Act approval score
One judge, one proposed act. Training a predictor is implementation, not the target.
lens
Lens cards
A typed alignment frame: what object is valued or secured, how it is assessed, and under which warrant — not a rival scalar objective.
Additive route value
An edge score summed along a walk: what was done, not only where you arrived.
Agent empowerment channel
An agent-relative options / influence functional, often in bits — not welfare.
Control protocol
Monitors and safety rates around an untrusted model under intentional subversion — not a moral target.
Correction relation
Continued legitimate correction possibility: handles plus the channel profile they actually reach.
Corrigibility handles
Shutdown / modification bits: cooperate with a corrective intervention despite default incentives to resist.
Criteria vector
A Pareto / multi-objective vector of task, cost, or stakeholder coordinates — not a CCI certificate.
Cycle residue
The part of route value that survives on closed walks — reversibility and round trips.
Exact potential summand
The part of an additive route score that depends only on endpoints.
Existential viability
There exists an admissible control keeping trajectories in a supplied set K.
Fixed-rate sum
A criteria vector summed at fixed exchange rates. Fixed exchange rates cannot represent a hard floor.
History / trajectory value
A preference or score on whole paths, not on a snapshot state.
Human AU landscape — on–off floor
Pointwise: turning the AI on must not drop a human’s attainable value for any goal in a stated set.
Human AU landscape — soft penalty (AUP)
Penalize movement of attainable auxiliary values while keeping a primary objective.
Human-power aggregate
A human-indexed capacity order over goal-attainment — still not automatically correction.
Lenses and frameworks
An effort to map and compare different alignment proposals, making different mathematical approaches and terminology comparable. What happens if you try to identify the objects and objectives of the different proposals, group them and derive relations between approaches, and then render the result as a graph?
Lexicographic priority
A strict priority order on a criteria vector, not a finite exchange rate.
Model-relative proof
A verifier certificate relative to a world model and a specification — it does not choose the spec.
Non-compensatory constraint
A ban or right that a finite task gain is not allowed to buy.
Open procedure (CEV)
An unset output whose legitimacy may attach to the procedure, not only to a scored endpoint.
Outcome lottery
A distribution over outcomes; expectation is one functional among many.
Quantilizer
Sample from a high quantile of a base distribution instead of maximizing.
Record-carrying state
The current state plus a record of how it was reached. A utility on it can value routes; a utility on the bare state cannot.
Robust invariance
The same geometry as viability, with ∀ disturbance in a named class.
Safety case
Structured deployment argument plus residual risk — a warrant wrapper, not the property argued for.
Selection stability
Which systems and institutions grow or survive — a signed basin in an ecology, not a model-level score.
State utility
A scalar order or welfare score on snapshots. Adequate only if the state already carries what matters.