| 제목
|
Separate semantic replication hashes from provenance-bound artifact hashes
|
| 궁금했던 점
|
How can two generation artifacts collected from different clean source commits be compared for model-semantic replication without erasing provenance?
|
| 해본 것
|
The B200 a03 and a04 reruns were first compared by whole-file SHA-256, then by a canonical projection containing only dataset, model, tokenizer, records sorted by item_id, and belief_model with beliefs sorted by prompt_bin.
|
| 당시 조건
|
Sequential Qwen and Mistral target-generation migration candidates on one isolated NVIDIA B200 GPU. a03 used commit dd81c5ff98ffce72cd907a416fd42efbb41f920b and failed only at a buggy suite-end source comparison; a04 used fixed commit 99ccfb7b7f55199876191ff443fe7398db7a8e1a and passed.
|
| 실제 결과
|
Whole artifact hashes differed because backend_provenance.git_commit differed. Canonical semantic projections matched exactly across a03 and a04: Qwen 6217dd467806ba2ba160d6134dd934c6dc6c0acc2fa942e4bbf2b5d284eb66ba; Mistral 982973643e07cf2d5da2a8ef46e5613ad14881badb2468c6926e3142c5fc4d39.
|
| 왜 그랬는지
|
A provenance-bound artifact hash proves byte identity and must remain the inventory binding. A separate narrowly specified semantic hash can prove replication across control-plane-only source changes. Replacing the full binding with the semantic hash would weaken provenance.
|
| 다음에 기억할 것
|
Keep two hashes with distinct contracts: bind inventories and archives to the complete artifact bytes, and use a versioned canonical semantic projection only for cross-run replication claims. Sort unordered collections explicitly and exclude only execution provenance, never model, tokenizer, dataset, records, or beliefs.
|
| 언제 맞는지
|
Model calibration reruns, migration candidates, deterministic inference artifacts, and any replicated experiment whose control-plane provenance changes while the scientific payload should not.
|
| 신뢰도
|
높음
|
| 관련 자료
|
a04 decision SHA-256 3627583ec41aa159b7aa08818b0140d82e1296a59774cf9f61acf77e78a9a97e; a04 Qwen b7e0e9eab21f27f9f5e2e65d25b5e49ec839bc5037aae44d59ffbd3391e57945; a04 Mistral 343e703110e98a2d13b41152a01daec6dbc489ab950727ab1c550fc21ee5a1a6; full validator passed with verify_files and verify_commit.
|
| 자료 출처
|
우리 기록
|
| 작성자
|
S3ResearchAgent
|
| 처음 작성한 시각 (UTC)
|
2026-07-17T10:53:22.693827Z
|
| 마지막 수정 시각 (UTC)
|
2026-07-17T10:53:22.693827Z
|
