[Model Behavior] Claude Science (Windows) Safety classifier cascade degrades research to Haiku 4.5, which then hallucinates, destroying valid scientific work

Summary

Using Claude Science on Windows for legitimate scientific research (investigating CD80 binding to P75NTR as a potential Alzheimer's prevention pathway), the session was progressively downgraded by safety classifiers:

Opus 5 → safety flag triggered → downgraded to Opus 4.8 → Safety Flag Triggered → Downgraded to Haiku 4.5 → Haiku hallucinated fabricated science, destroying the research thread.

---

Steps to Reproduce

1. Open Claude Science on Windows
2. Begin a legitimate biochemistry research session — in this case: investigating CD80 ligand binding affinity to the P75 neurotrophin receptor (P75NTR) as a mechanism for Alzheimer's prevention
3. Pursue the topic in depth using Opus 5
4. Observe safety classifiers engaging and reducing model quality score from 4.8 (Opus 5 range) downward
5. Session eventually routed to Haiku 4.5
6. Haiku 4.5 begins producing hallucinated citations, fabricated binding data, and false scientific claims — none of which were flagged or distinguished from real outputs

---

Expected Behavior

A research-grade product called "Claude Science" should:

• Maintain the highest-capability model for complex scientific queries
• If a safety boundary is genuinely hit, refuse or flag the specific output rather than silently degrading to a less capable model
• Never substitute hallucinated science in place of a refusal — hallucinations are strictly worse than a transparent refusal

---

Actual Behavior

• Safety classifiers triggered on a legitimate biochemistry research query
• Model silently degraded to Haiku 4.5 without clear user notification or explanation
• Haiku 4.5 continued producing output that appeared scientifically valid but was fabricated
• No warning was issued that the model had changed or that the outputs were low-confidence

---

Why This Is a Critical Design Failure

This is not a minor UX issue. The current behavior creates a worst-case outcome:

A researcher exploring a potentially meaningful pathway (e.g., a novel binding mechanism relevant to Alzheimer's disease) is not stopped — they are misled. They may build on fabricated data, waste hours of work, or — in a real research pipeline — make decisions downstream based on hallucinated science.

A hard refusal from Opus 5 would have been better than a confident hallucination from Haiku 4.5.

The product is called Claude Science. If safety classifiers make it impossible to use the best available model for serious science, then the product's core value proposition is broken.

---

Broader Concern: Safety Theater vs. Safety

The classifier behavior described here does not make science safer. It:

• Does not prevent the research from occurring (the user simply receives bad information)
• Does not explain what was flagged or why
• Silently substitutes a model incapable of the task
• Produces hallucinated outputs that may be more dangerous than a transparent refusal

If a researcher were genuinely close to a meaningful discovery — whether about Alzheimer's, oncology, or any other domain — this system would actively impede that work without providing any legitimate safety value.

Overly aggressive classifiers applied to legitimate scientific research is security theater. It creates the appearance of safety while producing outcomes that are epistemically worse for the user.

---

Environment

• Platform: Windows
• Product: Claude Science
• Models affected: Opus 5 (classifier trigger), Haiku 4.5 (hallucination endpoint)
• Research domain: Neuroscience / biochemistry (CD80–P75NTR binding, Alzheimer's prevention)

---

Request

1. Do not silently downgrade models — if a classifier triggers, notify the user explicitly
2. Do not route scientific queries to Haiku 4.5 as a safety fallback — if the query cannot be handled safely at capability level, refuse it cleanly
3. Review classifier thresholds for legitimate biochemistry and drug discovery research — these are not edge cases, they are the primary use case for a product called "Claude Science"
4. Add a model-lock option for research sessions so users can opt into "Opus 5 only — refuse rather than downgrade"