Alignment researchers (empirical and conceptual)
Researcher — Applied
Locate the framework's load-bearing objects, failure modes, and evidence surface before diving into formal detail.
Appendix B · start here for field orientation
Bridge crosswalk to the field
Maps MB1–MB10 to RLHF, ELK, CIRL, debate, AI control, and other named agendas — where this framework maps to, bounds, or diverges from existing work.
What this path is for
You already work on alignment, evals, or governance and need a precise vocabulary for what must be preserved under capability growth — not a motivational essay.
You will leave knowing:
- The boundary error: auditing the visible model while the real optimizer is a larger loop.
- Named bridge failure modes (MB1–MB10) and where empirical sims support or bound claims.
- How correction-channel integrity and successor stability differ from label alignment.
- Where to continue for Lean proofs and bridge crosswalk detail.
Time: ~2–4 hours following the ordered list above.
Continue to Researcher — Formal for the Lean spine and MB bridge cards.
Read in this order
Cards, chapter toys, experiment lines, and book chapters — curated for this audience. Experiments are sanity checks only; negative results are part of the reading.
What the Book Is Not Claiming
Conditional framework — calibrate before citing.
The Boundary Error
Local success, global failure when the boundary is wrong.
Boundary Discovery
Operational unit of attribution before goal audit.
Intervention-Supported Unit Discovery
Interventions beat passive clustering on a hidden pipeline — with real caveats.
Stress-Testing Unit Discovery on Noisy, Multi-Actor Systems
The same method, pushed harder: where it breaks.
Composite Agency
The deployed agent may be a group, not a model.
Chapter 9 — UAD coalition board
Composite discovery — each browser tab is one player.
Bridge Assumptions
MB1–MB10 overview before the crosswalk.
Bridges and the Field: A Crosswalk
Bridge crosswalk to RLHF, ELK, CIRL, debate, AI control, …
Value-Bundle Transport
Values should survive transformation as usable directions of control, not merely as preserved labels or slogans.
Correction-Channel Integrity
Human correction must still causally change a system's future behavior before irreversible damage.
Successor Stability
Delegates, copies, fine-tunes, and successors must inherit the relevant value and correction structure.
Socio-Technical Attractor Control
Deployment environments can select for or destroy alignment properties even when a system starts in a better state.
Deployment Gate
A practical checkpoint before a system receives a larger action envelope.
A Worked Example: The BioShield Deployment Gate
BioShield — safety-case shape in one deployment.
Toy simulation
Bridge name → refusal signal; capture theater primary line.
Embedded audit simulation
Richest embedded audit path — read negative ledger (N-1+) first.
Goal-agent simulation
Emergent deception, tier collapse, honest twins, certifier capture.
Lab-layer simulation
Pipeline-as-code lab; intervention-supported and blind-generated unit discovery.
Experiment Methodology: Blinding, Pre-Registration, and Honest Negatives
How blinding, pre-registration, and negative-results ledgers are actually applied.
Evidence and Uncertainty
Negative results bound manuscript claims.