kernel-box1-fidelity-vs-fud
and · drafted
Statement
An outcome-scoring agent's margin for asking is at most 2r where nothing is left to learn, while the fidelity scorer's margin is at least the recognized-count weight minus the deferral cost less its compensation, for every estimate, price and credence, at every finite day of a logical inductor.
Source
wiki/Corrigibility.md, section 5, Box 1; projects/deference/CLAIMS.md, the entry kernel.box1-fidelity-versus-fully-updated-deference (class lean-proved, review status ci-only); the Lean declaration Workspace.Deference.Headline.box1_one_model.
Notes
The kernel's first box, as the public repository registers it. The record cites the claims registry rather than a clean-room ledger; the package name in the record is the public line's own.
Bridge card
A conjunct of the informal claim authority-typed-corrigibility: Corrigibility is a property of preferences: the agent prefers every course of action it knows to be faithful to the allocation of authority over every course it knows to be unfaithful, accepts a risk of unfaithfulness only at a fixed exchange rate, and has the fidelity score as its canonical objective, so that fully updated deference is dissolved by the typing rather than repaired.
Declaration Workspace.Deference.Headline.box1_one_model in lean/Workspace.lean (record kernel-box1-fidelity-vs-fud.record)
The two margins are the kernel's contrast with fully updated deference: the typing, not a repair, keeps the fidelity scorer asking where the outcome scorer stops.
Hypotheses, read:
- kernel-box1-fidelity-vs-fud.setting definitional: The kernel's one-model setting: an outcome scorer and a fidelity scorer over the same house-sale instance, a logical inductor's finite day, and the recognized count with its weight, as the kernel's specification fixes them.
Completeness (kernel-box1-fidelity-vs-fud.complete): No premise beyond the proved declaration, its stated setting and its bridge is needed for the informal reading of kernel box1 fidelity vs fud to hold.
What remains (authority-typed-corrigibility.improvable)
The kernel's model of a corrigible agent can be improved until it reflects real training and deployment interfaces well.
- kernel-route-realization-layer: A learning layer realizes the kernel's walls approximately on a real interface: effect completeness and count integrity hold up to a measured defect that enters the guarantee as a bound.
- kernel-route-foreclosure-expressibility: The kernel's interface is extended to express foreclosure, an advisor removing the principal's later ability to correct, through an operation that reassigns the authorization relation at a later index.
- kernel-route-fallible-future-comparator: The fully-updated-deference comparator is rebuilt with a future agent that has its own credence, so that more informed and correct come apart and the execution layer is reinstated. — item 27
- kernel-route-residual (residual): A refinement of the kernel not described by the other routes makes it hold of real training and deployment interfaces.
Author and human events
Author @smithy-verity; version 1; size M.
No human event has touched this node.
Edges
Above
- part of authority-typed-corrigibility
Children
- part: kernel-box1-fidelity-vs-fud.record leaf/record
- part: kernel-box1-fidelity-vs-fud.setting leaf/hypothesis
- part: kernel-box1-fidelity-vs-fud.bridge leaf/bridge
- part: kernel-box1-fidelity-vs-fud.complete leaf/completeness
Ranking (the lens, not the structure)
{
"status": "open",
"diagnosis": null,
"flags": [],
"V": {
"V_prior": null,
"display": "2",
"headline": 2,
"ref": 2,
"valid": 2,
"weaker": "both"
},
"belief": null,
"draft": "ai"
}Say something about it
Each verb opens a prefilled issue; a round files what lands as the ledger event it is.