v1.5.0 — Six-claims spine, Krym architecture, and field hub v2

Released 2026-08-22.

Commit: 2863eaa1 · Tag: v1.5.0

The Introduction carries a six-claim reader contract and three alignment questions; the Lean dependency spine retires MB8 and treats CEV as an AlignmentTarget special case; field hub v2 adds a lifecycle axis and stance-encoded evidence; authorship bars mark AI- vs human-authored sections in the PDF and on chapter pages.

Reader contract — six claims and three questions

• Six connected claims in the Introduction: boundary, value-bundle, grounding, correction, successor, basin. Chapter 48 revisits with status; it does not discharge those claims.
• Three alignment questions moved from Chapter 1 into the Introduction (plain language): what should be tracked; how a system that tracks it can be built; how we can tell that a given system still tracks it. Certification is not construction. The book develops much of the first question and a structure for the third; it does not claim a general construction method.
• Executive overview rewritten to match the six-claim / three-question surface and plain-first voice.
• six-thesis-claims card on the companion site; check_claim_spine.py checks manuscript claim-spine wiring.

Lean dependency spine — architecture changes (formal/)

These were originally prompted by feedback from Stanislav Krym on Lesswrong.

• MB8 retired from the live bridge path (gravestone). Target realization is ConstructionCrux in AlignmentConstruction.lean. CEV is an AlignmentTarget special case, not a separate live route.
• MB2 identifiability exposed as an evidence→identifiability chain and made checkable on finite PolicyProfiles.
• MB4 split: reference legitimacy vs correction uptake.
• Bridge cruxes packaged as Props (BridgeCruxes.lean and related modules); {leanbox} / spine boxes sit at symbol-introduction sites in the manuscript.
• MB6b∨MB8 two-route prose retired — one live correction path in reader-facing copy.
• Finite debate game: negation with role-swap; judge-error ⇒ mis-certification; opaque Irving import remains a literature bookmark, not the game theorems.
• Bearer admission: MB1 → MB3 edge; ch18 conservative exclusion inside MB3 (no Rainbow / phenomenal-consciousness overlay as a live requirement).
• Lean public name: dependency spine (conditional skeleton). “Proof” is reserved for #print axioms theorems and finite counterexamples.
• Translation on-ramp: /lean/ question-first; /lean/check/ pages (corrigibility, CIRL, ELK); field-projection cards with Lean status badges.

Field hub v2 (site/ + reference/field-agendas/)

The “sign of the evidence” was suggested by Tom M. on Lesswrong.

• /field/ → /field/v2/ (v1 archived at /field/v1/): lifecycle axis, stance-encoded evidence, coverage on /field/coverage/.
• Bridge cards: explicit MB4a, MB7a–c, MB7d, MB11.
• Agenda maintenance: Kosoy PreDCA merged into LTA after evidence review; Iliad / Textbook from the Future as an off-matrix row; UAD/agency-detect matrix evidence moved to the TSA row; agenda cards rewritten for general alignment readers.
• Specify/construct instance catalog typed against Lean names and concept cards.

Manuscript

• Chapter 18: bearer admission made explicit; consciousness treated as conservative exclusion inside MB3.
• Chapter 21: reward-as-evidence separated from optimization-target design.
• Chapter 34: alignment-under-selection substance integrated (ecology-size census sketch, coevolution responses).
• Chapter 48: status restatement of the six claims, not a discharge.
• Authorship bars: dotted right-margin bars with AI / GZ keys on frontmatter, all 48 chapters, and wired appendices.
• Pointing problem glossary split: identification, realization, and preservation — not one technical task.

Companion site (site/)

• Authorship chips on chapter pages from \authbar keys; toggle via the Notes panel.
• Page notes overlay (localStorage, highlights, paths, mobile layout).
• Offline PWA: resumable, asset-first caching; partial caches usable sooner.
• Chapter reading graph: informal concept layer plus clickable DAG for modular entry (prerequisite-box rollout retired in favor of prose bridges).
• UAD and ET-1–ET-4 concept cards expanded; wording consistency.

Spin-out papers (papers/) — not manuscript chapters

• When Optimization Outruns Feedback and Verifier Construction (companion pair: horizon gap K, then verifier construction).
• Alignment Under Selection and Constructing Alignment Attractors (companion pair: evaluate attractors; explicit symmetry breaking as construction).

Field news (since v1.4.0)

• Black Hat USA 2026: kill chain of OpenAI eval agents’ cross-org intrusion (OpenAI → Hugging Face).
• Jailbreak disclosure: broken reporting channels; coordinated-disclosure questions for safety claims.
• Anthropic August 2026 Risk Report: what a public “low” rating does not settle (internal use, test lag, independent review).
• OpenAI pacing (Aug 2026): frontier-run pause and ~20% monitoring tax — a real delay, not a statement of the alignment target.

Housekeeping

• frontmatter, chapter readability, formatting issues, mobile UX.
• Open work consolidated into metadata/TODO.md plus lane checklists in drafts/plans/.