rs-wireheading-through-pushers
leaf/step · drafted
Statement
Under unconditional reflection on next-day belief states, a logical-induction agent's evaluation of an action weights future value beliefs by which pushes the action elicits, so actions that elicit favourable pushes from the overseers score highly and the agent cannot represent a push as caused by its own deception.
Source
The BLI thread's Consequence 3 (projects/deference/faf-cleanroom-2026-10-02/sources/corrigibility/imported-chats/2026-09-14__bli-thread-conditioning-vs-prior-overrides__claude-ai-paste.md: Armstrong's motivated value selection and Everitt's reward tampering in logical-induction clothing); projects/deference/corrigibility-note-dump-2026-10-01/corrigibility-bli-thread-and-early-questions-audit.md, section 1.5 and row 2.7 (the mapping to run 2's retention criterion; absorbing prior overwrites into conditioning has a price unless reflection is restricted along causal lines).
Notes
Threatens: the BLI formulation of the thesis. Its fix, the causal restriction of reflection, exists in a finite non-LI model (the July 2026 stop-gradient result) and not in Bayesian logical induction; the lift task is the given parent. A risk story the corrigibility work itself produced: it shows a probability and a status and enters the comparison through the plans whose tables cite it. No outcome class is tagged (tagging a failure class is a human judgment under the protocol's rule on classes), so its failure mass goes to the default trajectory until a human tags it.
A root of corrigibility-landscape, the sub-frame whose plans it threatens; the top frame keeps only stories at the level of its own question.
Author and human events
Author @smithy-verity; version 1; size M.
No human event has touched this node.
Edges
Above
- part of corrigibility-landscape
- given of lift-stop-gradient-to-bli
- root (risk-story) of corrigibility-landscape
- given lift-stop-gradient-to-bli
Ranking (the lens, not the structure)
{
"status": "open",
"diagnosis": null,
"flags": [
"AI number",
"given lift-stop-gradient-to-bli",
"interval not a bound (negative evidential edge)"
],
"V": {
"V_prior": null,
"display": "2",
"headline": 2,
"ref": 2,
"valid": 2,
"weaker": "both"
},
"belief": {
"J": 0,
"mid": "7/20",
"range": [
"0/1",
"17/20"
],
"width": "1/1"
},
"draft": "ai",
"root": {
"status": "open",
"Pr": {
"hi": "7/10",
"lo": "0/1",
"mid": "473/1600"
},
"value": null,
"cost": "85/1",
"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.