mp-juuuns/evidence-first-research-memory
Evidence-First Research Memory — Synthetic Demo Index This is the small, fully synthetic artifact for the Evidence-First Research Memory portfolio project. 한국어 안내 이 Dataset은 Evidence-First Research Memory 포트폴리오의 작고 완전한 synthetic 데모 artifact입니다. 실제 연구 코퍼스가 아니라, LLM 에이전트가 검색한 뒤 원문 근거를 검증하는 retrieval 구조를 재현하기 위한 공개 예제입니다. Files events.jsonl — four canonical synthetic source records index.sqlite — derived SQLite FTS5 and vector index manifest.json —… See the full description on the dataset page: https://huggingface.co/datasets/mp-juuuns/evidence-first-research-memory.
Evidence-First Research Memory — Synthetic Demo Index
This is the small, fully synthetic artifact for the Evidence-First Research Memory portfolio project.
한국어 안내
이 Dataset은 Evidence-First Research Memory 포트폴리오의 작고 완전한 synthetic 데모 artifact입니다. 실제 연구 코퍼스가 아니라, LLM 에이전트가 검색한 뒤 원문 근거를 검증하는 retrieval 구조를 재현하기 위한 공개 예제입니다.
Files
events.jsonl— four canonical synthetic source recordsindex.sqlite— derived SQLite FTS5 and vector indexmanifest.json— input/index hashes, vector dimensions, and limitationsSHA256SUMS— checksums for the derived files
파일 안내 (한국어)
events.jsonl— 정본이 되는 4개의 synthetic 원본 기록index.sqlite— FTS5와 vector 검색을 위한 파생 SQLite 인덱스manifest.json— 입력·인덱스 hash, vector 차원, 제한사항SHA256SUMS— 파생 파일 무결성 확인용 checksum
Contract
The canonical source is events.jsonl; the SQLite index is derived and can be rebuilt. Search results carry a source path plus an exact line range. An LLM should retrieve candidate evidence, fetch the cited source range, then reason from that source.
사용 계약 (한국어)
정본은 events.jsonl이고 SQLite 인덱스는 재생성 가능한 파생물입니다. 검색 결과에는 원본 경로와 정확한 라인 범위가 포함됩니다. LLM은 후보를 검색한 다음 그 라인의 원문을 가져와 확인하고, 확인된 원문을 바탕으로만 답변해야 합니다.
Scope and limitations
This artifact contains only synthetic records. It contains no private research corpus, original PDFs, credentials, local paths, or model weights. The vectors use signed_hashing_v1, a deterministic portable demonstration backend—not a trained embedding model.
공개 범위와 한계 (한국어)
이 artifact에는 synthetic 기록만 들어 있습니다. 비공개 연구 코퍼스, 원본 PDF, 자격 증명, 로컬 경로, 모델 가중치는 포함하지 않습니다. 벡터는 학습된 임베딩 모델이 아닌 deterministic signed_hashing_v1 portable 데모 backend입니다.
The implementation, rebuild command, tests, and optional MCP adapter are in the linked GitHub repository.
