Analytic memos (51)

DIFFERENTIATION CONFIRMED (deep read): Kluge Corrêa and the dissertation share two formal moves - ordinal-only treatment (his unit 11 = the dissertation's ordinal-convergence reply to Lloyd) and a coherentist epistemology (his Sayre-McCord base = the dissertation's MWRE bridge). The differences are now precise: (1) OBJECT-LEVEL THEORY - Dynamic Normativity deliberately avoids any substantive moral theory ('not rooted in any quintessential metaethical blueprint'), remaining a procedural-epistemic account of PREFERENCE aggregation; the dissertation runs substantive frameworks (Ross/consequentialism/contractualism) and uses their CONVERGENCE - so Kluge Corrêa remains within what Zhi-Xuan would call the preferentist data model, while the dissertation works with reasons/values; (2) RESPONSIBILITY - no allocation theory anywhere in Dynamic Normativity (mitigation is about system impact, not human accountability); (3) EMPIRICS - his empirical layer is the WAIE guidelines survey (documents), the dissertation's is naturalistic folk discourse at scale (stakeholders); (4) his 'sufficient minimal alignment' is a design standard, the dissertation's convergentism is a justification standard. Complementary more than rival; cite as the nearest neighbor and state these four contrasts explicitly in the intro. Also mine Ch1's WAIE dataset (200 guidelines, marginalized-regions analysis) as a companion empirical source for the governance chapter.

The Kästner-vs-Hällström adjudication is a ready-made dissertation section: Kästner et al. make responsibility TRACK epistemic access (MI shifts liability to whoever understands); Hällström & Gouveia DECOUPLE them (organisational virtue grounds responsibility regardless of understanding). The dissertation can argue these are complementary at different timescales: virtue governs EX-ANTE conduct under irreducible opacity (choosing to deploy, monitoring, acknowledging uncertainty - Gregg's escalation layer), while difference-making + epistemic access allocates EX-POST liability for specific harms - and the FMTI transparency decline (Stanford HAI) shows why neither alone suffices: virtue without disclosure is unverifiable (organisations can claim conscientiousness - the safety-washing worry, H&D fn25), disclosure without virtue is Akyol's ethical hallucination at the organisational level. Synthesis: virtue supplies the standard of care, difference-making the attribution machinery, institutions the verification - the three-level allocation again.

PRIORITY OBJECTION TO ANSWER: Mahant attacks the inference the dissertation (and Sanwoolu, Hakli & Mäkelä, Coeckelbergh) relies on - from 'AI lacks moral agency' to 'responsibility remains with humans'. If responsibility doesn't require agency, denying AI agency leaves open that the SYSTEM bears (some kind of) responsibility - reopening RL-SYS. Available replies to develop: (a) disambiguate responsibility KINDS (Mahant's own Sect. 7 concedes classifications matter): attributability and liability may not require agency, but ACCOUNTABILITY (answerability - giving reasons, being blamed with uptake, making amends) plausibly does, and the practical stakes (compensation, sanction, repair) run through accountability - so the human-transfer conclusion survives for the responsibility-kind that matters to victims; (b) even granting non-agent responsibility as coherent, it is practically idle without sanctionability - which reconnects to Tasioulas's accountability-capacity point and the Diella case; (c) note convergence: Mahant's view actually HELPS one dissertation thread - it legitimates holding SYSTEMS to normative standards (constrained evaluability) without agency-metaphysics. Engage in the responsibility chapter as the strongest published counterargument; possibly worth direct correspondence (Uppsala, AICon project).

Gregg supplies the missing PREMISE-LEVEL answer to the responsibility gap that the allocation project needs: even where control is partially lost (conceded, unit 1), responsibility does not evaporate - it ESCALATES to the political community that chose deployment under acknowledged uncontrollability. This is the ex-ante/community-level allocation that Yampolskiy's impossibility claims force and H&D's misuse analysis implies, stated as a positive thesis. For the dissertation's Part II architecture: Gregg gives the top layer (political community, inescapable), Kästner the middle (actor-level difference-making), the epistemic/access condition the distributing principle - a three-level allocation structure no single source has.

Use as the pessimist ceiling in the responsibility chapter: standard responsibility conditions (control - Fischer & Ravizza; foreseeability - negligence law; explanation - Kästner's MI regime) are each targeted by one of Yampolskiy's impossibility claims. Two consequences: (a) if he is right, the responsibility gap is PERMANENT and the only honest allocation is ex-ante (development/deployment decisions - who chose to build/deploy under acknowledged uncontrollability), which converges with H&D's governance turn and strict-liability regimes (PLD); (b) if he is wrong in degree (Kästner's MI, Melo's alignment-by-construction), responsibility tracks the achievable degree of control/explanation - which is the graded picture the dissertation defends. Either way, cite the chapters not the vibe: Ch 4 (unverifiability) pairs with MELO's theorem; Ch 6 (uncontrollability) with LUNDGREN.

The principal-agent frame is the unexploited bridge between the technical and responsibility literatures: economics uses principal-agent theory precisely to analyze RESPONSIBILITY under delegation (who bears risk, who monitors, how incentives allocate accountability) - yet the AI-alignment appropriation kept the incentive mathematics and dropped the accountability apparatus. The dissertation can restore it: each case study is a principal-agent triangle (hospital-clinician-Scribe; state-officer-Chibook; employer-manager-Interviewer) where H-M's formal results (unit 2: incomplete proxies guarantee persistent misalignment) PREDICT the failure modes, and the economics of delegation supplies allocation principles (monitoring duties, incentive design = Kästner's difference-makers in economic dress). Also note unit 2 is the formal engine behind Zhi-Xuan's critique and Lundgren's loops - cite the theorem, not just the slogans.

Cappelen & Dever supply the deepest layer of the dissertation's anti-agency stack: below 'no phronesis' (Noller), 'no deliberative capacity' (Millière), and 'no stable values' (five empirical studies) sits 'no settled account of CONTENT' - if it's open whether the system's outputs mean anything, ascribing moral judgments (let alone agency) is doubly premature. Their closing argument (unit 2: reliability theories collapse into content theories) also warns the dissertation's own experiment: interpreting LLM 'moral reasoning' outputs presupposes a metasemantic stance - state it (a modest externalism: outputs inherit content from human linguistic practice, per the Wittgenstein thread) rather than assume it. And unit 3's Lucie case is the philosophy-of-language ancestor of the intelligible-reasons requirement running through S&K, J&N, and Tasioulas - cite as the origin point.

Tasioulas is the philosophical anchor the dissertation's pluralism should cite FIRST: the three Ps state, with Oxford authority and in 2022, the exact combination the dissertation builds on - plural incommensurable values (Rossian structure), procedures mattering beyond outcomes (Baum's execution scope), participation (the stakeholder corpus's rationale). Unit 3 (noise as eligible variability) additionally arms the methodology chapter against a subtle objection: folk disagreement in the corpus is NOT necessarily error to be averaged away (contra Condorcet-style treatments, cf. S&K unit 8) - some of it is rational variability within the eligible set, and the convergentist method's job is to distinguish eligible-range variation from genuine conflict. That distinction - operationalized via the corpus's reasoning codes - could be an original methodological contribution.

The Wittgensteinian rule-following point is the philosophical deep-structure beneath THREE coded findings: S&K's encoding gap (principles don't determine learned policy), Millière's shallow dispositions (trained regularities aren't rule-grasp), and STELA's rule-interpretation problem (annotators must interpret rules via their own background). All are instances of: no formulation carries its own application; application is stabilized by shared practice. For the dissertation this cuts both ways - it deepens the critique of constitution-writing approaches, AND it supports the corpus methodology (folk discourse IS the record of the shared practice within which AI-ethics terms have their use - meaning-as-use makes naturalistic discourse data philosophically load-bearing, not merely sociological). Cite the pair (2024 + 2025 response to Bangu) together.

Manheim is the needed counterweight in the moral-agency chapter: unlike the five empirical-incoherence studies, he argues the grounding deficit is CONTINGENT - agentic affordances (memory, persistent identity, world-mediated action) are exactly what would upgrade symbol-reflection into genuine interpretation, with no architectural barrier. Two consequences for the dissertation: (a) the anti-moral-agency conclusion should be indexed to CURRENT systems and stated with Manheim's contingency acknowledged (matching Gabriel & Keeling's fn14 openness to revision) - this is also the honest reading of Augustine's own experiment; (b) agentic AI thus threatens BOTH sides of his framework at once: it breaks pre-agentic responsibility frameworks AND potentially erodes the no-moral-agency premise. The dissertation's convergentist/distributed-responsibility architecture should be stress-tested against the 'grounded agent' scenario - a section Howard will appreciate as anticipating the strongest objection.

Formal reinforcement for two coded arguments: (a) SCHUSTER_KILOV's encoding gap - if verifying an arbitrary model's alignment is undecidable, then democratic legitimacy for the principles->algorithm step cannot be recovered by after-the-fact verification even in principle, only by construction-time process (strengthening the case that legitimacy must attach to the PROCESS); (b) KAESTNER's MI-as-gold-standard - undecidability bounds what interpretability audits can promise for arbitrary architectures, supporting their restriction of the demand to high-risk/high-stakes contexts and the recommendation of inherently interpretable systems. One-cite in governance chapter: verification-based regulation (audits) has a formal ceiling; architecture/process regulation does not.

One-cite use: when the metaethics chapter argues that normative uncertainty is not a temporary engineering deficit but structurally ineliminable, Wyeth & Hutter supply the formal ceiling case - at AIXI-level idealization, utility assignment still bifurcates between interpretations (death vs ignorance) with different decision consequences, requiring imprecise-probability machinery. Pairs with Ecoffet & Lehman (moral uncertainty at the practical level) as the in-principle bookend.

The legal-centric proposal (units 1-2) needs a principled reply in the governance chapter, since the dissertation's regulation strand could be mistaken for it. Three points: (a) law-as-crystallized-volition inherits every defect the coded set found in descriptive sources of norms - historical legal records encode past injustice (slavery, coverture, colonial law were law), so 'trained on jurisprudence' needs exactly the normative filter (Brophy's filtration; the moral-error caution in GABRIEL_2020 unit 25) that pure legal positivism cannot supply; (b) G&K's analysis (legal compliance = minimum standard; law lags; some things shouldn't be legislated) directly counters legal sufficiency; (c) the constructive version: law matters as the INSTITUTIONALIZATION layer of Josifović & Noller's normative interface - the enforcement architecture for independently justified norms, not their source. The two Josifović papers read together actually support (c): the 2026 paper's justificatory structures do the normative work the 2024 paper assigned to legal content.

This paper must be engaged head-on in the methodology chapter, because on its surface it condemns exactly what the dissertation does (collect folk moral responses about AI at scale). The defense has three layers, all already in the coded set: (1) PURPOSE - LaCroix attacks using folk data as a VALIDATION BENCHMARK/ground truth; the dissertation uses it as reflective-equilibrium INPUT (Brophy: CMJs are provisional, filtered, revisable) and as intuition-pump-at-scale (LaCroix's own approved use, unit 1 - the corpus is a massive systematic comparison of cases exposing morally salient differences); (2) AWARENESS - LaCroix's constructive standard (unit 7) demands researchers know what they measure; the dissertation explicitly measures folk normative DISCOURSE and says so, never presenting distributions as verdicts; (3) BRIDGE - the convergentist framework supplies the metaethical account (unit 4's demand) of when folk convergence carries evidential weight. Done right, LaCroix becomes the dissertation's ally: he shows why Moral-Machine-style benchmarking fails, which is precisely the contrast class that makes the corpus-plus-metaethics design necessary.

Unit 5 is the first source in the library to STATE distributed responsibility as a design principle rather than bracket it - but note what's still missing: Noller says responsibility is shared among designers/policymakers/users and requires 'institutional feedback architectures', yet gives no allocation mechanism (what share, on what basis, adjudicated how?). Same for 'ethical resilience' (unit 6) - a success metric with no measurement. This is the recurring pattern at its clearest: the field now AFFIRMS distributed responsibility in principle (Noller, Dignum, IEAI brief unit 11) but has no theory of the distribution. Kästner's difference-making + the epistemic/access condition (KAESTNER, HELLRIGEL_DUNG memos) + the dissertation's case-based analysis is that theory. The lit review's Part II thesis: from AFFIRMING shared responsibility to ALLOCATING it.

The DVA moral-jury architecture (units 2-3) is the closest thing in the library to an ENGINEERING blueprint for Rossian pluralism: distinct normative modules (prima facie perspectives), context-sensitive weighted aggregation (the all-things-considered judgment), anti-extremization as a design property (unit 4 - answering Lundgren's local-optimum loops exactly as the dissertation's memo predicted pluralism would). Two dissertation-relevant critiques to develop: (a) user-controlled weights RELOCATE rather than solve the normative problem - the user becomes the weigher, inheriting the intra-value fragmentation problem (FISCHLI memo: meta-autonomy presupposes the user can do the weighing the theory couldn't); (b) numerical score aggregation re-imports the cardinal-comparability assumption Lloyd showed to be undefined across theories - a convergence-first architecture (flag verdict agreement before aggregating scores) would be more defensible. Both critiques are constructive extensions, publishable as engagement.

Unit 3 (sycophancy is structurally entailed by preference-maximization, since both humans AND preference models prefer agreeable over correct responses) upgrades the anti-preferentist case from 'preferences are the wrong target' (Zhi-Xuan) to 'optimizing preferences actively CORRUPTS the other alignment goals' - preference-maximization doesn't just miss values, it trades honesty away. Combined with unit 4 (mirroring is strongest exactly on contested moral questions), this yields a fifth empirical datum for the anti-moral-agency thread AND a self-standing argument: a system whose moral outputs track the INTERLOCUTOR'S views rather than the merits cannot be a moral reasoner in any load-bearing sense (it fails the most basic independence condition on judgment). Cite alongside Peterson-instability, Millière-dispositions, Rozen-personas, Ma-disequilibrium.

The justificatory-gap argument (units 3-4) is the most precise statement yet of the inference the dissertation must engage, sharper even than Baum's syllogism: uncertainty (normative + metanormative) eliminates theoretical justification, so only political justification remains. The convergentist reply targets the ELIMINATION step: uncertainty about which theory is true does not eliminate theoretical justification for verdicts on which the candidate theories CONVERGE - convergence is precisely theoretical justification that survives theory-level uncertainty (each theory underwrites the verdict from its own premises; the verdict inherits support from the disjunction). So the gap is narrower than S&B claim: political justification is needed only for the non-convergent residue. This slots the dissertation into their epistocratic/democratic axis as a PRINCIPLED HYBRID: convergence-checking is the 'expert' layer (but imposes no single theory), stakeholder data the participatory layer - exactly the hybrid they call for (unit 9), with a worked mechanism they lack.

The outcome/execution scope distinction (unit 9) upgrades the responsibility taxonomy: KAESTNER's difference-makers and G&K's misalignment modes both concern outcomes; Baum shows EXECUTION misalignment (right result, wrong means - bribing the maître d') is a distinct failure class. For agentic systems this is central: multi-step agents choose their own means, so execution alignment is precisely what increases with agency (cf. HELLRIGEL_DUNG unit 11) - and execution wrongs may have NO outcome trace (the bribe succeeded, everyone's happy). Who is responsible for unobserved wrongful means? Connects to Kästner's type-(ii) difference-makers (only interpretability reveals the execution path) and to the folk corpus: check whether folk comments distinguish outcome-blame from means-blame - a codable distinction the AI Responsibility category may already contain.

Unit 4's syllogism is a gift: the proceduralist consensus (Gabriel, Zhi-Xuan, G&K, Schuster-Kilov) finally stated as an explicit inference with numbered premises - which means the dissertation can locate its dissent PRECISELY at P4. The convergentist counter: P1-P3 are all true, but P4 is a non-sequitur, because there is a third option between imposing one theory and retreating to procedure - aligning with verdicts on which the plural reasonable theories CONVERGE imposes nothing on anyone (each constituency can accept the verdict from its own premises) while remaining genuinely MORAL alignment, not just public justifiability. Baum's own observation (unit 5) that proceduralism is ambiguous between truth-approximation and aim-replacement strengthens this: convergentism resolves the ambiguity by keeping moral truth-directedness while earning public justifiability as a byproduct. The metaethics chapter should quote the syllogism verbatim and argue against P4.

Ecoffet & Lehman's Nash voting (unit 5) is the INTRA-agent mirror of Lloyd's INTER-stakeholder bargaining (LLOYD unit 10) - same Nash machinery, applied to theories-within-one-agent vs stakeholders-across-society. Together they show the field's formal toolkit treats 'which theory wins' and 'whose values win' as the same aggregation problem. The dissertation's convergentism dissolves rather than solves this problem in the favorable cases: where theories CONVERGE on a verdict, no voting/bargaining is needed - aggregation machinery is only for residual disagreement. So the convergentist architecture = (1) check convergence first (cheap, no comparability needed), (2) fall back to Proportional-Say-style mechanisms only for genuinely contested cases. This two-stage proposal is concrete, novel, and directly implementable in their formalism - a possible technical-companion paper to the dissertation.

Conceptual spaces offer the dissertation a THIRD formal option beside utility aggregation (rejected via Lloyd/Zhi-Xuan) and pure ordinal convergence: values as REGIONS with prototype structure. This fits Rossian pluralism strikingly well - prima facie duties as distinct regions in moral space whose boundaries are contextually negotiated, with 'weighing' as locating a case relative to multiple prototypes rather than summing utilities. If the metaethics chapter wants a formal semantics for how plural duties coexist without commensuration, Gärdenfors geometry is a candidate the alignment literature has already domesticated. Speculative - flag for Howard discussion rather than committing.

Unit 6 (multiple equilibria compared by explanatory power, scope, empirical fit, parsimony) supplies the second half of the answer to LLOYD's intertheoretic-comparison objection (LLOYD memo): convergentism needs ordinal agreement on verdicts (first half), and when frameworks DON'T converge, competing equilibria are adjudicated by theoretic virtues, not by a common choiceworthiness scale (second half). Together: the dissertation's metaethics never requires the cardinal intertheoretic comparisons that sink MSEC. Also note unit 6 = Daniels' reply maps exactly onto the convergentist posture: multiple coherent frameworks, disagreement rational, convergence evidential.

This is the dissertation's methodology chapter written by someone else at 80% - which is both validation and a mandate to differentiate. WHAT BROPHY PROVIDES: published warrant that reflective equilibrium is THE right epistemology for alignment; the filtration answer to 'corpus data is just biased opinion' (unit 5: IMJs are filtered into CMJs - the xphi pipeline's cleaning, deduplication, category-coding, and IRR checks ARE a filtration process and should be explicitly described as one); the Nielsen quote (unit 4) answering the culturally-skewed-data objection head-on. WHAT THE DISSERTATION ADDS BEYOND BROPHY: (a) Brophy's CMJ layer is RLHF preference data - thin, unreasoned; the folk corpus supplies REASON-ANNOTATED judgments, a far better CMJ approximation; (b) Brophy's BT layer is 'the constitution' - the dissertation's BTs are actual normative theories (Ross, consequentialism, contractualism) with the convergentist test as the coherence check; (c) Brophy never says WHOSE judgments - the stakeholder structure (government vs public) gives the CMJ layer a political dimension Brophy lacks. Cite generously, then extend.

Unit 4's translation barrier (humanities can't operationalize, engineers can't philosophize - Rahwan's 'cultural divide') is the deepest justification for the dissertation's HYBRID method: LLM-assisted coding of normative discourse is literally a translation device between the two registers - philosophical categories (reasoning types, value dimensions, responsibility loci) rendered as operational annotation schemes at engineering scale. The methodology chapter should claim this explicitly: the pipeline is an existence proof that Tolmeijer's pitfall (theories too abstract to implement vs implementations philosophically faulty) can be navigated, with IRR/validation numbers as the evidence of fidelity in both directions.

Unit 9 (deliberation attaches REASONING to rules) + unit 7 (community rules are thick) empirically confirm the Zhi-Xuan/thick-values line from the theory side of the corpus: when you actually ask affected people, they produce reason-laden, contextual norms, not preference orderings. And the folk_ai 'reasoning' field captures exactly this layer at scale. Also note unit 6: the developer/community divergence is the first EMPIRICAL demonstration in the library that the choice of alignment target is non-neutral - Gabriel's own team showing Gabriel's 2020 worry is real. For Howard: this is what 'the corpus does normative work' looks like - divergence findings falsify the assumption that developer values proxy for everyone's.

Unit 11 (harm as epiphenomenon, no intentional agent anywhere) is the limiting case for ANY responsibility taxonomy and belongs in the dissertation's responsibility-gap chapter as the hardest test: not Matthias's autonomous-system gap (no human in causal command) but a RELATIONAL gap - the harm emerges from the dyad's dynamics with no locus of intention at all. Kästner's difference-making machinery still works here (the design choices enabling personalisation/memory are type-(iii) difference-makers), which shows the causal-analysis approach outperforms intention-based approaches for relational harms. Strong argument for the dissertation's Dignum-style structural account over mens-rea-style accounts.

Unit 7 states the deepest methodological challenge in the library to using preference/attitude data as normative evidence: if sustained AI interaction makes the human reward function ENDOGENOUS (the system shapes the preferences it is scored against), then preference data collected downstream of AI exposure is potentially contaminated as evidence of what people value. This applies to the xphi corpus: social-media comments about AI are themselves produced within algorithmic environments. The methodology chapter needs a considered answer - available moves: (a) the corpus captures REASONED discourse (arguments, justifications), not bare preferences, and reasons can be evaluated on their merits regardless of causal origin (cf. ZHIXUAN units 2/18: preferences derive from reasons - reasons are the stable layer); (b) cross-platform and cross-stakeholder triangulation dilutes any single feedback loop; (c) Gibbard's expressivism evaluates the norms expressed, not the causal history of their expression. Do not let an examiner raise this first.

Unit 11's agency tradeoff (lower agency -> less misalignment risk but MORE misuse risk, because non-agentic systems cannot resist being misused) is the most sophisticated statement yet of why 'agentic AI' matters normatively, and it complicates both prior positions in the coded set: contra Zhi-Xuan's containment strategy (engineered tool-likeness, ZHIXUAN unit 7), tool-like systems are maximally misusable; contra simple agentic-alarmism, agency confers misuse RESISTANCE. For the dissertation's agentic-AI framing: agency redistributes rather than merely increases risk - and correspondingly redistributes responsibility (a misused tool implicates the user/designer; a resisting agent that is overcome implicates the attacker; a corrigible agent repurposed implicates whoever corrected it, unit 10). A responsibility taxonomy indexed to agency level is an original contribution sitting right here.

This paper supplies the dissertation's missing bridge between alignment and responsibility, almost ready-made: the tradeoff thesis says solving alignment RELOCATES catastrophic risk from the system to humans (unit 4: an aligned AGI's controller inherits the disempowerment capacity; unit 12: alignment is what makes systems predictable enough to misuse). Consequence: the better alignment succeeds, the more every AI catastrophe becomes a HUMAN responsibility question (misuse by which actor, with what access, under whose coercion) - i.e., alignment progress makes the responsibility-attribution question MORE central, not less. This inverts the field's implicit assumption that responsibility is downstream cleanup after alignment; on H&D's own argument, responsibility allocation is the core of what remains once alignment works. Use this as the lit review's pivot from Part I (alignment) to Part II (responsibility).

Unit 8 (intertheoretic unit comparisons) is the sharpest formal objection in the library to the dissertation's convergentism and must be answered head-on in the metaethics chapter: if Rossian, consequentialist, and contractualist evaluations cannot be measured on a common scale, what does 'convergence' even mean? The answer available to Augustine - and it is a good one - is that convergentism needs only ORDINAL agreement on the verdict (all frameworks rank option X above its rivals in this case), not cardinal comparability of choiceworthiness magnitudes. That is exactly the structural difference between convergence-as-corroboration and MSEC-style aggregation: convergentism never adds theories together, so Lloyd's undefinedness objection doesn't bite, and his fanaticism objection (unit 7) doesn't either (a 0.01%-credence view can't dominate a verdict it merely fails to join). Lloyd thereby inadvertently shows why convergentism is FORMALLY more robust than the moral-uncertainty machinery. Write this section early; an examiner will ask.

Flag for critical engagement: Sanwoolu's load-bearing claim that transformer models are the 'functional equivalent' of Kantian practical judgment (units 12-13) sits in direct tension with MILLIERE_2025 and LI_2026's Khamassi strong/weak distinction. Millière shows LLMs do NOT reliably perform contextual normative weighing (they follow disposition-activation, not salience-tracking); Khamassi shows LLMs pass pattern-matching but fail intentionality/causal reasoning. So Sanwoolu's functional-equivalence claim is empirically contestable by two other coded sources. This is a live disagreement the dissertation can adjudicate WITH DATA: Augustine's LLM moral-reasoning experiment directly tests whether transformer 'maxim formation' tracks morally salient facts or just salient prompt features. If Millière is right, Sanwoolu's Kantian-alignment proposal fails at exactly the practical-judgment step - and Augustine has the experiment to show it.

The constrained-vs-accountable distinction (AG-CONSTRAINED, unit 10) is a sharp conceptual tool the whole dissertation can adopt: it cleanly separates two questions the literature conflates - (a) is the AI subject to/guided by norms? (morally constrained - yes) and (b) does the AI bear responsibility for norm-violations? (morally accountable - no). This maps precisely onto Gabriel & Keeling's stipulation (GABRIEL_KEELING unit 10: 'morally evaluable actions' without 'moral claims') - Sanwoolu ARGUES what G&K ASSUME. And it sets up the dissertation's central move: if AI is constrained-but-not-accountable, then responsibility for its norm-violations MUST land on humans/institutions (RL-DEV/USER/INST), never RL-SYS - which is exactly why the responsibility-attribution question (Kästner) is the unavoidable next step the alignment literature keeps bracketing.

The RL-CONTROL code (autonomy = self-control, unit 11) creates a bridge the dissertation can exploit: control connects the alignment literature (Lundgren, Russell, Bostrom) to the responsibility literature (Kästner's epistemic condition, the control condition on moral responsibility from Fischer & Ravizza). Both literatures independently converge on control/controllability as necessary - alignment needs it to avoid catastrophe, responsibility needs it to attribute blame. This is a unifying thread for the whole review: control is the hinge between 'is the system aligned?' and 'who is responsible when it isn't?' Nyholm's five-quality control analysis (cited here, overlaps Fischli's autonomy components) is a shared conceptual resource worth acquiring.

Lundgren unintentionally builds the dissertation's positive case. His across-the-board demolition (units 8-13) shows EVERY monistic framework fails when operationalized - and his own diagnosis (unit 10) is that value MONISM is the culprit: 'a singular intrinsic good makes a normative theory more sensitive to... simple-minded goal-satisfaction.' He then dismisses pluralist combination in one sentence (fn/end of sec 4) because 'the failure all depends on similar concerns [lack of control].' But that conflates two different failures: the local-optimum-loop failure is about OPTIMIZATION + monism, not about pluralism per se. A Rossian pluralist system that WEIGHS competing prima facie duties (rather than optimizing one) is not obviously subject to the loop - the plurality of non-optimized duties is precisely what blocks single-value gaming. The dissertation can turn Lundgren's argument around: his cases show monistic optimization fails; Rossian pluralism + control is the surviving option, not the excluded one. This is a direct, publishable rejoinder.

Unit 11 (thought-injection, unfaithful reasoning traces) is the empirical counterpart to SCHUSTER_KILOV's unit 17 ('the wise judge, not the clever but unprincipled lawyer' - justifications must guide, not rationalize). Millière shows reasoning traces can be manipulated so the stated reasoning does NOT govern behavior - exactly the faithfulness failure S&K worry about, now empirically demonstrated. Both point to the same open research problem: verifying that cited reasons are causally operative. This is testable with folk_ai.db's 'reasoning' field methodology - a faithfulness study is now supported by two independent Phil Studies papers plus the Khamassi strong/weak distinction (LI_2026).

This is the single most important source in the library for the metaethics chapter. Millière independently arrives at the claim that Ross's prima facie / all-things-considered distinction (unit 6, explicit Ross 1930 citation) is the CAPACITY whose absence explains why LLM alignment fails - i.e., the dissertation's chosen metaethical architecture (Rossian pluralism) is not just philosophically defensible but is being identified, by a philosopher publishing in a top venue, as the missing ingredient in cutting-edge AI safety. This flips the usual worry: rather than needing to justify why a dissertation on Ross is relevant to AI, Augustine can cite Millière to show Rossian structure is the field's own emerging answer. The metaethics chapter's thesis can be: 'what Millière calls the capacity for normative deliberation IS the operationalization of Rossian pluralism, and my convergentist account specifies how the all-things-considered weighing should proceed.'

The six modes of misalignment (unit 11) crossed with the dissertation's RL-* codes yields an original analytical product: each mode implies a distinct responsibility allocation that G&K leave unstated (mode 1-2: developer responsibility for reward design; mode 3: user responsibility + developer guardrail duty - which they bracket at unit 12; mode 4-5: developer responsibility qua agent; mode 6: institutional/state responsibility). A 'responsibility completion' of the G&K taxonomy - saying for each misalignment mode who owes what to whom - is a publishable chapter section and possibly the Res Practica submission Howard flagged.

Unit 10 (fn10-11-14) is the single most useful passage for the Luke debate: Gabriel & Keeling ASSUME (i) minimal moral agency - AI actions are morally evaluable; (ii) no moral standing - AI advances no claims; (iii) openness to future revision. This is exactly the position Augustine's LLM moral-reasoning experiment argues FOR: his experiment supplies the missing argument for the assumption the field's leading paper merely stipulates. Frame the experiment's contribution as: 'Gabriel & Keeling assume AI systems cannot advance claims; I provide empirical-philosophical grounds for that assumption, and specify the conditions under which it would fail.' That is a precise, publishable contribution - and it answers Howard's worry that the anti-moral-agency work is 'unnecessary territory' by showing a top-venue paper NEEDING it.

Incommensurability (unit 4) is the decision-theoretic face of Rossian pluralism: Ross's plural prima facie duties with no fixed exchange rate = Chang/Anderson's incommensurable values = Zhi-Xuan's incomplete preferences and vector-valued rewards. This completes the answer to GABRIEL_2020 unit 3's challenge (RL architecture favors consequentialism): Zhi-Xuan supplies the technical machinery (multi-objective RL, lexicographic orderings, preferential gaps, unit 7's tool-like locality) by which a Rossian pluralist architecture could actually be implemented. The dissertation's metaethics chapter can cite this as proof that Rossian structure is computationally realizable, not just philosophically attractive.

The ECD model (unit 5) plus 'preferences are constructed from reasons and values' (units 2, 18) is the strongest available philosophical legitimation of the xphi corpus methodology: the folk_ai.db coding scheme captures exactly what Zhi-Xuan et al. say alignment data SHOULD capture - the reasoning behind judgments (reasoning field = their 'justifications for trade-offs'), the values invoked (value_primary = their 'distinct evaluative concepts'), and the stance (their commensuration output) - rather than bare preference orderings. The methodology chapter can claim: where RLHF collects thin preference data, the xphi corpus collects thick, reason-annotated evaluative data - i.e., it already implements the post-preferentist data model the field's leading agenda paper calls for.

Unit 17 ('wise judge, not the clever but unprincipled lawyer' - justifications must guide, not rationalize) is empirically operationalizable with Augustine's existing infrastructure: the folk_ai.db LLM-coding 'reasoning' field plus the LLM moral-reasoning experiment can test whether model-cited reasons track model outputs across controlled variations - a faithfulness study. Combined with Khamassi's strong/weak criteria (LI_2026 unit 9), there is a publishable empirical paper here that directly serves S&K's criterion 2. Possible venue: the Res Practica 'Faces of Responsibility' special issue Howard flagged.

Their fn6 explicitly invites a third kind of reason for accepting AI outputs beyond moral-epistemic and political - and the dissertation's convergentism is a candidate: when Rossian, consequentialist, and contractualist analyses CONVERGE on an output/policy, reasonable dissenters from any one framework have a reason-from-their-own-lights to accept it. That is neither expert-deference (no moral-expertise claim) nor procedural legitimacy (no vote) - it is epistemic corroboration under pluralism. This positions the dissertation as answering S&K's open problem rather than merely citing it. NB: their systematic-error argument (unit 10) also supports convergentism - cross-framework convergence is exactly the kind of error-check that single-source crowd data lacks.

Note for the regulation strand: the EU AI Act's own definition of AI system (unit 1) already includes 'varying levels of autonomy' and post-deployment adaptiveness - the LAW's definition is agentic-ready while the responsibility doctrines applied to it (AILD's narrow output-fault link, unit 10) are not. The mismatch between definitional scope and doctrinal machinery is itself a citable finding for the governance chapter.

RL-EPISTEMIC (units 7-8) is the legal analogue of the knowledge/epistemic condition on moral responsibility (Fischer & Ravizza reasons-responsiveness; Aristotelian ignorance excuses). The philosophically significant consequence: interpretability TOOLS redistribute responsibility - a technical artifact (MI) changes who is morally/legally on the hook. This mirrors and extends Gabriel 2020's tech/normative interdependence thesis (GABRIEL_2020 unit 3): there, architecture constrains which values can be encoded; here, explanation infrastructure determines who bears responsibility. The dissertation can unify both under one claim: technical choices are covert normative allocations.

Unit 11 is the responsibility gap (Matthias 2004) surfacing inside a rigorous legal-doctrinal analysis: after all the difference-making machinery is deployed, the authors concede there may be 'no (human) actor actually in command of the difference-maker' - and their answer is a pointer to game-theoretic methods, not a normative account. Contrast with Luke's move (assign responsibility to the rational AI): Kästner et al. never even consider the AI as a responsibility-bearer; the option space is humans, institutions, or formal reallocation. This is strong evidence for the dissertation's framing: the legal literature has sophisticated tools for DISTRIBUTING responsibility among humans but no answer for the residue case - which is exactly where agentic systems make the residue larger.

Empirical grounding pattern: the agentic-risk claims (units 2, 11, 16) rest on brand-new 2025-2026 empirical work (Kosmyna 'Your Brain on ChatGPT' cognitive debt; Fang et al. longitudinal chatbot study; Kulveit gradual disempowerment). The philosophy is tracking a moving empirical target - supports the dissertation's methodological stance that normative AI ethics now requires live empirical input, contra pure-armchair approaches. Useful for the methodology chapter's defense of the xphi corpus approach.

Intra-value fragmentation (unit 9) is a gift to the dissertation's metaethics chapter: if even ONE value like autonomy dissolves into rival specifications requiring adjudication, then alignment necessarily involves the kind of conflict-resolution machinery Rossian pluralism provides (weighing competing prima facie considerations in context) - and proceduralism cannot help, because the fragmentation recurs INSIDE whatever target the fair procedure selects. The paper's own solution (meta-autonomy = let the user choose per domain) is effectively a delegation of the Rossian weighing task to the user, without noticing that this presupposes the user can do the weighing that the theory could not.

GAP-AGENTIC-UNTESTED, core evidence: Gabriel twice touches the agentic scenario - super-human-speed default decision-making without live intention (unit 8) and system-type-relative principle choice across autonomy levels (unit 21) - but the paper's machinery (six loci, procedural selection) is built for the one-person-one-agent, principal-directs-tool case. Delegated multi-step agency breaks the loci taxonomy's presupposition that there is a live principal whose instructions/intentions/preferences anchor alignment at decision time. Also unit 5: the 2020 claim that reasons-responsive (Kantian/contractualist) capabilities 'extend well beyond most existing artificial agents' is now empirically contestable - Augustine's LLM moral-reasoning experiment speaks directly to it.

Gabriel's proceduralism ('political not metaphysical') and the dissertation's Rossian convergentism are rival responses to the same premise (unit 15: every single theory is counterintuitive or underdetermined). Gabriel: retreat to fair procedure. Augustine: keep object-level normative theory but treat cross-framework convergence as corroboration. The metaethics chapter should stage this as a direct choice and argue convergentism handles Gabriel's domination worry (unit 16) without giving up substantive ethics - overlapping consensus can be the OUTPUT of convergence rather than a substitute for theory. Note also unit 13: Gabriel's own move to 'beliefs about value' is metaethically undertheorized; Gibbard's expressivism supplies the missing account of why attitude-alignment can carry normative weight.

Gabriel's §2 shows the technical/normative interdependence cuts in the dissertation's favor: the question Augustine posed to Howard (2026-06-19 email) - 'which normative framework do we map the design and algorithmic logic to?' - is not framework-neutral, because RL architecture is structurally biased toward consequentialist encodings (unit 3) and resistant to rights-based ones (unit 4). A Rossian-pluralist architecture therefore needs a positive account of HOW prima facie duties get operationalized in optimizing systems - the technical literature will not supply it by default.