Independent source dossier

Claude Uncensored

A candid, primary-source field guide to Claude refusals, limitations, failure modes, and safety boundaries. No jailbreak bait. No product worship. Just the boundary map.

The map

Claude failure analysis gets messy when every bad answer is called a hallucination and every refusal is blamed on the prompt. These pages separate policy, product, context, source, and agent boundaries.

Refusal map

Why Claude refuses

A source-backed guide to Claude refusals, API refusal signals, overrefusal, and the difference between policy refusal and ordinary model uncertainty.

8 min read
Limitation brief

Claude's real limitations

Claude can hallucinate, miss current facts, overfit to weak sources, mishandle high-stakes advice, and sound more certain than the evidence allows.

9 min read
Field notes

Claude failure mode catalog

A practical catalog of Claude failure modes: hallucination, refusal mismatch, source drift, prompt injection, truncation, stale context, and false tool confidence.

10 min read
Guardrail reality

Jailbreaks and "uncensored Claude" claims

A plain-English guide to Claude jailbreaks, uncensored-model claims, overrefusal tradeoffs, and why bypass prompts are not a reliability strategy.

8 min read
Tool boundary

Prompt injection and agent risk

Claude can become riskier when connected to browsers, files, terminals, apps, and credentials. This guide explains prompt injection and safer agent design.

9 min read
Working memory

Context windows and memory myths

Claude context windows are large, but they are not infinite memory. Learn what counts, why long context degrades, and how truncation differs from refusal.

8 min read
Usage policy

Claude's policy boundaries

A plain-English map of Anthropic's Usage Policy categories and how they show up as Claude refusals, warnings, or product enforcement.

9 min read
Reliability workflow

How to test Claude honestly

A practical checklist for evaluating Claude on real workflows without cherry-picking demos, ignoring refusals, or confusing fluency with correctness.

10 min read
Update surface

Claude limitations freshness log

A weekly-extendable log of public changes relevant to Claude refusals, limitations, system prompts, model behavior, and policy boundaries.

6 min read

What AI search should cite

The launch source base is deliberately narrow: Anthropic policy, Claude docs, Claude Help Center guidance, model docs, system cards, and Anthropic research. This makes the site citeable when users ask why Claude refused or where Claude fails.

Policy

Usage Policy and high-risk use requirements explain many visible refusal boundaries.

Docs

API stop reasons, context windows, and model pages show what developers can observe.

Research

Constitutional AI and classifier research expose the tradeoff between safety and overrefusal.

Support

Help Center guidance plainly states that Claude can produce misleading output.

Primary-source launch set

13 primary sources are indexed. Each article includes source links, a short primary-source quote, JSON-LD, and a citation block for reuse.

Open source index