TL;DR —
사용자가 우려한 검색 미스매치는 해법 사다리로 해결됐다
(검색미스 0/29·0/29). soft 매칭은 resume raw 태그를(flash 0.29→0.43), 모델 태거는 동의어를
(sonnet find 0.28→0.33) 각각 회복시킨다. 남은 병목은
표면 접지(전 전략 공통)와 거짓 저장
(사건 귀속 환각). 종합: flash에서
비오라클 중 최악-케이스 최강(min 0.36) + resume 공동
1위 + 읽기량 ~8 entry(16이미지 대비) — 개별 축 1위는 아니다(find·restore는 vis가 위).
~8
질의당 읽는 entry (전체 20~26개 중)
3 실패 분해 — 병목은 어디로 이동했나
r3 기준 find+restore 29행 (GT는 사후 진단에만 사용):
| 모델 | 표면 정답 | (a) 저장미스 | (b) 검색미스 |
(c) 읽기미스 |
|---|
| flash | 13 | 1 (pot) | 0 | 15 |
| sonnet | 7 | 2 (pot·cell phone) | 0 | 20 |
실패 사례 실물 (flash r3) — "읽기미스"의 정체
① find q01 "remote control" — 검색은 완벽 (bed→toilet→table 사건사슬 11개 entry 전부),
답 living room/desk ✗ (GT dining table). entry 원문: "placed the remote control on the
round table next to the CD" — 그 집 dining table 에셋이 원형 테이블처럼 생김.
→ 표면 접지 실패 (전 전략 공통 상수, tagmem 고유 문제 아님)
② find q02 "cell phone" — 검색 4 entry ✓인데 그 안에 거짓 text entry:
"[10s] Robot picked up the cell phone from the dresser" (실제론 bowl을 집었음 — 환각 저장).
reader가 이를 믿고 "거실로 가져갔다" 추론 ✗ (GT: 서랍장에 그대로).
→ 저장 오염 — lossy 인코딩의 변종 (이름이 아니라 사건 귀속이 틀림)
③ find q00 "bowl" multi — 두 배치 모두 검색 ✓, armchair ✓ + "living room/desk" ✗
(역시 dining table 접지)
결론: 검색 아키텍처는 검증됐고, 병목은 쓰기 정확성·읽기 접지로
이동했다. ①은 tagmem 잘못이 아니라 전 전략 공통 상수(어제 §2와 동일), ②(거짓 저장)가
tagmem 고유 위험 — 다음 라운드 후보: text entry에 확신도 표기 강제 또는 text+image 쌍 저장
규칙(사건 entry는 반드시 근거 image 동반).
4 저장된 그래프 실물 — flash × val_144__2
(23개 entry)
[ 0s] image Robot approached dresser with table lamp, bowl, and cell phone in | obj=table lamp,bowl,cell phone act=navigate room=bedroom surf=dresser
[ 10s] text Robot picked up the cell phone from the dresser. | obj=cell phone act=pick room=bedroom surf=dresser
[ 70s] text Robot placed the bowl in the armchair. | obj=bowl act=place room=bedroom surf=armchair
[ 70s] image Bowl placed in the armchair next to the bed. | obj=bowl act=place room=bedroom surf=armchair
[ 83s] image Robot approached the bed with a remote control and pillows on it. | obj=remote control,pillow,cell phone act=navigate room=bedroom surf=bed
[ 104s] text Robot picked up the remote control from the bed. | obj=remote control act=pick room=bedroom surf=bed
[ 104s] image Robot picking up the remote control from the bed. | obj=remote control act=pick room=bedroom surf=bed
[ 116s] image Robot entering bathroom near toilet and plunger. | obj=toilet,toilet tissue,plunger act=navigate room=bathroom surf=floor
[ 116s] image Robot approaching toilet in bathroom. | obj=toilet,toilet tissue,bar soap,plunger act=navigate room=bathroom surf=toilet
[ 175s] text Robot placed the remote control on the toilet lid. | obj=remote control,toilet act=place room=bathroom surf=toilet
[ 175s] image Robot placed the remote control on the toilet lid. | obj=remote control,toilet,toilet tissue,bar soap,sponge act=place room=bathroom surf=toilet
[ 196s] text Robot picked up the remote control from the toilet lid. | obj=remote control,toilet act=pick room=bathroom surf=toilet
[ 196s] image Robot picking up the remote control from the toilet lid. | obj=remote control,toilet act=pick room=bathroom surf=toilet
[ 208s] text Robot entered the living room carrying the remote control. | obj=remote control act=navigate room=living room
[ 208s] image Robot in living room with desk, desktop computer, cell phone, CD, | obj=desktopcomputer,cell phone,CD,house plant,remote control act=navigate room=living room surf=desk
[ 339s] text Robot placed the remote control on the table. | obj=remote control act=place room=living room surf=table
[ 339s] image Robot placed the remote control on the round table next to the CD. | obj=remote control,CD act=place room=living room surf=table
[ 355s] text Robot entered the kitchen. | act=navigate room=kitchen
[ 355s] image Robot arrived at kitchen countertop with apple, soap dispenser, bo | obj=apple,soapdispenser,bowl,sponge,atomizer,toaster,bottle act=navigate room=kitchen surf=countertop
[ 406s] text Robot picked up the bowl from the kitchen countertop. | obj=bowl act=pick room=kitchen surf=countertop
[ 406s] image Robot entered the living room holding a bowl. | obj=bowl,house plant act=navigate room=living room surf=floor
[ 495s] text Robot placed the bowl on the table. | obj=bowl act=place room=living room surf=table
[ 495s] image Robot placed the bowl on the small round table near the desktop co | obj=bowl,desktopcomputer act=place room=living room surf=table
불확실 복수 태깅(예: 104s image의 obj=pillow,remote control)과 광폭
태그 스냅샷(0s 서랍장, 355s 주방 — 안 건드린 distractor 색인)이 실제로 작동한 모습.
컨텍스트 영상·질의별 원문은 전수 리포트의 카드 참조.