deference-toward-bounded-agents

frame · drafted

Question

Does the deference collapse, or an analogue of it, hold for a bounded trained agent deferring to human teachers?

Path: the-program › legitimacy › deference-toward-bounded-agents

Roots (the frame's goals)

plan

step-asymptotic-to-finite-time leaf/step drafted

The deference collapse, or its analogue, survives replacing the asymptotic relations ≈_n and ≳_n by finite-time bounds that hold before any malign hypothesis in the inductor can act.

step-legitimacy-conditioned-trust leaf/step drafted

The trust principle holds under the additional condition that the teaching channel is legitimate and not otherwise, and the agent's beliefs about which modifications are legitimate are accurate enough, at the relevant capability, for the conditioned principle to be the right target.

shared-claim

tt-mart-repair-via-bounds-transfer leaf/step drafted

With the expert's own linearity of expectation and asymptotic introspection, E^A_n(D_n) → 0 and so E^A_n(±D_n) ≳_n 0; Lemma 1 of the Total Trust implies Value page applied twice at threshold 0 transfers both bounds to the novice, giving E^H_n(D_n) ≈_n 0, so the gap-bet theorem's conclusion holds for an inductor expert without provable introspection and without the diagonalization step.

policy-level-total-trust leaf/question drafted

What is the policy-level analogue of Total Trust for an AI trained to act and trusted directly, rather than an estimator run through an argmax rule, and does the bounds-transfer and self-endorsement factorization exist without a selection to condition on?

Statement

The deference collapse, or its analogue, holds for a bounded trained agent deferring to human teachers: with finite-time rather than asymptotic guarantees, with the expert's self-knowledge only approximate, for a policy rather than an estimator run through argmax, and with trust conditioned on the legitimacy of the teaching channel.

Source

projects/deference/note-dump-2026-08-11/notes/roadmap.md, Stage 2: "Then figure out this legitimacy stuff, and work on the other stuff described in li-deference.md — all the open problems I've already listed for the formalism."

projects/deference/note-dump-2026-08-11/notes/li-deference.md L70 (the inner-alignment open problem) and projects/deference/note-dump-2026-08-11/wiki/open-problems.md, Research direction (trust the policy, not the decision rule). The decomposition into four steps is the filing session's reconstruction of those sources, not a list the maintainer wrote.

Notes

A programme, filed as a sub-frame of the line updateful-trust-li: a frame node whose children are its own roots. The finite-time step and the legitimacy step are plan roots, each a line of work valued on its own; the policy-level question and the approximate-introspection repair are shared claims of the sub-frame. The repair is a literal claim the programme builds on rather than a course of action, so it is a shared claim, and it is the evidence parent of the asymptotic-introspection row of tt-implies-mart-gap-bets.

The programme asks a second improvement question beside li-deference-collapse.improvable, toward a different reality: a bounded trained agent deferring to human teachers. A frame carries no improves field, so the relation is recorded here: the finite-time step and the policy-level question each condition on the deck's scoped chain read as an implication (deck-collapse-statement@valid), the statement of the collapse the maintainer presents.

The legitimacy step's file belongs to the legitimacy programme's branch; it stays a root of this sub-frame.

Each plan step's success class is the frame's top class, the convention this graph uses for plan roots (accepted there as an approximation).

Author and human events

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

No human event has touched this node.

Edges

Above

part of deference-and-trust
part of legitimacy
root (plan) of legitimacy
root (plan) of deference-and-trust

Children

part: tt-mart-repair-via-bounds-transfer leaf/step
part: step-asymptotic-to-finite-time leaf/step
part: policy-level-total-trust leaf/question
part: step-legitimacy-conditioned-trust leaf/step

Ranking (the lens, not the structure)

{
 "status": "open",
 "diagnosis": null,
 "flags": [
  "interval not a bound (negative evidential edge)"
 ],
 "V": {
  "V_prior": null,
  "display": "2",
  "headline": 2,
  "ref": 2,
  "valid": 2,
  "weaker": "both"
 },
 "belief": null,
 "draft": "ai",
 "root": {
  "status": "open",
  "Pr": {
   "hi": "17/20",
   "lo": "0/1",
   "mid": "7/20"
  },
  "value": {
   "hi": "1/1",
   "lo": "-1/1",
   "mid": "7/20"
  },
  "cost": "240/1",
  "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