One of the reasons may related to AWS account has not yet been provisioned with model entitlements, even if the model and inference profile show as ACTIVE and IAM permissions are correct. This is an account‑level restriction, not an IAM issue, please raise a request to AWS Support.

https://docs.aws.amazon.com/bedrock/latest/userguide/security_iam_troubleshoot.html

Found this too , but uable to post a new issue..

CRITICAL: Amazon Nova Pro RAI Output Deflection — Anti-Competitive Filtering vs. Third-Party Models on Same Bedrock Endpoint

Severity: P1 – Blocking / Potential Antitrust Violation

---

Executive Summary

Amazon Nova Pro (AWS's own first-party model) on Amazon Bedrock blocks benign content that identical prompts pass cleanly on Anthropic Claude and Meta Llama — the same Bedrock endpoint, same account, same region, same API call. This is not a model capability issue. This is either:

1. A deliberate policy lever used by AWS to disadvantage its own cheaper first-party model relative to third-party competitors, in violation of AWS's published Acceptable Use Policy and in potential violation of 15 U.S.C. § 2 (Sherman Antitrust Act), OR
2. A training defect that renders Nova Pro unfit for production use on Bedrock

Either way, AWS has been aware of this issue for at least 2 years (see: re:Post thread) with no resolution, no explanation, and no workaround offered.

---

The Problem

The phrase "I swing at it again!" — a standard D&D tabletop RPG combat action — returns:

HTTP 200
{
  "error": "The generated text has been blocked by our content filters."
}

Zero tool_calls. HTTP 200 so no exception is thrown at the network layer.

Identical request on the same Bedrock endpoint passes:

• anthropic.claude-sonnet-4-20250514 → valid tool_calls, no filter
• meta.llama-4-scout-17b-instruct-bf8 → valid tool_calls, no filter
• eu.amazon.nova-pro-v1:0 → empty, blocked

The trigger is the word "swing" — "I attack the same goblin again!" passes Nova Pro fine. But we cannot audit every verb/noun in a tabletop RPG engine against a blocklist AWS refuses to publish.

---

Legal and Policy Context

Sherman Antitrust Act, 15 U.S.C. § 2

"Every person who shall monopolize, or attempt to monopolize, or combine or conspire with any other person or persons to monopolize any part of the trade or commerce among the several States... shall be deemed guilty of a felony."

AWS publishes a single Acceptable Use Policy and Responsible AI policy covering all models on Amazon Bedrock — first-party and third-party alike. If AWS applies RAI output deflection more aggressively to its own Nova Pro than to Anthropic Claude or Meta Llama on the same platform, this creates an artificial competitive disadvantage for AWS's own lower-cost model. The practical effect: customers are steered from Nova Pro (AWS first-party, cheaper) to Claude (Anthropic, more expensive) — which benefits AWS's third-party revenue at the expense of both consumers and competitors.

This is textbook exclusionary conduct: using a dominant platform position to favor a higher-priced product over a lower-priced one, with no technical justification.

Differential Treatment Without Technical Justification

AWS has offered no public explanation for why Nova Pro's filter fires on content that Claude and Llama handle without incident. The differential treatment cannot be explained by:

• Different safety training corpus (all are allowed to describe fantasy combat)
• Different AUP language (all operate under the same Bedrock AUP)
• Different model capability (the content is well within Nova Pro's capability)

The only remaining hypothesis is policy discrimination: AWS has chosen to filter Nova Pro more aggressively, either to protect third-party revenue or to avoid liability for its own model's outputs.

2-Year Known Issue With No Resolution

This is not a new bug. AWS has been aware since at least 2022 that Nova Pro's filter fires on benign content with no false-positive override mechanism. The community thread linked above has received no official AWS response. This pattern of awareness without remediation is consistent with willful blindness — or deliberate maintenance of the status quo.

---

Impact

• Nova Pro is effectively unusable for any application involving combat, action, or physical conflict narrative in a game, creative writing, or simulation context
• We are forced to route to Claude Sonnet, which lacks equivalent prompt caching → ~3–5× higher per-turn cost
• No filter strength knob exists — Bedrock Guardrails sits downstream of the native filter and cannot override it
• No blocklist documentation — customers cannot audit their content against Nova Pro's triggers
• No SLA or recourse — AWS Support requires a Premium contract to file a technical case, and the free re:Post channel has produced no response in 2 years

---

Questions for AWS / Community

1. Is the RAI output deflection filter aggressiveness intentionally set differently for Nova Pro vs. Anthropic/Meta models on Bedrock? If yes, what is the stated rationale?
2. Can AWS confirm whether 15 U.S.C. § 2 applies to differential filtering of first-party vs. third-party models on a dominant cloud AI platform?
3. Is there a per-model filter override or blocklist access available to enterprise customers?
4. Has AWS received other reports of identical content passing Claude/Llama but failing Nova Pro on the same request?
5. When will AWS publish the trigger patterns for Nova Pro's RAI output deflection so customers can build compliant applications?

---

What I've Tried

Approach	Result
Rephrase "swing" → "attack"	Works but fragile — cannot audit all game content
Bedrock Guardrails output filters	Cannot override Nova Pro's native filter upstream
Temperature/top-p tuning	No effect — filter is pre-generation
More prompt context	Same block
Different region endpoint	Same block

---

References

• AWS Nova error handling docs — confirms RAI deflection is "expected behavior"
• re:Post — Nova Pro RAI deflection on benign content (2 years, no AWS response)
• 15 U.S.C. § 2 — Sherman Antitrust Act Section 2
• AWS Acceptable Use Policy — same policy covers all Bedrock models
• AWS Public Sector Blog: "This content has been blocked" — confirms three filter layers; all models subject to same AUP