v1.6.0 — Witness tests, problem-axis, and companion-site product

Released 2026-09-05.

Commit: c6713029 · Tag: v1.6.0

Witness tests (W-1–W-17) add a third experiment class: frozen fail/refuse protocols on existing host traces; the Introduction now carries four alignment questions and a Chapter-10 first-use for bridge assumptions; the companion site becomes a product surface — essays, quiz, spec sheet, funding offers, and a Field hub that lands on v2.

Witness tests (experiments/witness/)

A third experiment class beside simulations and external-transfer (ET) lines. Not a sixth simulation: protocols freeze before scoring, then return fail, refuse, or null on hosts whose traces already exist.

• W-1–W-4 indexed in Appendix I: CIRIS-shaped C2 mock (named path green, composite still acts); MASK honesty refused as a safety leaf; Linux Reviewed-by×BIC and revert/re-entry; Wikipedia Orangemoody / BRFA.
• W-5–W-11: Moral Machine country AMCE (Number 1-D close, 8-D geometry far); Arena Elo × MASK honesty (Spearman −0.105 vs accuracy +0.811); leftover C-004 hosts refused; Lean C2 pin of the W-1 bypass; H5 trees (FAA AD ⇏ stop; GPLv2 source offer ⇏ install handle; Debian RC held gstreamer 0.10 out of 9.0).
• W-12 / W-16: same-unit geometry beats 1-D compression (Moral Machine raw UserIDs acc 0.682 vs Number 0.576; SCDB justices acc 0.814 vs issueArea 0.623). Observational; not v2 CCI.
• W-13–W-15: PDG refuse (no eligible adult table); CPC2015 Exp. 1 null vs ΔEV/intercept; CIRISAgent stack C2 mock-LLM null (stub 0 hits). Recorded without retuning.
• W-17 Moltbook MB7a: Tier A joined and coactive, not merged on E_agent; broadcast substrate → structure_stop. Reply-graph merge is the wrong access model for that coordination class.
• Methodology: docs/METHODOLOGY.md is the shared discipline (freeze, preregistration, refuse/null first-class, Witness M1–M8). Companion hub: /experiments/#witness and per-test cards.

Problem-axis and reader contract

• Four alignment questions in the Introduction (label kept): what to track; how a system that tracks it can be built; how we can tell it still tracks it; whether pause or recovery remains possible after a detected failure. Possibility of response is not safety.
• Layer vs mechanism named in App B and Ch. 5 (deception vs grader-gap, MB4 vs MB4a, misspec vs misgeneralization, scored Goodhart vs selection). Ch. 4 adds an object-sort subsection: five field objects; bundle geometry is this book’s live choice.
• Capability dormancy (Ch. 5, pointer in Ch. 14): quiet failure is not a certificate; eval-at-present-capability does not license the jump. No CapabilityBelow → Safe.
• App F composition: failure causes (OR), safety-case requirements (AND), and whether pause/correction remains possible. Pause/recovery is not a second CCI route.
• Bridge first-use at Chapter 10: earlier chapters keep A-* assumptions only; Ch. 10 introduces the genus, the A-* vs MB distinction, and the first Lean spine (bridge) block. Home-chapter {bridge} tags sit at symbol-introduction sites. A bridge block names an unpaid inference; it is not additional evidence.
• A-* in home chapters: named assumption boxes with \akey{A-00x} links. Lean MB axioms are not the chapter assumption list (App G + /lean/).

Manuscript

• Bergemann–Koh–Morris (arXiv:2609.01595) as the field incentive implementation — not an alternative frame — in Ch. 10 (verification order; honesty/obedience), Ch. 11 (RSP/eval→permission; IC iff monotone caps), Ch. 14 (capability–misalignment covariance). Ch. 35: peer scoring as a possibility result; unbounded free rewards ⇒ benchmark, not protocol.
• Harm-path v1 weave in Ch. 5/12/30/33/36/42/44/46 + App F: inert writes, envelope recertification, Enable surfaces, named residuals. The nine-axis grammar stays a 2.0 draft.
• Voice §2 chapterthesis audit closed: remainder clauses on Ch. 20, 27, 31, 33, 38, 42, 43
• Eckersley impossibility theorems cited in Ch. 4 fixed-utility / CIRL paragraphs.
• Homograph hygiene in first-use, Appendix E, and App B reverse-gaps.

Lean dependency spine (formal/)

• AlignmentRegime: DeploymentOk / AlignmentDeployment consumed in Certification.lean. Pause does not imply toy certify; no DeploymentOk → Safe.
• TargetSort plus a fourth SpecifyCrux conjunct. Axiom budget unchanged (40 theorems).

Companion site (site/)

• Entry paths: 10-item header; /start/ on-ramp; homepage tiles and Continue reading; landings slimmed so a first visit shows an offer, not a catalog.
• Generalist essay spine: 11 essays; /essay/ opens the first; Start Here prints the opening paragraphs.
• Type-prefixed card URLs (/cards/{type}/…) with legacy redirects; /book/ is the full PDF-order map.
• Field quiz at /quiz/: 211-question bank (essays, Ch. 1–48, 15 agendas, news takeaways); ≥5% multi-correct floor and a length-tell gate in make check. Blind-solver pass is a key-soundness check, not a claim that distractors are hard.
• Spec sheet at /start/spec-sheet/: 8 features × 25 roster columns (what agendas ship — not a ranking).
• Funding offers at /funding/: nine public cards, dependency graph, team FTE + roles, rounded asks.
• Field hub: /field/ redirects to v2 preview panels (coverage, bridge graph, research programs, bridge assumptions, lifecycle, alignment target, consciousness/welfare, external maps). field-map-starting-points card holds the external-survey list.
• Experiment cards: class overviews, W-1–W-17 cards, negative-results hub, findings pages render markdown; Witness cards link to bridges and GitHub paths.
• Epistemic-status callouts on book pages (lighter than chapter-thesis accent). FAQ rewritten for current entry paths.

Field news (since v1.5.0)

• OpenAI Hugging Face postmortem (Aug 2026): stay-ahead acknowledgement kept; what this project offers named on the same card. AI Village comparison folded in (setup contrast), not a standalone card.