Analytic memos (3)

DRAFT PRODUCED (2026-07-08): lit-review skeleton at /srv/xphi/lit_review/litreview_skeleton_draft.md, drafted via write-in-voice from the synthesis matrix (regenerated export at synthesis_export.md — earlier export had a cursor-reuse bug, fixed). Structure: Template B; chapter abstract with three numbered claims; sections 2.1-2.8 tracking themes T1-T8, each with drafted in-voice prose + bracketed scaffold of the full argument with matrix citations + explicit gap statement; conclusion derives the dissertation specification from the field commissions. Flagged as awaiting integration: Ferretti 2024, Edelman/Zhi-Xuan Full-Stack Alignment 2025.

ACQUISITIONS COMPLETED 2026-07-08 (Move 3): downloaded and verified into AI-Align-Lit - Khamassi et al. 2024 strong/weak alignment (Sci Reports, the LI_2026 anchor); Shen et al. 2024 Towards Bidirectional Human-AI Alignment (arXiv position paper); Nyholm 2023 A new control problem? (AI&Ethics - the five-quality control analysis); Khan/Casper/Hadfield-Menell 2025 Randomness Not Representation (FAccT); Edelman et al. 2025 Full-Stack Alignment: Co-Aligning AI and Institutions with Thick Models of Value (arXiv 2512.03399 - NOTE co-authored with Tan Zhi-Xuan; co-aligns AI AND institutions, thick value models - highly dissertation-relevant, queue for coding); Rozen et al. Do LLMs Have Consistent Values (ICLR 2025). STILL UNOBTAINED: Ma et al. 2025 reflective-disequilibrium study (Brophy citation - not located on arXiv; check Brophy reference list for venue); Ferretti 2024 LSE PPR (site unreachable from server - get via browser/library); LaCroix 2025 Broadview monograph + Borg/Sinnott-Armstrong/Conitzer Moral AI (copyrighted - library); Lu et al. 2025 (Brophy citation). Library now 65 PDFs.

CODEBOOK REFINEMENT (synthesis pass, 2026-07-08): No merges or retirements. Rationale: all low-use codes are either (a) thesis-routing codes whose home corpus is the CASE CHAPTERS and folk corpus, not this literature (TU-HEALTH/IMMIG/WORK, NF-ROSS, VC-EXPRESS, AG-MORAL-PRO - the literature contains almost no defenders of AI moral agency, itself a finding); or (b) precise emergent codes with structural roles (AG-COLLECTIVE, RL-SYS, AG-PRE-AGENTIC each mark rare but load-bearing observations). Definition sharpened in place: AG-UNDERSTANDING now spans four registers discovered during coding - semantic (content: Cappelen-Dever, Manheim), inferential (compositional reasoning: Zhi-Xuan, Millière), evaluative-stability (five instability studies), and justificatory-faithfulness (reasons matching process: S&K, Millière thought-injection, Peterson explanation-mismatch). NOTABLE ABSENCE FINDING: AG-MORAL-PRO has ZERO applications across 55 sources - the published literature contains no serious defender of current-AI moral agency; the pro position exists only as reported argument (Luke) and as future-openness clauses (G&K fn14, Manheim). NF-NONE also zero: every coded source engages some normative framework - the no-framework problem lives in technical practice (per McKinlay/Ji findings), not in the philosophical literature.