the core math holds up: the EA model has a unique parity fixed point (alpha p, alpha p) reached from anywhere, and the AA threshold analysis is internally consistent as far as I checked. Second, the paper's claim that \"even a slight asymmetry\" causes permanent exclusion is doing more work than the model actually supports. The affinity rule is a step function -- if x_A > x_B, every non-admitted member of A gets a fixed epsilon boost, no matter how small the lead. That discontinuity, not the smallness of epsilon, is what makes parity unstable and drives the separation result.\n\nWhat is genuinely new: most dynamic fairness work starts from pre-existing asymmetry or imposes statistical constraints. Here, the two groups are identical, the allocation rule is meritocratic, efficient, and equalizes selection rates (Theorem 1 uniqueness is a nice touch), and stochastic shocks are the only source of divergence. Theorem 2 is straightforward and clean. Theorem 4's threshold epsilon_tilde = 2 alpha (1-p)/(1-2 alpha) and the closed-form fixed point (2 alpha (p - epsilon) + epsilon, 0) give a concrete, falsifiable characterization. The appendix proofs are largely complete, and the richer simulations match the qualitative story. No code is shipped, but no constants are fitted to data; the results are derived from explicit transition rules.\n\nThe soft spots are real but addressable. The main one is Model 2 itself: the affinity boost is a uniform, fixed-size effect for the entire leading group, not scaled by the size of the lead, not saturated, not dependent on actual high-type contacts. The map T is discontinuous on the diagonal, so the theorem statements carefully restrict to starting points with x > y and do not confront the tie branch. Calling this a \"slight asymmetry\" is misleading -- the mechanism is a hard sign threshold. If real network effects are graded or saturate, the separation result may not survive, and the paper only pays lip service to this in the discussion. There is also a clear typo in Section 2: the over/under-subscribed regime conditions are swapped (N X(t) < C is labeled over-subscribed when it should be under-subscribed). The appendix uses the correct definitions, so it is fixable but will trip up a reader. Theorem 5's definition of T_delta is loose; the proof uses one-step separation, which is a different event than the absolute separation stated in the theorem.\n\nWho this is for: researchers in dynamic fairness, feedback loops, and intergenerational mobility who want a stylized, provable baseline result. Treat the entrenchment result as a proof-of-concept under a specific affinity rule, not as a generic consequence of arbitrarily small continuous asymmetry. I would send this to peer review with a request to fix the Section 2 typo and to add an explicit discussion of the step-function assumption and its robustness to lead-dependent or saturating feedback. That is work the authors can do, and the EA parity theorem plus the AA threshold are worth having on record.","headline":"A clean theoretical result with a heavy modeling assumption: the entrenchment conclusion is real but depends on a step-function affinity boost, not on the size of epsilon.","tokens_in":32997,"tokens_out":3240,"would_cite":true,"duration_ms":31514,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proves that a fair, meritocratic admission rule produces permanent, self-reinforcing inequality once the leading group receives even a tiny affinity boost.","keywords":["fairness","meritocratic selection","intergenerational mobility","feedback loops","affinity advantage","fixed points","stochastic dynamics","college admissions"],"falsifier":"Measure the actual transition rate of non-admitted members of the leading group as a function of the lead size: the model assumes a constant ε even when the lead is tiny. If the measured boost vanishes as x_A−x_B approaches zero, or declines with group size N, then the threshold ε̃ overstates the regime of permanent exclusion and the system will not converge to (2α(p−ε)+ε, 0).","tokens_in":32130,"feed_emoji":"🎓","tokens_out":6901,"duration_ms":57886,"temperature":0.7,"pith_summary":"This paper studies whether a college-admission rule that is fair at every step—always prioritizing qualified candidates while keeping selection rates equal across groups—stays fair over generations. It models two identical groups and proves that without group-level feedback, the expected dynamics converge from any starting point to a unique fixed point where both groups have the same high-type fraction, αp; temporary gaps caused by randomness decay. The central finding is that introducing even a tiny 'affinity advantage' (a fixed extra probability ε for non-admitted members of the group that currently has more high types) destroys this parity: for every ε>0 there is a separating fixed point, and for ε above a threshold the trailing group is entirely excluded from college. This matters because it shows that static fairness metrics do not guarantee long-run fairness, and that pure chance can harden into permanent stratification.","feed_headline":"Tiny affinity edge makes fair admissions permanently unequal","feed_subtitle":"Even from identical starts, a small network effect for the leading group can lock in disparity for good.","key_machinery":"The expected-dynamics map T(x,y) = (E[X_A(t+1)|x,y], E[X_B(t+1)|x,y]) and its fixed points, which characterize long-run average outcomes for each group. The analysis splits into over-subscribed and under-subscribed regimes and hinges on the affinity parameter ε: the threshold ε̃ = 2α(1-p)/(1-2α), where α is the college capacity fraction and p is the probability that admission makes an individual high-type, marks the boundary between full exclusion of the trailing group and a partially separated equilibrium with a closed-form lower bound on the gap.","core_discovery":"Under the Equal Advantage model, the expected dynamics have a unique fixed point (αp, αp) that is reached from any starting state, so a selection rule that is meritocratic and fair at each step restores parity on average. Under the Affinity Advantage model, the unique fixed point has x_A > x_B for every ε>0; when ε ≥ ε̃ = 2α(1-p)/(1-2α), the fixed point is exactly (2α(p-ε)+ε, 0), meaning the leading group holds all high types and the trailing group is completely absent from the admitted population. The paper derives lower bounds on separation for smaller ε and verifies in a continuous-ability simulation that the same qualitative behavior persists.","pith_inferences":["If real affinity effects depend on the size of the lead or require direct contact with high-type individuals, the constant-ε threshold is likely to soften: inequality may persist only above a minimal lead size, a prediction testable in the richer simulation by setting ε proportional to x_A−x_B.","The model identifies a concrete intervention target: reducing ε (the group-level network boost) below the threshold avoids complete exclusion, while capacity expansion α can shrink but not eliminate separation—suggesting anti-network policies may matter more than quotas.","In high-stakes, small-cohort settings (fellowships, internships, youth athletics), deliberate randomization of scarce slots could stop the initial stochastic lead from forming, since the feedback loop amplifies any early gap.","The sharp threshold ε̃ = 2α(1-p)/(1-2α) gives designers a quantitative calibration target, though the paper's lower bounds leave the intermediate-ε regime only partially characterized."],"forward_implications":["Static fairness is not enough: a rule that is fair at every step can still produce permanently unequal outcomes once group-level affinity feedback exists.","Any positive ε guarantees a long-run separation under the Affinity Advantage model; for ε at or above the threshold, the trailing group is completely excluded from the program.","Under Equal Advantage, parity returns on average, but the time to recover grows sharply as p approaches 1, so more effective programs take longer to erase accumulated advantage.","In small populations, stochasticity alone can create large temporary disparities from symmetric starts, and these are worse at low capacity α.","The continuous-ability simulation reproduces the model's qualitative predictions, indicating the results do not depend on the binary-type simplification."],"fun_headline_variants":["Tiny edge in admissions locks in permanent inequality","Fair meritocracy? Not with a slight affinity bias","One epsilon: how a small preference breaks fairness","When fair admissions still yield permanent gaps"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The persistence result rests on the assumption that every non-admitted member of the leading group receives the same fixed probability ε of becoming high-type, no matter how small the lead, how large the group, or whether the individual has any contact with high types; if the real boost is weaker, saturates, or requires direct contact, the separating fixed point need not exist.","fun_headline_variants_meta":{"raw":{"variants":["Tiny edge in admissions locks in permanent inequality","Fair meritocracy? Not with a slight affinity bias","One epsilon: how a small preference breaks fairness","When fair admissions still yield permanent gaps"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00048,"raw_usage":{"total_tokens":2241,"prompt_tokens":799,"completion_tokens":1442,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":543,"completion_tokens_details":{"reasoning_tokens":1396}},"tokens_in":543,"tokens_out":1442,"duration_ms":8011,"temperature":1.0,"reasoning_tokens":1396,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T10:56:22.453999+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the actual transition rate of non-admitted members of the leading group as a function of the lead size: the model assumes a constant ε even when the lead is tiny. If the measured boost vanishes as x_A−x_B approaches zero, or declines with group size N, then the threshold ε̃ overstates the regime of permanent exclusion and the system will not converge to (2α(p−ε)+ε, 0).","supporting_citations":[],"review_version":1}