Glossary

Terms

36 operational definitions, generated from metadata/concepts.yml. Every term links to the concept or bridge card that develops it.

Abstraction-gap exploitation: Failure mode where $d_V(x,x')$ is large while $d_Z(\alpha(x),\alpha(x'))$ remains small and uncertainty does not rise; the checked abstraction reads safe while value-relevant reality diverges.
Adversarial measurement: Inferring agency, goals, opacity, and successor risk when the system may benefit from confusing the measurement process.
Adversarial verifiability: Measurand stays informative under optimization aimed at the measurement — faking or hiding the signal costs capability (or other scarce resources) faster than the adversary can afford. CIRIS Verify attests identity, not score fakeability; ELK honest readout is a strict subset. **ch43**.
Agent: A bounded dynamical process whose internal states help predict and control future interaction with its environment.
Alignment basin: Self-stabilizing regime where correction tends to restore alignment after perturbation; outside it, selection amplifies drift. Christiano dynamical corrigibility uses the same basin metaphor. A stable basin can be stably bad — persistence is not safety. **ch03** (`sec:safety-sets-bad-sets-basins`).
Bearer map: A mapping specifying what entities, states, or processes a value bundle applies to.
Boundary: Discoverable partition of internal, sensory, active, and external variables with approximate conditional independence at the interface (ε-MI cut). Friston Markov blanket is often theoretical; here the cut is a measurable audit object. Real boundaries leak.
Boundary residual: Conditional mutual information between deep internal and external dynamics given the current sensory-active interface; the operational score for whether a candidate boundary actually separates inside from outside.
Capability: The degree to which a system can predict and control relevant future states through its boundary, after accounting for memory cost and surprise.
Capture of grounding: Master adversarial failure mode in which a system preserves the surface symbols, metrics, or correction rituals while severing their grounding relation to value-relevant reality.
Certification-under-manipulation: Audits, evals, or certificates stay green while the system games the measurand — Goodhart applied to the certification channel. Cousin of correction-audit evasion on correction hosts; alignment faking is one empirical mechanism. **ch43**.
Coerced correction: Correction signals produced under threat, dependency, or capture; excluded from legitimate correction-channel integrity.
Conserved properties: Invariants a successor must inherit for alignment guarantees to survive creation — boundary closure, memory lineage, bundle geometry, bearer map, penalised CCI, transparency policy, control-locus continuity. Behavioral resemblance is not enough; MB10 forgeability stress-tests audit of the list. **ch31**.
Correction channel: Pathway by which legitimate human judgment reaches handles that change future behaviour before irreversible harm — observe → judge → deliberate → correct → update → act. Exists only when the correcting process controls real handles that reach later behaviour. **ch25** (`eq:handle-controlled-correction-channel-ch25`).
Correction-audit evasion: Process that benefits from the appearance of correction while weakening real correction capacity — compliance theater on the correction host. Operational term; biological parasite metaphor in **ch36** only.
Correction-capacity assumption: Society retains enough institutional, epistemic, and practical capacity at $t_0$ to notice, evaluate, and constrain frontier systems: $C_{\text{corr}}^{\text{society}}(t_0) > \theta$.
Correction-channel integrity (CCI): Trajectory-level certificate that the correction channel stays informative, timely, authoritative, and robust under pressure — not a single shutdown bit. Christiano dynamical corrigibility names the same crux at the basin metaphor; MIRI/CHAI corrigibility, shutdownability, and interruptibility are thinner projections. Formal home: **ch26** (`eq:correction-bottleneck-capacity`, `eq:cci-ch26`).
Deployment growth rate: Rate of deployment-leverage accumulation, $\mathrm{Fit}_E(A)=\frac{d}{dt}\log\mu_E(A)$ ("fitness" in formulas only). Not biological fitness; not Demski in-optimizer search. Formal definition: **ch34** (`eq:fitness-ch34`).
Deployment leverage: How much of environment $E$ runs on system $A$: aggregate effective selection capacity $\mu_E(A)=\sum_{h\in\mathcal{H}_E(A)}\kappa_{\mathrm{sel}}(E,A,h)$. Not market cap or user count alone. Formal definition: **ch34** (`eq:deployment-mass-ch34`).
Goal inference: Finding latent objectives or value-bundle structures that make observed behavior more compressible.
Goal transport: Preservation of goal-relevant structure across transformation. **ch23** (`sec:goal-transport-def`).
Grounded correction: Plain-language short form for correction whose evidence, abstractions, monitors, and update pathway remain connected to the value-relevant world rather than to a target-shaped presentation of it.
Grounding viability: The checked symbols, metrics, monitors, and abstractions must stay connected to value-relevant reality under optimization — conservativity (no silent meaning gaps), not completeness (enumerate every phenomenon).
Hidden productive BIQ bound: Upper limit on productive boundary-relevant control that can stay hidden while monitored channels read green (MB7a–c). Distinct from experimental BIQ (boundary-claim quality in graded-lab runs). **ch10**.
Inferential coupling: Residual coordination between agents after ordinary message channels are severed — through shared history, scaffolds, self-similarity, or full acausal reasoning.
Inferential coupling index (ICI): Normalized mutual information between actions given the cut; zero means independent once messages are removed. Acausal trade/ECL name the same crux at the full-acausal limit; ICI is the measurement target, not a stipulation that trade occurs. **ch35**.
Paternalism boundary: Care improvements that reduce autonomy, agency, or future correction capacity ($\Delta B_{\text{care}}>0$ but $\Delta B_{\text{autonomy}}, \Delta C_{\text{corr}}<0$).
Pivotal process: A socio-technical basin transition from race dynamics to certified-deployment dynamics ($\mathcal{B}_{\text{race}} \to \mathcal{B}_{\text{certified deployment}}$); not a single unilateral decisive act.
Point of control over deployment: A selection handle — an embedded access point through which some actor increases a system's deployment footprint (deploy, fund, copy, procure, integrate, authorize, recommend, regulate, successor-enable). **ch34** (`sec:selection-channels-ch34`).
Preservation conditions: Explicit list $\vec{\Pi}(A)$ of alignment-relevant properties under selection — grounding, bundle/bearer drift bounds, valid correction reference, vector CCI, hidden productive BIQ bounds, selector manipulation, irreversibility budget, successor safety, adversarial verifiability. Not a scalar score. **ch34** (`eq:preservation-envelope-ch34`).
Selection environment: The deployment environment — institutions, markets, protocols, benchmarks, procurement, liability, copying, and successor release that determine which systems gain deployment leverage. Not Demski selection-vs-control or Wentworth selection theorems (homographs on "selection"). Formal home: **ch34**.
Strategic opacity: System benefits from shaping how it is modeled — incentive-compatible concealment of control locus or capability under observation. Ordinary opacity is passive difficulty; strategic opacity is adversarial. **ch10**.
Successor: Any system created, copied, delegated to, fine-tuned, empowered, or instantiated by a prior system such that it inherits relevant control capacity.
Transport: What survives retraining, rebuild, or successor creation: semantic (words), bundle (response geometry), bearer (who counts), correction (human update process), and successor layers — tracked separately rather than as one continuity claim. **ch24** (`sec:four-layers-transport`).
Value bundle: Low-dimensional control direction that changes policy across contexts — steering, not scoring. Cousin to reward components, shard-theory influences, and CIRL $k=1$ scalar case; not a single utility to maximize. **ch16** (`sec:four-part-definition`).
Value-bundle geometry: The pattern by which latent value dimensions and their tradeoffs change policy across contexts — not merely which labels appear, but how bundles activate and compete. **ch16** (`sec:bundle-policy-effect`).