Theme

Hidden reasoning may be rebuildable from outputs alone

Explore this theme’s evidence over time →

Sources

Attribution and reporting linked by this record; links do not imply independent verification.

Welcome to August 10, 2026 (theinnermostloop.substack.com)

If the reasoning traces labs withhold can be reconstructed from the answers they publish, then withholding them protects nothing. The question of what is extractable from a deployed model becomes the question of whether closed weights mean anything.

Development timeline

Newest reporting dates first. Reporting dates may differ from event dates.

2026-10-05: OpenAI reports disrupting a model-distillation campaign

2026-10-03: OpenAI accuses Moonshot of model extraction

2026-08-16: A watermark hides in the randomness between equally good words

2026-08-12: Researchers learn to extract hidden reasoning traces from three model families

2026-08-10: Research suggests hidden reasoning traces can be rebuilt from outputs alone

Knowledge graphopen in browser →

Cite or correct this record

Citation formatCitation text

Hidden reasoning may be rebuildable from outputs alone
Innermost Loop KB — secondary record
Reporting date: unknown
KB permalink: https://nicholsn.github.io/innermost-loop-kb/themes/extractability-is-existential
Attributed sources:
- Welcome to August 10, 2026 — https://theinnermostloop.substack.com/p/welcome-to-august-10-2026
This citation does not imply independent verification of linked-source claims.

Suggest a correction — opens a GitHub draft for you to review and submit.