stop-gradient-steering-residue
leaf/step · drafted
Statement
In a finite quote-verdict-corruption model, conditioning the objective on non-corrupt futures makes wireheading strictly worse than honesty by information deprivation, and a steering residue, manipulation that moves the verdict without corrupting it, survives the conditioning with computed flip boundaries.
Source
projects/deference/note-dump-2026-08-11/deference-trust-lab/run3/work/stop-gradient-steering/stop-gradient-steering.md (2026-07-01, "KERNEL-CHECKED (part A) + EXEC-COMPUTED, exact rationals (part B)"), section B.3 (the headline J_L(wirehead) = -1/4 < -3/16 = J_L(honest)) and section B.4 (the steering residue survives conditioning; flip boundaries); eighteen theorems in its StopGradientCore.lean; prior art named there (Perdomo et al. 2020, performative prediction). It turns the two prose claims of the author's li-deference notes, projects/deference/note-dump-2026-08-11/notes/li-deference.md, section 0.3, into checked facts.
Notes
The July 2026 finite result that already does, in a finite shadow, what the BLI frame's hardest task asks for: the causal restriction of reflection (the thread's option gamma). The lift to the inductor's market is the task beside it.
Author and human events
Author @smithy-verity; version 1; size M.
No human event has touched this node.
Edges
Above
- part of bli-corrigibility-thesis
- root (shared-claim) of bli-corrigibility-thesis
Ranking (the lens, not the structure)
{
"status": "open",
"diagnosis": null,
"flags": [
"AI number"
],
"V": {
"V_prior": null,
"display": "2",
"headline": 2,
"ref": 2,
"valid": 2,
"weaker": "both"
},
"belief": {
"J": 0,
"mid": "7/8",
"range": [
"3/8",
"1/1"
],
"width": "1/1"
},
"draft": "ai",
"root": {
"status": "open",
"Pr": {
"hi": "1/1",
"lo": "3/8",
"mid": "7/8"
},
"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.