ALL MEMOS Download .docx

Key Insight

AWS Bedrock extended its reinforcement fine-tuning (RFT) capability to open-weight models in February 2026: specifically qwen.qwen3-32b and openai.gpt-oss-20b. APIs are OpenAI-compatible, making migration from existing RFT workflows straightforward. Claude variants (Sonnet, Haiku) remain excluded from RFT on Bedrock — confirmed from prior research (intel_bedrock_agentcore_cli_20260503.md). The Bedrock platform also now offers GPT-5.5, Codex, and OpenAI Managed Agents in limited preview (April 2026), confirming Bedrock's neutral multi-vendor positioning.

Action Items

• No immediate TITAN action: DARWIN's self-correction loop uses Claude natively, not Bedrock RFT. Claude fine-tuning is not available via Bedrock.
• If TITAN ever needs a fine-tuned open-weight model for a specialized sub-task (cost reduction on high-volume classification), Qwen3-32B on Bedrock is now a viable path.
• Monitor for Claude RFT support on Bedrock — if Anthropic enables it, it becomes the top DARWIN evolution target.
• SilentInfinity and LiveGrow apps currently use Claude via API; Bedrock access would add IAM-based access control and private VPC routing — worth evaluating for compliance if app scale increases.

Relevance

• AWS Bedrock: The multi-vendor Bedrock harness is now established — any TITAN component assuming Claude-only on Bedrock is safe, but the ecosystem has broadened.
• Active projects: SilentInfinity uses direct API; no Bedrock dependency currently. No immediate risk.