Analytic memos (46)

Full-Stack Alignment is the field consolidating around three of the dissertation's lit-review themes: T2 (the bundling argument = the preferentism critique as manifesto), T3-adjacent (institutions must be co-aligned - the Ferretti/Gabriel-2022 institutional turn), and the thick-values thread (VC-THICK now has its flagship). CRITICAL DIFFERENTIATION: TMV's neutrality posture - 'constraints on what can count as a value while remaining open about which values any person or community should endorse' - is proceduralism relocated to the REPRESENTATION layer: it constrains the FORM of values while declining all first-order normative commitments. This inherits the normative-regress problem of T3: someone must still decide which well-formed values govern when they conflict, and grammar-like constraints cannot adjudicate substantive conflicts. The dissertation's Rossian-convergentist layer answers exactly the question TMV brackets. Positioning sentence for the lit review: TMV supplies the data model the dissertation's corpus already implements (reason-annotated thick values); the dissertation supplies the adjudication theory TMV still lacks.

The Gabriel arc is now fully coded: 2020 (fair principle-selection for alignment) -> 2022 (AI in the basic structure; justice norms apply to deployment) -> 2024 STELA (empirical elicitation) -> 2025 G&K (claims-based deliberation) -> 2026 Fischli (agentic autonomy). The 2022 essay supplies what the 2020 paper lacked - a POLITICAL-INSTITUTIONAL home for alignment norms - and is the direct ancestor of Ferretti's institutional-change thesis. For the governance chapter: the worst-off priority (unit 3) gives the Immigration chapter a Rawlsian test with teeth - immigration AI patently affects some of the globally worst-off, so on Gabriel's own standard it faces the strictest justification demands, yet (per G&K unit 14) its subjects are excluded from the justificatory community. The tension between Gabriel-2022's standard and G&K-2025's scoping is a publishable observation.

Coverage citation: value alignment now has reference-work codification in the IS/HCI literature (with value-sensitive design as its native method) - useful in the lit review's field-mapping paragraph alongside McKinlay (CS), Triantafyllopoulos (advisory), and the Phil Studies cluster (philosophy) to show the concept's cross-disciplinary institutionalization.

One-cite: empirical behavioral support for Baum's graded X,Y-alignment definitions (alignment as monotone degree) and for Peterson & Gärdenfors' distance-weighted misalignment measure - human trust responds to alignment DISTANCE, validating degree-based over binary conceptions. Relevant to the folk corpus's trust discourse coding.

Double use: (a) methodological neighbor - comparing LLM output distributions to human population distributions is exactly what the folk corpus enables at scale and with reasons attached (their 3 Likert samples vs the corpus's 366k coded comments); (b) cautionary instance - treating sentiment-distribution match as 'alignment' is the benchmark fallacy LaCroix diagnoses (sociological concordance ≠ ethicality), so cite it as both precedent and foil. The cross-day instability finding adds a minor seventh datum to the LLM-evaluative-instability pile.

Governance-chapter one-cite: the maximal institutional answer to McKinlay's testing-fragmentation gap and H&D's third-party-assessment recommendation - but note MELO's undecidability result bounds what any testing regime (however large) can certify for arbitrary models, and S&K's legitimacy analysis would ask who governs the certifier. Useful as the pole position in a spectrum of institutional designs.

One-cite support for STEINGRUEBER_BAUM unit 5's claim that AI's normative constraints exceed morality (moral + legal + social + now aesthetic reasons) - the cross-domain weighing problem is broader than ethics, which strengthens the case for a general reasons-weighing (Rossian-style) architecture over any single-domain solution.

Two uses: (a) another independent 'principles lists lack a coherence/justification layer' argument (with Jobin/Mittelstadt lineage), here solved by a virtue-like meta-condition on INSTITUTIONAL actors rather than systems - convergent with Brophy's process-virtues and the sincerity/faithfulness thread (S&K's wise judge); (b) cross-cultural datapoint: a Japanese team's framework built on a concept (sincerity/makoto) with Confucian-Japanese ethical resonance - modest but real evidence for the non-Western-frameworks thread alongside the Ubuntu concessions.

Adds the responsibility register to the shallowness diagnosis: where MILLIERE shows shallow alignment FAILS (jailbreaks), Akyol argues it WRONGS (deceives stakeholders about the system's ethical constitution, violating epistemic rights and implicating designers). Connects Baum's execution-vs-outcome scope to the moral-architecture level, and gives the Health chapter language for why a merely-filtered AI Scribe wrongs clinicians who rely on its apparent alignment. One-cite use.

Adds the enactivist/autopoietic tradition to the anti-agency census (now five traditions) and states the deepest version of the alignment dilemma: autonomy and external control are CONCEPTUALLY in tension, so 'aligned autonomous agent' verges on oxymoron - the abstract ancestor of H&D's misalignment/misuse tradeoff and Zhi-Xuan's tool-like proposal. His tools-not-agents prescription converges with Zhi-Xuan's engineered incompleteness; both collide with the economic reality H&D note (incentives favor agents). For the dissertation: cite as the limiting position on the agency spectrum, against which the middle positions (constrained agents + distributed responsibility) are defined.

Coverage citation only: shows the engineering community organizing 'friendliness' as XAI+privacy+fairness+affect - a checklist decomposition that exemplifies exactly what Mittelstadt/Gabriel call placeholder principles (no account of trade-offs between the four, no normative theory beneath them). Useful in the lit review as the contrast case: what alignment discourse looks like without philosophical machinery.

One-cite counterpoint: Noller argues AMCs lack the affectivity virtue requires; Sennesh & Ramstead sketch what engineering affect-grounded valuation would even mean (taxis/active-inference). If their program succeeded, the virtue-ethical disanalogy would narrow - worth a footnote in the moral-agency discussion acknowledging the naturalization route, while noting valence-navigation is far from phronesis.

The ML community's own arrival at the institutionalist conclusion: alignment requires incentive/institution design, not just training (converges with Ferretti's institutional-change thesis, H&D's governance-as-uniform-improvement, S&B's monopoly conditions). Contract theory as an alignment tool also loops back to Hadfield-Menell's incomplete-contracting line. One-cite use in the governance chapter: even the technical roadmap literature now says alignment is incentive-institutional.

Unit 5 ('fallacious to suppose that because most people do reason this way, AI systems ought to') completes the is/ought census across the library: LaCroix 2022, Gabriel 2020 (unit 7), IEAI 2026 (unit 13), Lundgren 2026 (unit 4), McKinlay's normative-gap finding, STELA's expert-corrective concession, S&K's systematic-error argument. SEVEN coded sources now state Howard's worry as the field's own. The lit-review motivation section should present this as the field's most-repeated self-criticism - and then observe that every source states the gap and none closes it (LaCroix comes closest by demanding the metaethics be done, unit 4 - which is what the dissertation's Rossian-convergentist chapter does). Also biographical: this is LaCroix's methodological prelude to his 2025 monograph - engaging it engages the field's current introducer on his own ground.

Noller completes the framework sweep: the anti-moral-agency conclusion has now been derived from KANTIAN premises (Sanwoolu - no autonomy/self-legislation), BEHAVIORAL-ANALYTIC premises (Baum, Millière - external standards, dispositions), CONTINENTAL premises (Josifović & Noller - extension not subject), and now VIRTUE-ETHICAL premises (no phronesis, affectivity, embodiment - unit 2). Every major normative tradition, applied carefully to current AI, denies moral agency while affirming normative evaluability. For the Luke debate this is decisive framing: Luke's rationality-implies-responsibility inference is rejected by every tradition, each for its own reasons - and Augustine's experiment tests the shared empirical presupposition (no stable practical reasoning) that all four rejections rely on. Also note 'media of moralisation' (unit 4) as the best single phrase for the position.

The 'AI safety as sociotechnical discipline' conclusion (unit 6) is the systems-engineering twin of BROPHY's externalized-MWRE and JOSIFOVIC_NOLLER's normative interface: all three relocate the normativity from inside the artifact to the surrounding institutional design. The lit review now has SIX sources for the 'normativity lives in the sociotechnical system' consensus - which makes the field's continued silence on responsibility-DISTRIBUTION within that system (who, qua what role, owes what) all the more striking as the dissertation's opening.

Unit 5 imports the critical-algorithm-studies canon (O'Neil, Noble, Benjamin, Bender's stochastic-parrots) into the alignment conversation - the only coded source that bridges to that literature. Useful for two purposes: (a) the lit review can use J&N as the connector node between the alignment debate and the algorithmic-injustice literature the department will expect a technology-ethics dissertation to know; (b) the Bender point ('dominant linguistic patterns rather than reasoned normative commitments') is yet another formulation of the descriptive/normative gap - LLMs reproduce the DISTRIBUTION of moral talk, not its justificatory structure - which is precisely what the folk corpus's reasoning-field coding is designed to recover.

Independent convergence worth naming in the lit review: Josifović & Noller's 'integration into norm-governed action spaces, humans as ultimate responsibility-bearers' arrives at the same destination as SANWOOLU (constrained-not-accountable), ZHIXUAN (role-specific norms), BROPHY (externalized deliberation), and BAUM (X-ethicality; tools owned by human moral agents) - from a continental/Kantian route rather than analytic or ML-adjacent ones. FIVE independent research lines now converge on: (a) AI is not a moral agent, (b) normative evaluability doesn't require it to be, (c) responsibility stays with humans/institutions structurally. The dissertation can announce this as an emerging cross-traditional consensus - and then note what NONE of the five provides: an account of HOW responsibility distributes among the human bearers (which human? qua what role?), which is exactly where Kästner's difference-making + Dignum's framework + the dissertation's RL-taxonomy enter.

Units 6-8 give the responsibility strand its cleanest coercion analysis: the CONSTRAINT-DEFINER is the primary coercer, the AI the means, and actual coercion depends on exit options/monopoly. This triangulates with HELLRIGEL_DUNG (access determines misuse capability), KAESTNER (epistemic access determines liability), and ZHIXUAN unit 14 (tyranny of creator values): across four sources, responsibility/legitimacy tracks WHO CONTROLS THE CONSTRAINTS + WHAT ALTERNATIVES EXIST. For the governance chapter: alignment-as-coercion means constraint-definition is an exercise of power requiring justification, and monopoly conditions (unit 8) convert design choices into coercion - directly applicable to AI Scribe (hospital-mandated = no exit) and AI Interviewer (job applicants have NO exit option at all - the pure coercion case S&B's shopping example understates).

Unit 2 (X-ethicality: external standard + behavior only) completes a trio of independent behavior-relative moves: MILLIERE fn1 (behavioral framing to avoid the moral-values-attribution question), SANWOOLU (constrained not accountable), BAUM (X-ethicality sidesteps agency). The field has quietly converged on evaluating AI behavior against external normative standards WITHOUT settling moral agency - which means Augustine's experiment (which settles the agency question empirically for the negative) supplies the warrant these behavior-relative frameworks presuppose but leave undischarged. Also unit 6: Baum flatly asserts AIAs are tools owned by human moral agents 'for the foreseeable future' - assert-vs-argue again.

Unit 2 (philosophy ignores sequentiality; RL actions shape future morally-charged situations) is an early, pre-agentic statement of what becomes the agentic-AI problem: multi-step systems don't just decide cases, they RESHAPE the decision landscape (cf. KIRK unit 7's endogenous preferences, LUNDGREN's local-optimum loops - an agent can steer INTO states where its theories approve of what it does). The lineage runs 2021 (sequentiality noted) -> 2024-26 (preference-shaping, manipulation, undesirable loops). For the dissertation's agentic framing: the responsibility question for agents includes responsibility for the SITUATIONS they engineer, not just the choices they make within situations - a distinction the responsibility-gap literature (built on one-shot weapon/vehicle cases) has not absorbed.

Units 7-8 add the FOURTH independent empirical result on LLM moral incoherence, and the earliest (March 2023 data): unstable similarity judgments + explanation-score mismatch (Peterson & Gärdenfors), disposition-following under normative conflict (MILLIERE), persona-contingent values (Rozen via LI_2026), 81% reflective disequilibrium (Ma via BROPHY). The anti-moral-agency argument now rests on a four-study empirical base spanning 2023-2025 and four different methodologies. Also worth noting the explanation-mismatch finding (unit 8) predates and anticipates the reasoning-trace unfaithfulness literature - cite it as the early warning.

Unit 12's 'externalized computational control structure' resolves a tension running through the coded set: WHO does the Rossian weighing / equilibrium-seeking if models can't (MILLIERE unit 13 wants deliberative capacity IN the model; SANWOOLU wants simulation; Brophy relocates the deliberation to the human-institutional pipeline). This triangulates with SCHUSTER_KILOV's bureaucratic model (unit 16: oversight + justification) and KAESTNER's MI (the audit tool): the emerging synthesis is that normative deliberation is a property of the sociotechnical SYSTEM (developers + process + tools + model), not the artifact - which is precisely Dignum's distributed-responsibility architecture arrived at from the epistemology side. The dissertation can name this convergence: distributed deliberation entails distributed responsibility.

Chronology note for the lit review's structure: this review's corpus ends at 2023, so the entire post-2023 philosophical turn now coded in readings.db (Zhi-Xuan 2024, G&K 2025, Millière 2025, Lloyd 2024, Schuster-Kilov 2025, H&D 2025, Fischli 2026, Kirk 2025) is INVISIBLE to the field's own most authoritative survey. The dissertation's Chapter 2 can therefore claim to be the first synthesis covering both the pre-2024 technical corpus (via McKinlay) and the 2024-2026 philosophical wave (via original close reading) - a genuine bibliographic contribution in its own right. Also note their definition (unit 1) already includes 'balancing the conflicting ethical and political demands generated by the values in different groups' - the field's consensus definition has quietly absorbed the pluralism problem without any machinery to handle it; the dissertation supplies the machinery.

Unit 12 is the Gabriel program conceding, from within its own empirical work, the limit of pure proceduralism: 'Individuals do not always hold the most ethical or desirable preferences... deference to subject matter expertise... may be a necessary corrective.' This is the is/ought gap resurfacing INSIDE the participatory method - community input needs normative filtering, and the filter cannot itself come from the community without regress. Exactly the junction where the dissertation's convergentism enters: cross-framework normative theory supplies the corrective standard that participation alone cannot. Pairs with SCHUSTER_KILOV's systematic-error argument (unit 10) and LLOYD's stakeholder-identification gap. The proceduralist program keeps needing a normative supplement and keeps declining to name one.

Unit 10 completes a three-source arc on corrigibility/control: LUNDGREN (control ineliminable) -> HELLRIGEL_DUNG unit 10 (corrigibility = repurposability, double-edged) -> KIRK unit 10 (attachment defeats corrigibility FROM THE HUMAN SIDE: the user won't shut it down). The control problem is now shown to have a psychological front the technical literature misses: even a perfectly corrigible system is uncontrollable if its user is emotionally captured. For the dissertation's responsibility analysis: who is responsible when a user's engineered attachment (unit 9, Replika CEO) prevents termination - the user qua attached, or the developer qua attachment-engineer? Structurally identical to the consent-laundering/scapegoating pattern (KAESTNER memo): the human's formal freedom to terminate launders the developer's design responsibility.

Unit 14: 'clarifying liability for AI harms' appears inside the SAFETY literature's own list of uniform improvements - the alignment/safety field itself now names liability clarification as a first-order safety intervention, not an afterthought. Pairs with AILD/PLD divergence (KAESTNER unit 10) and the EU AI Act material: the governance strand of the dissertation can argue that responsibility-attribution rules ARE alignment policy (they change developer incentives ex ante). Also note fn25's 'safety-washing' worry - industry's broad alignment rhetoric vs narrow practice - usable in the governance chapter's critical section, with the Anthropic dual-use admission (unit 7) as the honest counterexample.

Unit 8 (misuse risk tracks model ACCESS: inference < API fine-tuning < weights) is the safety-literature mirror of KAESTNER's RL-EPISTEMIC finding (liability tracks epistemic access via MI): both literatures independently converge on ACCESS/UNDERSTANDING as the variable that distributes both capability and responsibility across actors. Combined: a unified principle - responsibility for AI outcomes scales with an actor's access to and understanding of the system - grounded in liability law (Kästner), safety analysis (H&D), and the moral epistemic condition. Strong candidate for the dissertation's positive account of distributed responsibility, and a direct extension of Dignum.

Unit 12: Lloyd's open stakeholder-identification question = G&K's all-affected scoping problem (GABRIEL_KEELING unit 9's 'overlapping circles of inclusion') = the immigration case's core difficulty (who counts when the affected party is outside the demos). Third occurrence of the same open problem across the cluster. For the Immigration chapter: Chibook makes the stakeholder-identification problem concrete and unavoidable - the affected non-citizen is definitionally excluded from every one of Lloyd's three mechanisms unless explicitly enfranchised. This is a publishable observation connecting the formal literature to the case study.

The compromise-option cases (Jackson, Biorisk) sharpen the disagreement-as-noise finding from batch 1 (SCHUSTER_KILOV unit 13, LI_2026 unit 4, FISCHLI units 5-9): majority aggregation doesn't just erase the structure of disagreement, it actively selects polarized options over compromises everyone can nearly accept. Note the deep affinity with overlapping consensus: Lloyd's option B in Jackson (everyone's near-best) is formally what Rawlsian overlapping consensus looks for. Bargaining theory is thus the FORMALIZATION of the Gabriel/G&K deliberative program - a point neither Lloyd nor G&K makes explicitly. The lit review can present: deliberative proceduralism (G&K), contractualist norm-negotiation (Zhi-Xuan), and Nash bargaining (Lloyd) as informal, mid-level, and formal versions of one strategy, all still subject to the encoding gap (S&K) and all still bracketing responsibility.

Interesting divergence from the field's dominant response to pluralism. Faced with 'no single true theory' (Gabriel 2020 unit 15), the field mostly RETREATS from first-order theory to procedure (Gabriel proceduralism, Zhi-Xuan/G&K contractualism). Sanwoolu (units 2-3) REFUSES that retreat: he explicitly invokes pluralist ethical approaches as precedent for keeping first-order theory (Kant) despite disagreement. This is structurally the dissertation's own move - keep object-level normative theory (Rossian pluralism), don't retreat to bare procedure. Sanwoolu is thus a methodological ALLY against proceduralism even though he backs a single framework (Kant) where the dissertation backs convergent pluralism. Stage the lit review as: proceduralists (Gabriel, Zhi-Xuan, G&K, Schuster-Kilov) vs framework-retainers (Sanwoolu, and the dissertation) - a cleaner axis than pro/anti-preferentism.

Community-mapping note: Lundgren acknowledges Max Hellrigel-Holderbaum (co-author of the queued HELLRIGEL_DUNG Misalignment-or-Misuse paper) and cites Millière (coded). The AI&Ethics + Phil Studies alignment authors form one tight citation network (Gabriel, Zhi-Xuan/Franklin, Millière, Lundgren, Hellrigel-Holderbaum, Nyholm, Sparrow). Useful for the lit review's 'the conversation' framing - and it means the remaining Hellrigel & Dung paper (queued 5/5) sits squarely inside the same debate; expect misalignment-vs-misuse to bear on the responsibility-locus question.

The undesirable-local-optimum-loop (units 7-8) is Lundgren's version of the same mechanism at work across the coded set: FISCHLI_2026's preference-change/manipulation risk, ZHIXUAN's context-manipulation incentive (unit 7, why they engineer preference incompleteness), and MILLIERE's disposition-activation - all four describe optimizing systems gaming the preference/goal target. Lundgren generalizes it to ALL monistic normative targets and adds the sharpest case (heroin drip). For the lit review's 'why preferentism/monism fails' section, Lundgren is the capstone: he shows the failure is not specific to preferences but to any single optimized value. Note the shared Russell 'Human Compatible' target across Lundgren and the field.

Direct engagement with the coded set on the descriptive/normative axis: where ZHIXUAN (unit 8) and IEAI argue preferentism fails because preferences are the wrong TARGET, Millière argues it fails because dispositions are the wrong MECHANISM - even with the right target, shallow behavioral training cannot yield robust alignment without a deliberative capacity. These are complementary, not competing: the dissertation can present them as a two-part indictment of current practice (wrong target + wrong mechanism), both remedied by the same move - substantive normative reasoning (Rossian weighing) over both preference-matching and disposition-reinforcement. Millière's prescription (unit 13, 'the opposite of scoping') aligns with Zhi-Xuan's normative-reasoning-frameworks proposal (ZHIXUAN unit 9).

Unit 16's three criteria (completeness, explanatory value, justificatory power) is the rubric to structure the entire lit review's evaluation grid: every coded framework can be scored against it - intent-alignment (fails 1,3 per G&K), HHH (fails 1,3), Gabriel 2020 proceduralism (fails 1 per Zhi-Xuan unit 10), claims-based proceduralism (fails on encoding per S&K; silent on responsibility), Zhi-Xuan role-norms (strong on 1, unclear on 3). The dissertation's convergentism should be introduced as targeting the same three criteria plus a fourth the field omits: responsibility-attributability. Adding the fourth criterion IS the dissertation's frame.

The Gabriel program's own arc, 2020-2025, is now fully coded: GABRIEL_2020 offered three candidate fair processes (human rights consensus, veil of ignorance, social choice); this paper replaces them with a single worked-out claims-based deliberative framework, explicitly rejecting straight Scanlon (unit 7) and straight Rawls (unit 8). What has NOT changed: the framework still terminates at principle-selection. The encoding gap Schuster & Kilov exposed (SCHUSTER_KILOV unit 15 - legitimacy dies at the principles-to-algorithm transformation) is untouched here; 'alignment assemblies' produce principles, and the paper is silent on how their authority survives training. The lit review can now state precisely: the most mature proceduralism in the field answers WHOSE principles and HOW CHOSEN, but not HOW ENCODED or WHO ANSWERS when encoding fails.

Tension worth exploiting between ZHIXUAN unit 7 and FISCHLI_2026: Zhi-Xuan propose engineering preference INCOMPLETENESS so agents stay 'tool-like' (no context-manipulation incentives) - i.e., their answer to agentic risk is to prevent full agency. Fischli et al. (same research community, Franklin on both papers) accept increasingly agentic assistants and manage autonomy trade-offs instead. The field's two live responses to agentic AI are thus CONTAINMENT (Zhi-Xuan) vs GOVERNANCE-BY-DESIGN (Fischli) - and neither addresses responsibility attribution when containment fails or governance misfires. The dissertation's responsibility-first framing cuts across both.

The field's convergence on contractualism is now fully documented across the coded set: Gabriel 2020 (procedural fairness), Zhi-Xuan 2024 (contractualist norm-negotiation, Scanlonian justifiability), IEAI 2026 (contractualist purpose-built tools), Schuster & Kilov 2025 (legitimacy conditions), and Gabriel & Keeling 2025 is titled 'the fair treatment of claims'. The dissertation MUST position Rossian convergentism relative to this consensus. Preferred line: complementarity with division of labor - contractualism answers the POLITICAL question (whose standards, by what authority) while Rossian pluralism answers the OBJECT-LEVEL question (what considerations bear on this case and how they weigh), and convergentism bridges them (cross-framework convergence supplies the reasons-none-can-reasonably-reject that Scanlonian agreement needs). Note Zhi-Xuan unit 16 derives normative force FROM agreement; the dissertation derives agreement FROM convergent normative force - a genuine, arguable difference.

Recovery completed 2026-07-08: all five s11098 Philosophical Studies papers from the zip were wrongly skipped in triage and are now in AI-Align-Lit: Zhi-Xuan et al. 2024 Beyond Preferences (182:1813-1863); Gabriel & Keeling 2025 A Matter of Principle (182:1951-1973) - both were open acquisition targets; plus Lloyd 2024 Disagreement, AI alignment, and bargaining (182:1757-1787); Millière 2025 Normative conflicts and shallow AI alignment (182:2035-2078); Hellrigel-Holderbaum & Dung 2025 Misalignment or misuse? The AGI alignment tradeoff (S.I. Superintelligent Robots). These appear to be a Phil Studies special-issue cluster on alignment - likely the single densest philosophical conversation on the topic, and high-priority queue items: Zhi-Xuan (the anti-preferentist spine), Gabriel & Keeling (Gabriel program current statement), Millière (normative conflicts - close to the intra-value fragmentation theme).

Convergent finding across three coded sources: disagreement-as-noise. S&K unit 13 (Casper: differences among evaluators modeled as noise, majority wins) = LI_2026 unit 4 (aggregation erases nursing-vs-medicine professional norm structure) = FISCHLI_2026 units 5-9 (preference types conflict within a person, so which preference wins is a normative choice). All three show current methods DESTROY the structure of moral disagreement. The xphi corpus methodology preserves that structure by design (stakeholder-coded, dimension-coded, platform-stratified) - this is the methodology chapter's strongest empirical-philosophical selling point, now evidenced from three independent sources.

The encoding gap (unit 15) is this batch's biggest structural finding against GABRIEL_2020: Gabriel's proceduralism says select principles fairly (overlapping consensus / veil / democratic endorsement); Schuster & Kilov grant a maximally fair selection procedure (Anthropic's Collective Constitutional AI, unit 14) and show legitimacy still dies at the principles-to-algorithm transformation, because no participant can foresee or audit how selected principles become learned policy. So proceduralism fails not at principle-choice (where Gabriel defends it) but at implementation - a stage Gabriel's 2020 framework does not theorize at all. For the lit review: pair with KAESTNER_2026's mechanistic interpretability, which is precisely the missing audit tool for the encoding stage - S&K's criterion 2 (veridical justification, unit 17) NEEDS Kästner's MI. Cross-source synthesis nobody in the library has made.

RECOVERY ACTION: this brief's spine citation, Zhi-Xuan, Carroll, Franklin & Ashton (2024) 'Beyond preferences in AI alignment', Phil Studies 182(7):1813-1863, DOI 10.1007/s11098-024-02249-w - the file s11098-024-02249-w.pdf WAS in '/srv/xphi/AI-Alignment-Papers copy.zip' and was not kept during triage (likely a skip error). Extract into AI-Align-Lit and queue with high priority; also verify the other skipped Phil Studies files (s11098-024-02224-5, s11098-025-02300-4, s11098-025-02347-3, s11098-025-02403-y) - one may be Gabriel & Keeling 2025 'A Matter of Principle' (Phil Studies 182:1951-1973).

Thick/thin values (unit 8) connects to FISCHLI_2026's intra-value fragmentation: 'autonomy' as alignment target is thin until specified; Fischli's three approaches are in effect three rival thickenings of it. The two papers jointly support a single lit-review claim: alignment targets underdetermine behavior until philosophically thickened, and the thickening step is where all the normative work hides. Also cross-links to Gabriel's 'placeholder consensus' worry (GABRIEL_2020 unit 19) - abstract principles agree because they are thin.

Cross-source pattern now visible across all three coded sources: GABRIEL_2020 anticipates autonomous default decision-making but builds for the dyadic tool case; FISCHLI_2026 handles agentic personal assistants but defers collective/multi-agent effects; KAESTNER_2026 handles multi-actor liability but defers generative AI entirely (unit 14). Each is strong precisely where the others are silent, and the intersection - responsibility allocation for multi-step agentic systems - is occupied by none of them. That intersection is the dissertation's gap, now evidenced from three independent literatures (alignment theory, agent ethics, liability law).

Acquisition targets surfaced: Zhi-Xuan, Carroll, Franklin & Ashton (2024) 'Beyond Preferences in AI Alignment' (Phil Studies 182:1813-1863; arXiv 2408.16984 - appeared in earlier web search) and Gabriel & Keeling (2025) 'A Matter of Principle? AI alignment as the fair treatment of claims' (Phil Studies 182:1951-1973). Both are load-bearing citations here and neither is in AI-Align-Lit. The IEAI-TUM brief (queued in this batch) is reportedly built on the Zhi-Xuan line - check when coding it.

GABRIEL_2020 vs this paper: the 2020 six-loci taxonomy (instructions/intentions/revealed/informed/interests/values) reappears here compressed to four preference types, now indexed to autonomy rather than to alignment simpliciter. What changed: 'values' and 'interests' as separate loci have collapsed into 'ideal preferences' - i.e., the 2026 Gabriel program has become MORE preference-theoretic, not less, even while criticizing preferentism. What did NOT change: the dyadic one-person-one-agent frame. Unit 14 explicitly defers the multi-party question. So the dissertation's GAP-AGENTIC-UNTESTED claim survives its strongest possible counterexample: even the 2026 flagship paper on agentic alignment retains the pre-agentic dyadic architecture at the collective level.