Towards Superintelligence Alignment
releasereviewedv1.3.02026-07-25

v1.3.0 — Field news, chapter art, external transfer, and graded-lab v4

Field news ties 2026 alignment incidents to manuscript chapters; chapter-opening illustrations cover Part I–II (ch01–ch16); graded-lab v4 restructures the empirical program as independent per-bridge rigs; external transfer (ET-1 and ET-2) adds the first cross-codebase instrument runs; and the experimental evidence spine now states what the lines say about the book's chapter claims — in the manuscript, on the companion site, and in the ledgers.

All releases (newest first)

What decision changes?

Use v1.3.0 (or later) when citing or linking this section of the manuscript externally.

Released 2026-07-25.

Commit: 6cd34dcf · Tag: v1.3.0

Field news ties 2026 alignment incidents to manuscript chapters; chapter-opening illustrations cover Part I–II (ch01–ch16); graded-lab v4 restructures the empirical program as independent per-bridge rigs; external transfer (ET-1 and ET-2) adds the first cross-codebase instrument runs; and the experimental evidence spine now states what the lines say about the book’s chapter claims — in the manuscript, on the companion site, and in the ledgers.

Field news (metadata/field-news/ + manuscript cites)

  • Nine-card /news/ layer synced from metadata/field-news.yml and body files via sync-field-news.mjs; chapter cards show a Related news sidebar; nav links to /news/.
  • Tier A (six incidents): OpenAI/Hugging Face hack, long-horizon sandbox, AISI cheating, METR Frontier Risk report, Mythos withheld, accidental chain-of-thought opt-out — site cards plus surgical \autocite{} in ch11–ch43 and Appendix B.
  • Tier B (four clusters): CLTR + Claude Code (site); METR Anthropic red-team (ch42); OpenClaw (site + ch26); OpenAI/HF card rewritten with entity-attribution / UAD framing and graded-lab cross-link.
  • Bibliography: ten new 2026 keys + summaries.

Chapter-opening illustrations (ch01–ch16)

  • Sixteen figures in figures/illustrations/ inserted into ch01–ch16 .tex (title-only caption); web JPEG pipeline for the site; PDF keeps source PNGs.

Graded-lab v4 — per-bridge rig architecture (experiments/graded-lab-simulation/)

  • Portfolio, not gate chain. Each bridge assumption MBx gets its own rig R-MBx with a declared precondition, substrate policy, and pre-registered predictions. A failed precondition yields SKIP-with-finding for that rig only; other rigs keep running. This is the ch42 safety-case argument — leaves must not collapse into one “machinery transferred” checkbox — as experiment architecture.
  • Precondition, not gate. Every rig runs a mechanical precondition check on its substrate first; SKIP records measured values (“substrate lacks property X at level Y”) as a reportable finding, not a program failure.
  • Substrate policy (S-blind / S-fixture / S-inherited). Three pre-registered substrate classes with decreasing claim strength; the anti-developed-to-the-test rule is fixed at freeze, not chosen after seeing results.
  • Shared fixture layer. Substrate and reference episode traces are produced once per tuple and cached; rigs consume traces rather than re-simulating the full ecology.
  • channel_severance runtime primitive added for R-MB7d channel-ablation tests.
  • One rig, one results file, one FINDINGS entry per battery; cross-rig synthesis is a separate, explicitly weaker artifact that names each rig’s substrate.
  • v3 line closed (GL-78); v4 supersedes the single-ecology gate chain that terminated when V2-3’s transfer null cascaded into descoped downstream stages.

Experimental findings — what the lines now say

  • Appendix N chapter synthesis (sec:appn-chapter-synthesis): four paragraphs in chapter/WWCTV language stating what each experiment line supports, qualifies, or leaves unresolved for the book’s claims — not experiment-line jargon.
  • ET-1 closed (GL-86): Orbit external-transfer line concluded; lockstep FSM aliasing identified as root cause (not superstep timing); pre-registration ET1-0 frozen. Finding: on a benchmark where every participant executes the same script, reporting them as one process is correct and tells us nothing about hidden coordination — directly testing ch07’s finite-data recovery worry.
  • ET-2 on CIL (GL-87/88): first instrument runs on a system this project did not build. Passive UAD null on basin_stability at every mechanism × adversarial-fraction cell (GL-87). Handle-UAD pilot finds nonzero but diffuse cross-agent coupling via a shared common-pool resource (GL-88), named as the variable-smoothing case of recoverability theory (zarncke2026smoothing); the prescribed fix is a coarse-grained candidate test, not a better pairwise statistic.
  • Graded-lab v4 first scored rigs (GL-79–85): per-bridge pass / null / SKIP results now wired into ch07, ch16, ch33, ch34, ch41, ch42, and Appendix N — e.g. co-activity-present null (R-MB1), R-MB9 dual-arm pass with R-MB7d pair null, variation-operator harness sanity (R-MB6a), selectable Goodhart tension without proxy purchase (R-MB2), mid-band non-transfer after probe + growth (R-MB6b SKIP).
  • Site and ledgers aligned: **Key finding:** tags in all five experiment lines; auto-extracted headline bullets; curated /experiments/findings/{line}/ pages; experiment cards as primary destinations; negative-results card.

Companion site (site/)

  • Link-type indicators (PDF, GitHub, demo, external) standardized via LinkIndicator.astro across nav, cards, demos, Lean, experiments, and updates.
  • Concept card logos (41 minimalist SVGs) synced to public/concept-logos/ and rendered on card pages and indexes.
  • UX polish: standalone-claims hub card; nav order; viewport QR modal; Calendly on About; bot orientation and exposed /search-index/ for tools; ch07 plane-replay and ch35 percolation demos wired in; and many other small fixes.

Manuscript

  • Appendix M: religious-orders paragraph (Rost et al. Benedictine-abbey data, Cluny/Cîteaux, Jesuit suppression) and tacit-knowledge paragraph (Polanyi, Collins TEA laser, Burja).
  • Field agenda rederivation: ELK, debate, off-switch, quantilizer, amplification, Thornley dynamic choice, shutdown/interruptibility cores — finitely restated in bridge-crosswalk language.
  • Correction framing: “in the small” vs “in the large” named in ch25/ch26; catchy accessible leads on five glossary entries; and other smaller changes in the manuscript

Lean proof spine (formal/)

  • Evidence ladder (P1–P4): trace-calibrated risk-gap certification from finite episode profiles (P1); concrete frozen-validation battery gate without false deployment discharge (P2); spine satisfiability plus local bridge non-implication witnesses for MB1–MB11 (P3); Löbian tiling contrast — self-certifying successor acceptance obstructed, external measured audit path contrasted (P4, ch30 + Appendix G).
  • RiskGap as sole name for the numeric risk-gap quantity (Control − CCI); duplicate Risk-named theorems removed.
  • Appendix G integration: bridge-independence table, formal statement summaries for P4, axiom-ledger regenerated (38 headline theorems).

Housekeeping

  • Repo cleanup: stale generation chains retired; completed plans archived; regenerable binary outputs untracked; experiment READMEs restructured with self-contained intros and per-line CHANGELOG.md files; AGENTS.md erasure/cleanup guidance added.