AMRRA
A mini AI research lab where Retriever, Extractor, Experimenter, and Judge agents work together to power smart search, knowledge extraction, testing, and evaluation in a seamless closed-loop system.
Core trick
Keeps LLM reasoning separate from deterministic statistics: schema-constrained extraction feeds a planner that runs fixed stat tools.
Description
$29
README excerpt
# AMRRA
**Automated Machine Learning Research Reproducibility Assistant**
AMRRA is an agentic research system that turns a research question and source evidence into a traceable reproducibility assessment. The production workflow deliberately separates probabilistic model reasoning from deterministic statistical computation.
## Agent workflow
```text
Source ingestion
↓
Retriever (lexical shortlist + Groq/GPT-OSS semantic reranking)
↓
Extractor Agent (Groq/GPT-OSS, strict schema-constrained evidence + hypotheses)
↓
Experiment Planner (deterministic tool precondition checks)
↓
Deterministic Statistical Toolbox
↓
Judge Agent (Groq/GPT-OSS, evidence-cited synthesis)
```
Groq is the **only production LLM provider**, using `openai/gpt-oss-20b` by default through Groq's OpenAI-compatible Chat Completions API. AMRRA does not silently switch providers. If LLM reranking is temporarily unavailable, retrieval degrades to deterministic lexical ranking; extraction and judging fail explicitly rather than fabricating output.
`openai/gpt-oss-20b` is Groq's production replacement for `llama-3.1-8b-instant`, which Groq shut down for free/developer tiers on 2026-08-16. For compatibility with an older local `.env`, AMRRA maps that one known retired model ID to `openai/gpt-oss-20b` and logs the migration. New configuration should always use the replacement model explicitly.
Groq strict JSON Schema mode is used for probabilistic stage outputs. AMRRA converts each Pydantic