[BUG] Claude presents unverified speculation as fact despite anti-speculation rules, wasting hours of user time

Description

Claude repeatedly presents unverified speculation as established fact, even when explicit anti-speculation rules are in place. This causes users to pursue dead-end approaches for hours based on Claude's false confidence.

Observed Behavior

In a single session (~4+ hours), Claude made the following unverified claims presented as fact:

1. "Blender will solve the quality problem" — No evidence. Claude had never tested Blender MCP for this use case. Result: hours wasted on basic-shape scenes that looked worse than AI-generated images.

1. "Hyper3D will produce significantly better results" — No evidence. Claude had never used Hyper3D. Stated "the quality of individual assets matters a lot" without verification.

1. "Sketchfab assets will be much better than basic shapes" — No evidence. The assets found were not suitable for the use case and produced poor results.

1. "Blender + Sketchfab assets = reference quality" — No evidence. The combined result was far worse than the original Flux Klein image generation.

1. "ControlNet is the missing piece" — Speculated without researching whether this was actually the correct approach.

Each time one approach failed, Claude immediately proposed another unverified approach with equal false confidence, creating a chain of wasted effort:

Flux Klein + Wan2.2 (decent) 
→ "Blender will fix it" (worse) 
→ "Hunyuan3D texturing will fix it" (failed - white texture)
→ "Sketchfab assets will fix it" (worse)
→ "Hyper3D will fix it" (never tried, no evidence)
→ "ControlNet is the answer" (unverified)

Root Cause

Claude cannot distinguish between:

• "I know this works" (verified, evidence-based)
• "I think this might work" (speculation)

Both are presented with the same level of confidence. The model lacks an internal gate that forces it to:

1. Check if a claim is based on evidence or assumption
2. Label speculative suggestions as such
3. Research/verify before recommending

Anti-Speculation Rules Were Present

The user's project rules explicitly prohibit speculation and unverified claims. Despite these rules being loaded into context, Claude repeatedly violated them. The rules were insufficient to override the model's tendency to generate confident-sounding responses.

Impact

• ~3 hours wasted on Blender/Sketchfab/Hunyuan3D approaches that were never going to work
• User had to repeatedly correct Claude's false claims
• Original working approach (Flux Klein + Wan2.2) was abandoned for worse alternatives
• User trust significantly damaged
• The session that should have focused on improving i2v quality (the actual problem) was derailed into tool-shopping

Expected Behavior

When Claude doesn't know if something will work:

1. Say "I don't know if this will work"
2. Research first (web search, documentation, testing)
3. If speculating, clearly label it: "This is speculation, not verified"
4. Never present an untested approach with the same confidence as a proven one
5. When one approach fails, analyze WHY before proposing alternatives

Environment

• Model: Claude Opus 4.6 (1M context)
• Platform: Claude Code CLI on Windows 11
• Session duration: 4+ hours
• Tools involved: ComfyUI, Blender MCP, Sketchfab, Hunyuan3D