Precision over novelty
We do not invent a new term where an established security, privacy or ML term already explains the failure.
About the wiki
A public vocabulary for private boundaries.
Context Bleed Wiki exists to make a scattered, ambiguous conversation more precise — without overstating what the evidence can prove.
Independent project
Not affiliated with an AI model provider.
Language: English (UK)
01Editorial principles
How the wiki earns trust.
Evidence labels stay visible
Peer review, preprint, technical guidance and community observation are never presented as equivalent evidence.
Symptoms are not mechanisms
An odd response may motivate investigation, but it does not prove cross-user leakage or identify a system cause.
Corrections remain legible
Material changes receive a dated version note so definitions do not silently drift.
02Evidence ladder
Confidence has levels.
A
Replicated / peer reviewed
Controlled evidence or convergent independent findings.
B
Standards / primary guidance
Authoritative applied security or platform documentation.
C
Technical report / preprint
Useful methods or results that remain open to review.
D
Documented field observation
Reproducible operational evidence with stated limits.
E
Community report
A lead for investigation, not confirmation of mechanism.
03Governance
Open to correction. Resistant to capture.
Submissions may come from researchers, builders, vendors or affected users. Inclusion depends on source quality and relevance, not sponsorship. Future governance will publish named editorial roles, conflicts and a public change log.
Submit evidence or a correction