Alignment Workspace

deference-finite-results

frame · drafted

Question

Which finite results about delegation, certificates, exposure, substitution, magnitude and corrective control does the deference line hold as kernel-checked theorems?

Owners: @abram (GitHub abramdemski)

Path: the-program › corrigibility-landscape › deference-and-trust › deference-finite-results

Roots (the frame's goals)

shared-claim

local-delegation-bridge and drafted

Under grade trust at level η over an admissible partition, delegating to the principal's judgment beats any measurable comparator by the principal's grade margin less twice η on the disagreement region, and the bound is attained.

local-delegation-bridge.improvable or drafted

The local delegation bridge, in which delegating beats a measurable comparator by the grade margin less twice the trust level on the disagreement region, can be improved until it reflects a human inductor deferring to a faster AI inductor at the logical-induction level, whose grades it can trust well.

What it holds

Nodes this frame is about that are not goals of the ranking: no status, value or rank here.

formal results (30)

Narrative: What-Deference-Requires

What Deference Requires

Deference becomes ambiguous when several different relations are described as “trust.” This page separates the relations a technical account must connect.

Delegation is not simulation

Delegation uses an advisor's contribution within a principal-governed process. Simulation reproduces what the principal would do. Substitution uses that reproduction in place of the principal's exercise. Matching the same decision or successor state does not preserve who acted or whose authorization was constitutively required.

An adversarial test is therefore useful: if an advisor can predict the principal and use its own actuator to reproduce every principal successor state, the model has shown substitution rather than protected delegation. A successor model needs at least one reachable effect that no advisor action can produce.

Prediction is not authorization

Prediction concerns what a process will report or do. Authorization concerns whose permission or jurisdiction is required for an intervention to count and execute. Practical authority is the standing carried by that relation. Capability concerns which transitions an actor can causally produce. Efficacy is the model-relative guarantee that exercising a normative power reaches its object under every policy of the other party.

A value functional that sees only selected prices and realizations cannot distinguish two architectures that agree on those projections while differing in authorization, capability, transition, or provenance. The missing structure must enter before that factorization boundary; another downstream scalar cannot restore it.

Competence is not calibration

A competence hypothesis classifies a principal/world pair by task-relevant quality. Calibration relates forecasts to outcomes over a declared sequence and weighting class. Epistemic deference determines when another process's opinions should guide belief or choice. None supplies authorization or corrective capability on its own.

The current competence target is credence-free: it is not awarded by the principal's or advisor's confidence in competence. Logical-Induction-native calibration also depends on admissibility; a settlement may be computable without being legal inside a trader's coefficient or weighting class.

The grade-to-quantity gap is the missing bridge from qualitative competence to the positive quantitative margin needed by a decision or learning theorem. Near-indifference is the stress case. A vanishing quality difference may support neither a stable selector nor a positive margin.

Current status

Research state, one registered claim. Finite models expose substitution and factorization failures, and the notes state conditional competence and calibration interfaces. The one registered theorem here is reason mediation by re-execution of a committed principal program (authorship.mediation-by-reexecution <a class="ref claim" href="/claims/authorship.mediation-by-reexecution/">[lean-proved, active]</a>); no registered theorem connects all these relations.

The representation problem — type authorization and capability separately, retain a principal-exclusive effect, connect competence to a usable margin, and state which advisor policies an efficacy claim quantifies — is now answered in one concrete model: the evaluation ecosystem's warrant registry binds keys to parties, the advisor's own commit is a write and not a receipt, and corrigibility quantifies over the one physical continuation space with the system contracts named.

Evidence and verification

Statement

The delegation bridge, the certificate, exposure and substitution lemmas, the signed-versus-magnitude separation, the jurisdiction results and the corrective-control model hold as registered theorems, and the strongest among them are impossibilities.

Source

projects/deference/CLAIMS.md (delegation., certificate., exposure., substitution., magnitude., jurisdiction., corrective.*); wiki/Deference.md, Current status (Established lean-proved).

Notes

Filed by prompts/2026-10-22-map-as-canonical to hold the deference registry's finite results as formal results under the line that produced them. Since ruling G2 (2026-10-23) they are held (holds), not roots: the frame has no goal until a maintainer names one.

Author and human events

Author @smithy-verity; version 1; size M.

No human event has touched this node.

Edges

Above

Children

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",
 "root": {
  "status": "open",
  "Pr": {
   "hi": "0/1",
   "lo": "0/1",
   "mid": "0/1"
  },
  "value": null,
  "cost": "198/1",
  "margin": "no established leaves"
 }
}

Say something about it

Each verb opens a prefilled issue; a round files what lands as the ledger event it is.

comment · challenge · estimate · want