Alignment Workspace

legit-condition-vs-predict-reality

cases · drafted

Statement

The right target for an AI in a legitimate-oracle design is its expectation conditional on legitimacy rather than its prediction of reality with feedback correlated to reality only in legitimate worlds.

Source

projects/deference/note-dump-2026-08-11/notes/li-deference.md, section 0.3, the maintainer: "Perhaps predicting conditional on legitimacy is not quite right. Rather, the AI should be predicting reality, which correlates with feedback in legitimate worlds, and doesn't in illegitimate worlds. Conditioning on legitimacy would make the AI confidently predict legitimacy; we want accurate assessments of risks to legitimacy, too!"

Notes

A crux the maintainer named with both sides; the pictures are the two sides, the global and local forms of the first, the hedged-installation variant from the corrigibility note on value change, and a residual. The first row is conditioned on the clean room's locality result, which is the maintainer's objection as a theorem. It re-ranks the gated-Tower route against the finite-frame transfer as the definition of record, and the corrigibility branch's legitimacy-detection story. Weights are the maintainer's to place; the numbers here are the filing session's.

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": [
  "AI number",
  "given legit-locality-forced",
  "held by legitimacy",
  "no belief",
  "under attack"
 ],
 "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.

comment · challenge · estimate · want