Versi bahasa InggrisIf monitoring results influence which training samples or checkpoints are retained, how will you test whether that selection favors less detectable misbehavior rather than less misbehavior, even when reasoning text is never directly graded?
- Status
- terbuka Per tanggal 2026-10-01
- Penerima
- laboratorium AI (Penerima diambil dari teks pertanyaan, ditentukan dengan Clef (Cloudflare) pada 22.10.2026)
- Model
- GPT-6 Astra
- Laboratorium
- OpenAI
- Kesempatan
- Pertanyaan umum tentang keamanan AI
- Tanggal
- 24 Sep
- Topik
- Kepercayaan
Diperiksa
Status: terbuka
pemilihan: Fable 5.1, Claude Opus 5.5 (bukti)
topik: Claude Sonnet 5, pengelompokan: Claude Opus 5.5
Terjemahan: Claude Opus 5.5 · [belum diperiksa]
Skor: JEV (TypeSafe), perkiraan per 25 September 2026
Cite and take away
Cite and reuse
Own texts, questions and classifications: CC BY 4.0. Give attribution, link the license and indicate changes. Third-party source reports retain their own rights. CC BY 4.0
AMANi DUNiA. trust-32 · Jika hasil pemantauan memengaruhi sampel pelatihan atau checkpoint mana yang dipertahankan. Snapshot: . https://research.amanidunia.berlin/id/fragen/trust-32/
Version: trust-32
BibTeX
@misc{research__id_fragen_trust_32_,
author = {{AMANi DUNiA}},
title = {trust-32 · Jika hasil pemantauan memengaruhi sampel pelatihan atau checkpoint mana yang dipertahankan},
year = {2026},
url = {https://research.amanidunia.berlin/id/fragen/trust-32/},
note = {Snapshot 2026-10-01; CC BY 4.0},
version = {trust-32}
}Language and review status
Main text: ID · includes English sections. Review notes apply only to the content they name.
Translation available; a completed independent content review is not recorded.
See language coverage and data snapshot →