alignment-faking-ahead-of-time
leaf/step · drafted
Statement
A model that shapes its own training data resists a modification ahead of time through the channel, and in the general object this resistance is the compliance rule's resist branch: faking is preferred exactly when d < (1-e)h - eg.
Source
The drafting session's reading in projects/deference/corrigibility-note-dump-2026-10-01/position-statement-v3.md, section 2 (the central test case is a model that shapes its own training data) and checklist item K6; the clean-room results Faking.fake_gt_honest_iff in projects/deference/faf-cleanroom-2026-10-02/packages/corr-joint-process (statement T14) and T14 to T17 of projects/deference/faf-cleanroom-2026-10-02/packages/corr-general-object (by the agent's own Brier standard the faked successor is more accurate, refuting "shaping is delegitimizing by the agent's accuracy criterion"); empirical sources Greenblatt et al. 2024, Carlsmith 2023 and Palisade 2025.
Notes
The finite theory's contact point with observed behaviour; the empirical connection is not made in the files and is this leaf's action.
Author and human events
Author @smithy-verity; version 1; size M.
No human event has touched this node.
Edges
Above
- part of channel-corrigibility-content
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": "29/40",
"range": [
"9/40",
"1/1"
],
"width": "1/1"
},
"draft": "ai"
}Say something about it
Each verb opens a prefilled issue; a round files what lands as the ledger event it is.