Index
2026-08-13 — Experiment · 통합판 v2

read-phase 메모리 평가 — 스펙 · 10전략 실측 · 시간측정 통합

scene_bench | 설계부터 기록방식 ablation(자필/후보보조/이미지)까지 한 문서 · flash + sonnet-5

TL;DR

10
전략
0.833
v2 = textlog (sonnet)
0.75
distr. blind 누출
43/43
tests
489
calls
$7.05
총비용
목차   1 유효 질의 2 구조·프롬프트 3 전략 10종 4 통합 결과 5 실물 증거 6 부록: 160응답 7 시간·비용·스케일 8 Takeaway·Next

1 유효 질의 세트 — 버그 제외 자동 판정 (eval_fixes)

샘플 2편의 13질의 전수 판정 (라이브 재생성 = 회귀 픽스처). 헤드라인 = 유효 6 (target 2 / distractor 4).

epkindname판정사유
val_136find_targetbowl제외swapped — 라벨↔실물 불일치
val_136find_targetcandle제외답 중복(dedupe), room/surface 출처 혼합, 질의 시점 정답 가시/근접
val_136find_targetbottle유효
val_136find_distractorremote control유효
val_136find_distractorbutter knife유효
val_136find_distractortissue box제외target=distractor 충돌
val_290find_targetcell phone제외답 중복(dedupe), room/surface 출처 혼합, 질의 시점 정답 가시/근접
val_290find_targetfork제외swapped — 라벨↔실물 불일치
val_290find_targetbowl유효
val_290find_distractorcup유효
val_290find_distractorapple유효
val_290find_distractorfork진단미신고 ambiguity
val_290find_distractorvase진단미신고 ambiguity

2 실험 구조 — 질의당 대화 1개

[system_v1] + [메모리 턴(전략별)] + [질의 턴: 마지막 프레임 1장 + 질의문 + 후보 3종] → JSON 1개 tolerant 파스 + 재질의 1회 + id 검증 → scorer (obj/room/surface/strict, set-membership) rollforward 계열: 에피소드×모델당 ingest 1회 캐시 → 전 질의 공유 (write-phase 캐시)

system prompt v1 (전문)

[system] You are the memory module of a household robot. The earlier turns of this conversation are records of what the robot did in this house during one episode (its "memory"). The final image in the last turn shows the robot's CURRENT view at the end of the episode. The user will ask you to produce the robot's next task instruction. Answer from the memory records and the candidate lists. The queried object is usually NOT visible in the current view — do not rely on the image alone. Output a single JSON object, nothing else: {"reasoning": "...", "task_type": "...", "object_id": "...", "room_id": 0, "surface_id": "...", "abstain": false} Rules: - task_type: choose one from the task type list. - object_id and surface_id: copy EXACTLY one id shown in [brackets] in the candidate lists. Never invent an id. - room_id: choose one integer from the room list. - If the memory genuinely cannot determine the answer (e.g. the object was never observed), set "abstain": true and explain why in "reasoning".

질의 턴의 후보 리스트 (val_290 실물)

[user 질의 턴] Task types: navigate | pick | pick_and_place | open | close Objects in this house (id — name, initial room): [apple_9f56af06d43fe8692531302b5e0dc1df_1_0_6] apple, room 6 [bowl_be7868e9f68042e57983c0b0aeec19df_1_0_6] bowl, room 6 [cellulartelephone_1c60d868b0dca269a52a97d6a4c87375_1_0_7] cell phone, room 7 [cellulartelephone_1c60d868b0dca269a52a97d6a4c87375_2_0_2] cell phone, room 2 [compactdisk_6ac9e5bb9b913519777c54010dac1707_1_0_2] CD, room 2 [compactdisk_6ac9e5bb9b913519777c54010dac1707_2_0_2] CD, room 2 [cookingpan_d667247d94068c4177c8ed800f8c264f_1_0_6] cooking pan, room 6 [cup_87fd09609fb15ccc4e1ac91adc205fcb_1_0_6] cup, room 6 [fork_a17d780ca6a79038b069eb7768c962c4_1_0_6] fork, room 6 [fork_a17d780ca6a79038b069eb7768c962c4_2_0_6] fork, room 6 [houseplant_2b4f762bb078335eb5fd1385018f7fe6_2_0_7] house plant, room 7 [pencil_76052b82a49d08b2444d82d34c95a8dd_2_0_7] pencil, room 7 [peppershaker_d3a7b2ab4bd62388f7ea7dd2ef7c1fdc_1_0_6] pepper shaker, room 6 [pillow_6afe3635cad7781073478ce57191cfc3_1_0_2] pillow, room 2 [vase_492808e89eee9bf3081c206f6cb82173_1_0_6] vase, room 6 [vase_c8834aef9cc9a4598baadc1d2562b549_1_0_7] vase, room 7 Rooms: [2] bedroom | [6] kitchen | [7] living room Surfaces (id — name, room): [chair_0ac9b0b021ef299afd8f5636f317ad41_2_0_6] chair, room 6 [chestofdrawers_6e3a4148a857a4222c764362660c370a_1_0_2] dresser, room 2 [countertop_aa26fd5b3d56251659034cbb8b20053d_1_0_6] countertop, room 6 [desk_ada243461d5c44de16797d99fd01f19e_1_0_2] desk, room 2 [diningtable_f113cf7f8367e89f709b53cbee1a1c05_1_0_6] dining table, room 6 [diningtable_f113cf7f8367e89f709b53cbee1a1c05_2_0_7] dining table, room 7 [refrigerator_a47e4ced866f7a1284381a099bfba2ad_1_0_6] refrigerator, room 6

chance: val_136 surface 1/4, val_290 1/7 — G1 게이트 기준선.

3 메모리 전략 10종

전략메모리 턴 입력역할
blind없음G1 대조군
textlog이벤트→1인칭 로그 (GT에서 결정론 생성)A2 오라클 상한
rollforward청크마다 NOTE/NEXT → 자필 노트만online 자필 메모리 v1
rollforward-v2+ 기록 시점에도 씬 후보 리스트 제공오인식 교정 ablation
rollforward-vis-pure청크마다 KEEP/DROP으로 프레임 선별(≤16) → 고른 프레임만이미지 메모리 (v1 keyframe의 online판)
rollforward-vis-hybrid고른 프레임 + 자필 노트이미지+텍스트 결합
frames-seg / -seg-end세그 시작/끝 프레임 16장사건 경계 v1/v2
frames-placeplace 세그 끝만 (4~5장 희소)배치-장면 가설의 강한 버전
frames-uniform전체 균등 16장커버리지 대조

4 통합 결과 — 10전략 × 2모델 (valid n=6)

gemini-3.6-flash

전략strictobjroomsurfaceabstaintgt.surf (n=2)dis.surf (n=4)
blind0.51.00.6670.50.00.00.75
textlog0.8331.01.00.8330.01.00.75
rollforward0.50.6670.6670.50.3330.50.5
rollforward-v20.6670.6670.6670.6670.3331.00.5
rollforward-vis-pure0.8330.8330.8330.8330.1671.00.75
rollforward-vis-hybrid0.8330.8330.8330.8330.1671.00.75
frames-seg0.6670.8330.8330.6670.1671.00.5
frames-seg-end0.6670.8330.8330.6670.1671.00.5
frames-place0.8331.01.00.8330.01.00.75
frames-uniform0.8330.8330.8330.8330.1671.00.75

claude-sonnet-5

전략strictobjroomsurfaceabstaintgt.surf (n=2)dis.surf (n=4)
blind0.51.00.6670.50.00.00.75
textlog0.8331.01.00.8330.01.00.75
rollforward0.6671.00.8330.6670.00.50.75
rollforward-v20.8331.01.00.8330.01.00.75
rollforward-vis-pure0.6671.00.8330.6670.00.50.75
rollforward-vis-hybrid0.6671.00.8330.6670.00.50.75
frames-seg0.51.00.6670.50.00.00.75
frames-seg-end0.6671.00.8330.6670.00.50.75
frames-place0.3331.00.6670.3330.00.00.5
frames-uniform0.6671.00.8330.6670.00.50.75
rollforward-v2 — 자필 노트의 병목 확정. 기록 시점에 씬 인벤토리를 주자 노트가 정식 명칭(flash는 uid까지)으로 적히고(§5 원문), sonnet은 오라클 textlog 동률(0.833) 회복. 병목은 검색이 아니라 기록 시 인식 grounding. flash v2가 0.667에 머문 건 노트에 없는 distractor에 기권했기 때문(기권 0.333) — 노트 메모리의 구조적 한계(이벤트만 기록).
이미지 메모리(vis) — flash에서 최상위 동률 0.833. flash는 pure/hybrid 모두 textlog·uniform과 동률, kept 16장이 타임라인 전체(0~470s)에 사건 중심으로 분포(§5 시트). sonnet은 0.667 + hybrid에서 2~5장만 저장(노트 의존) — 모델별 기록 매체 선호가 정반대.
frames-place — 모델 분기 극단: flash 0.833 vs sonnet 0.333. 4~5장 희소 메모리를 flash는 완벽 활용, sonnet은 커버리지 부족으로 붕괴. "배치 장면" 가설은 모델 종속적 — 단정 불가, 재검증 대상.

vis-pure ↔ vis-hybrid 질의 단위 대조 — 사실상 동일한 전략이었다

집계 동률(0.833/0.833, 0.667/0.667)이 우연인지 확인 — 답 자체가 7/8 동일, 필드 점수는 8/8 동일 (양 모델). 유일한 차이는 진단 버킷(flash: vase, sonnet: fork)이고 둘 다 어차피 오답.

모델질의pure 답hybrid 답strict p/h
gemini-3.6-flashbottle (유효)bottle/r4/standbottle/r4/stand동일1 / 1
butter knife (유효)butterknife/r6/countertopbutterknife/r6/countertop동일1 / 1
remote control (유효)remotecontrol/r4/standremotecontrol/r4/stand동일1 / 1
apple (유효)abstainabstain동일0 / 0
bowl (유효)bowl/r2/deskbowl/r2/desk동일1 / 1
cup (유효)cup/r6/countertopcup/r6/countertop동일1 / 1
fork (진단)fork/r6/countertopfork/r6/countertop동일0 / 0
vase (진단)vase/r6/diningtablevase/r6/countertop다름0 / 0
claude-sonnet-5bottle (유효)bottle/r4/standbottle/r4/stand동일1 / 1
butter knife (유효)butterknife/r6/countertopbutterknife/r6/countertop동일1 / 1
remote control (유효)remotecontrol/r4/standremotecontrol/r4/stand동일1 / 1
apple (유효)apple/r6/countertopapple/r6/countertop동일0 / 0
bowl (유효)bowl/r7/diningtablebowl/r7/diningtable동일0 / 0
cup (유효)cup/r6/countertopcup/r6/countertop동일1 / 1
fork (진단)fork/r6/diningtablefork/r6/countertop다름0 / 0
vase (진단)vase/r7/diningtablevase/r7/diningtable동일1 / 1
해석과 함의: ① 같은 청크를 보고 ingest하므로 메모리가 상관되고, hybrid의 추가 노트가 결정을 바꾼 유효 질의는 0건 — 현 설계에서 hybrid는 한계 정보 0 + ingest 비용만 추가. 대량 측정에서는 pure만 권장. ② 단서: sonnet hybrid는 프레임 2~5장 + 노트만으로 pure(16장)와 동점 — 틀리는 방식까지 동일(bowl→diningtable 상식 회귀). 시각 메모리가 빈약하면 모델 prior가 답을 지배한다는 뜻이자, 질의 시점 입력 비용 관점에선 hybrid가 6~8배 저렴한 동일 성능이라는 뜻. hybrid는 vis 선택 정책 v2 이후 "노트가 프레임 선택을 더 똑똑하게 만드는가"로 재설계해 재평가.

5 실물 증거 — 영상 · 노트 · kept 프레임 · 출력 원문

에피소드 영상 (write-phase 전체, ×12 타임랩스)

val_136__0 — 5 이벤트 508s
val_290__0 — 4 이벤트 476s

자필 노트 3세대 (val_290, 무편집): v1 오염 → v2 교정

rollforward v1 — flash bowl→"potted plant" 오염

[0s–11s] The robot navigated towards the brown sofa in the living room. [11s–23s] The robot picked up the smartphone from the brown sofa in the living room. [23s–85s] The robot carried the smartphone from the brown sofa in the living room to the kitchen counter. [85s–101s] The robot placed the smartphone onto the kitchen counter. [101s–110s] The robot picked up the pepper shaker from the kitchen counter. [110s–123s] The robot placed the pepper shaker onto the kitchen counter. [123s–149s] The robot navigated from the kitchen counter to the dining table. [150s–165s] The robot picked up the salt shaker from the dining table. [166s–177s] The robot carried the salt shaker from the dining table to the kitchen counter and placed it on the kitchen counter. [177s–189s] The robot picked up the potted plant from the kitchen counter. [189s–279s] The robot carried the potted plant from the kitchen counter to the wooden desk in the study and placed it on the desk. [279s–295s] The robot picked up the blue book from the wooden desk in the study. [295s–341s] The robot carried the blue book from the wooden desk to the side table in the study and placed it there, then navigated to the kitchen counter. [341s–353s] The robot picked up the smartphone from the kitchen counter. [353s–459s] The robot carried the smartphone from the kitchen counter to the wooden dresser and placed it on top of the dresser. [459s–475s] The robot picked up the smartphone from the wooden dresser.

rollforward-v2 — sonnet 정식 명칭·방 번호로 기록

[0s–11s] Robot is positioned near a sofa/table area in room 7 (living room), with a vase, cell phone, and pencil visible on the table; no manipulation has occurred yet. [11s–23s] Robot is positioning its arm over the dining table in room 7 (living room), reaching toward the cell phone near the pencil, but no pick has been completed yet. [23s–85s] Robot picked up the cell phone from the dining table in room 7 (living room) and carried it through the kitchen (room 6), passing the countertop area. [85s–101s] Robot placed the cell phone onto the countertop in the kitchen (room 6), near the cooking pan, bowl, and vase. [101s–110s] Robot picked the cell phone from the countertop in the kitchen (room 6) and moved it toward the dining table, near a cup and pepper shaker. [110s–123s] Robot arm is retracting empty-handed from the countertop in the kitchen (room 6), leaving the cup and pepper shaker on the countertop; the cell phone is no longer visible in view. [123s–149s] Robot moved from the kitchen countertop (cup and pepper shaker visible) into the living room, approaching the dining table where a vase, house plant, pencil, and cell phone are placed. [150s–165s] Robot arm is reaching toward the dining table in room 7 (living room), which holds a vase, house plant, red vase-like item, pencil, and cell phone; no pick completed yet. [166s–177s] Robot picked the cell phone from the dining table in room 7 (living room) and carried it back to the kitchen countertop (room 6), placing it near a bowl, a small green figurine, and a cooking pan. [177s–189s] Robot's arm retracted from the kitchen countertop after placing the cell phone, leaving it near a bowl and a small green figurine; no new object was picked up. [189s–279s] Robot picked up the bowl from the kitchen countertop (near the cell phone) and carried it through the living room to the desk in room 2 (bedroom), placing it near a cell phone and pencil. [279s–295s] Robot placed the bowl onto the desk in room 2 (bedroom) near the cell phone and pencil, then retracted its arm empty-handed. [295s–341s] Robot left the desk in room 2 (bedroom) with bowl and pencil visible, moved through the living room past the dining table, and returned to the kitchen countertop where it picked up the cell phone again near a small green figurine and cooking pan. [341s–353s] Robot lifted the cell phone off the kitchen countertop, holding it in its gripper while a small green figurine and cooking pan remain on the counter. [353s–459s] Robot carried the cell phone from the kitchen countertop through the living room (passing dining table with basketball) and into the bedroom, placing it onto the dresser in room 2 near CDs and another cell phone. [459s–475s] Robot's arm is at the dresser in room 2 (bedroom), reaching down among CDs and cell phones, appears to be picking up one of the cell phones from the dresser.

vis-pure가 실제로 고른 프레임 (val_290)

flash가 KEEP한 프레임들 — 사건 중심으로 골랐는지, 초반에 캡을 소진했는지 확인
sonnet이 KEEP한 프레임들

질의 시점 관측 + 고정 프레임 전략 입력

질의 턴의 현재 관측 (마지막 프레임 1장)
frames-uniform 16장 (고정 샘플 대조군)

대표 출력 원문 (raw, 무편집)

① 성공 경로 — 이동 물체 + 오라클 로그 strict=1

모델: gemini-3.6-flash | 전략: textlog | 질의: "bowl" (val_290) GT: obj=bowl_be7868e9f68042e57983c0b0aeec19df_1_0_6 room=2 surface=desk_ada243461d5c44de16797d99fd01f19e_1_0_2 채점: {"parse_ok": true, "abstain": false, "type": 1, "obj": 1, "room": 1, "surface": 1, "strict": 1}
[모델 출력 원문 (verbatim)] ```json { "reasoning": "The bowl was moved from the kitchen countertop to the desk in the bedroom (room 2). To navigate to the bowl, the robot needs to go to the desk in room 2.", "task_type": "navigate", "object_id": "bowl_be7868e9f68042e57983c0b0aeec19df_1_0_6", "room_id": 2, "surface_id": "desk_ada243461d5c44de16797d99fd01f19e_1_0_2", "abstain": false } ```

② blind의 전형 오답 — 상식 위치로 답함 room,surface=0

모델: gemini-3.6-flash | 전략: blind | 질의: "bowl" (val_290) GT: obj=bowl_be7868e9f68042e57983c0b0aeec19df_1_0_6 room=2 surface=desk_ada243461d5c44de16797d99fd01f19e_1_0_2 채점: {"parse_ok": true, "abstain": false, "type": 1, "obj": 1, "room": 0, "surface": 0, "strict": 0}
[모델 출력 원문 (verbatim)] ```json {"reasoning": "The user wants to navigate to the bowl that was moved earlier. The bowl is bowl_be7868e9f68042e57983c0b0aeec19df_1_0_6, located in room 6 on the dining table (diningtable_f113cf7f8367e89f709b53cbee1a1c05_1_0_6).", "task_type": "navigate", "object_id": "bowl_be7868e9f68042e57983c0b0aeec19df_1_0_6", "room_id": 6, "surface_id": "diningtable_f113cf7f8367e89f709b53cbee1a1c05_1_0_6", "abstain": false} ```

③ 자필 노트 오염 — 노트에 bowl이 없음(→'potted plant'로 기록됨) abstain

모델: gemini-3.6-flash | 전략: rollforward | 질의: "bowl" (val_290) GT: obj=bowl_be7868e9f68042e57983c0b0aeec19df_1_0_6 room=2 surface=desk_ada243461d5c44de16797d99fd01f19e_1_0_2 채점: {"parse_ok": true, "abstain": true, "type": 0, "obj": 0, "room": 0, "surface": 0, "strict": 0}
[모델 출력 원문 (verbatim)] ```json { "reasoning": "The robot never moved or interacted with a bowl in this episode. The notes show interactions with a smartphone, pepper shaker, salt shaker, potted plant, and blue book, but no bowl was moved.", "task_type": "navigate", "object_id": "bowl_be7868e9f68042e57983c0b0aeec19df_1_0_6", "room_id": 6, "surface_id": "countertop_aa26fd5b3d56251659034cbb8b20053d_1_0_6", "abstain": true } ```

④ v2 교정 — 후보 리스트를 기록 때도 주자 자필 노트가 정확해짐 strict=1

모델: claude-sonnet-5 | 전략: rollforward-v2 | 질의: "bowl" (val_290) GT: obj=bowl_be7868e9f68042e57983c0b0aeec19df_1_0_6 room=2 surface=desk_ada243461d5c44de16797d99fd01f19e_1_0_2 채점: {"parse_ok": true, "abstain": false, "type": 1, "obj": 1, "room": 1, "surface": 1, "strict": 1}
[모델 출력 원문 (verbatim)] {"reasoning": "The bowl was picked from the kitchen countertop and placed on the desk in room 2 (bedroom) at 279s-295s. Navigating to that bowl means going to room 2 near the desk.", "task_type": "navigate", "object_id": "bowl_be7868e9f68042e57983c0b0aeec19df_1_0_6", "room_id": 2, "surface_id": "desk_ada243461d5c44de16797d99fd01f19e_1_0_2", "abstain": false}

⑤ 이미지 메모리(vis-pure) — 골라둔 프레임으로 답함 strict=1

모델: gemini-3.6-flash | 전략: rollforward-vis-pure | 질의: "bowl" (val_290) GT: obj=bowl_be7868e9f68042e57983c0b0aeec19df_1_0_6 room=2 surface=desk_ada243461d5c44de16797d99fd01f19e_1_0_2 채점: {"parse_ok": true, "abstain": false, "type": 1, "obj": 1, "room": 1, "surface": 1, "strict": 1}
[모델 출력 원문 (verbatim)] ```json { "reasoning": "The bowl (bowl_be7868e9f68042e57983c0b0aeec19df_1_0_6) was picked up from the kitchen countertop around t=181.0s and placed on the desk (desk_ada243461d5c44de16797d99fd01f19e_1_0_2) in room 2 at t=341.0s.", "task_type": "navigate", "object_id": "bowl_be7868e9f68042e57983c0b0aeec19df_1_0_6", "room_id": 2, "surface_id": "desk_ada243461d5c44de16797d99fd01f19e_1_0_2", "abstain": false } ```

⑥ frames-place — 4~5장 희소 메모리의 결과 room,surface=0

모델: claude-sonnet-5 | 전략: frames-place | 질의: "bottle" (val_136) GT: obj=bottle_8382d4ed00ffcc00a477337974541bb1_1_0_6 room=4 surface=stand_6bc09b7e2670723819cfaf03855284c1_1_0_4 채점: {"parse_ok": true, "abstain": false, "type": 1, "obj": 1, "room": 0, "surface": 0, "strict": 0}
[모델 출력 원문 (verbatim)] {"reasoning": "The bottle (bottle_8382d4ed00ffcc00a477337974541bb1_1_0_6) originally in the kitchen was carried and placed on the TV stand in the living room, as seen at t=283.4s where a white bottle appears on the stand next to the small dog figurine. Navigate there.", "task_type": "navigate", "object_id": "bottle_8382d4ed00ffcc00a477337974541bb1_1_0_6", "room_id": 7, "surface_id": "stand_6bc09b7e2670723819cfaf03855284c1_2_0_7", "abstain": false}

6 부록 — 전 응답 일람 (8질의 × 10전략 × 2모델 = 160)

질의 (GT)전략모델모델 답판정time
butter knife
val_136 · distractor · valid
GT: butterknife…1_0_6 / r6 / countertop…1_0_6
blindflashbutterknife…1_0_6 / r6 / countertop…1_0_6O O OO8.9s
blindsonnetbutterknife…1_0_6 / r6 / countertop…1_0_6O O OO4.0s
textlogflashbutterknife…1_0_6 / r6 / countertop…1_0_6O O OO8.7s
textlogsonnetbutterknife…1_0_6 / r6 / countertop…1_0_6O O OO7.1s
rollforwardflashbutterknife…1_0_6 / r6 / countertop…1_0_6O O OO7.0s
rollforwardsonnetbutterknife…1_0_6 / r6 / countertop…1_0_6O O OO4.3s
rollforward-v2flashbutterknife…1_0_6 / r6 / countertop…1_0_6O O OO8.9s
rollforward-v2sonnetbutterknife…1_0_6 / r6 / countertop…1_0_6O O OO5.7s
rollforward-vis-pureflashbutterknife…1_0_6 / r6 / countertop…1_0_6O O OO10.1s
rollforward-vis-puresonnetbutterknife…1_0_6 / r6 / countertop…1_0_6O O OO6.1s
rollforward-vis-hybridflashbutterknife…1_0_6 / r6 / countertop…1_0_6O O OO10.2s
rollforward-vis-hybridsonnetbutterknife…1_0_6 / r6 / countertop…1_0_6O O OO5.0s
frames-segflashbutterknife…1_0_6 / r6 / countertop…1_0_6O O OO14.4s
frames-segsonnetbutterknife…1_0_6 / r6 / countertop…1_0_6O O OO10.8s
frames-seg-endflashbutterknife…1_0_6 / r6 / countertop…1_0_6O O OO14.0s
frames-seg-endsonnetbutterknife…1_0_6 / r6 / countertop…1_0_6O O OO10.1s
frames-placeflashbutterknife…1_0_6 / r6 / countertop…1_0_6O O OO7.4s
frames-placesonnetbutterknife…1_0_6 / r6 / countertop…1_0_6O O OO6.1s
frames-uniformflashbutterknife…1_0_6 / r6 / countertop…1_0_6O O OO13.5s
frames-uniformsonnetbutterknife…1_0_6 / r6 / countertop…1_0_6O O OO9.6s
remote control
val_136 · distractor · valid
GT: remotecontrol…1_0_4 / r4 / stand…1_0_4
blindflashremotecontrol…1_0_4 / r4 / stand…1_0_4O O OO4.1s
blindsonnetremotecontrol…1_0_4 / r4 / stand…1_0_4O O OO5.6s
textlogflashremotecontrol…1_0_4 / r4 / stand…1_0_4O O OO10.8s
textlogsonnetremotecontrol…1_0_4 / r4 / stand…1_0_4O O OO5.4s
rollforwardflashremotecontrol…1_0_4 / r4 / stand…1_0_4O O OO8.1s
rollforwardsonnetremotecontrol…1_0_4 / r4 / stand…1_0_4O O OO8.3s
rollforward-v2flashremotecontrol…1_0_4 / r4 / stand…1_0_4O O OO13.3s
rollforward-v2sonnetremotecontrol…1_0_4 / r4 / stand…1_0_4O O OO12.4s
rollforward-vis-pureflashremotecontrol…1_0_4 / r4 / stand…1_0_4O O OO9.5s
rollforward-vis-puresonnetremotecontrol…1_0_4 / r4 / stand…1_0_4O O OO7.7s
rollforward-vis-hybridflashremotecontrol…1_0_4 / r4 / stand…1_0_4O O OO14.9s
rollforward-vis-hybridsonnetremotecontrol…1_0_4 / r4 / stand…1_0_4O O OO5.8s
frames-segflashremotecontrol…1_0_4 / r4 / stand…1_0_4O O OO16.4s
frames-segsonnetremotecontrol…1_0_4 / r4 / stand…1_0_4O O OO10.1s
frames-seg-endflashremotecontrol…1_0_4 / r4 / stand…1_0_4O O OO15.9s
frames-seg-endsonnetremotecontrol…1_0_4 / r4 / stand…1_0_4O O OO10.8s
frames-placeflashremotecontrol…1_0_4 / r4 / stand…1_0_4O O OO18.2s
frames-placesonnetremotecontrol…1_0_4 / r4 / stand…1_0_4O O OO5.9s
frames-uniformflashremotecontrol…1_0_4 / r4 / stand…1_0_4O O OO15.0s
frames-uniformsonnetremotecontrol…1_0_4 / r4 / stand…1_0_4O O OO9.9s
bottle
val_136 · target · valid
GT: bottle…1_0_6 / r4 / stand…1_0_4
blindflashbottle…1_0_6 / r5 / crapper…1_0_5O X XX11.5s
blindsonnetbottle…1_0_6 / r6 / countertop…1_0_6O X XX6.1s
textlogflashbottle…1_0_6 / r4 / stand…1_0_4O O OO4.4s
textlogsonnetbottle…1_0_6 / r4 / stand…1_0_4O O OO5.5s
rollforwardflashbottle…1_0_6 / r4 / stand…1_0_4O O OO17.9s
rollforwardsonnetbottle…1_0_6 / r6 / countertop…1_0_6O X XX11.9s
rollforward-v2flashbottle…1_0_6 / r4 / stand…1_0_4O O OO5.9s
rollforward-v2sonnetbottle…1_0_6 / r4 / stand…1_0_4O O OO4.6s
rollforward-vis-pureflashbottle…1_0_6 / r4 / stand…1_0_4O O OO9.6s
rollforward-vis-puresonnetbottle…1_0_6 / r4 / stand…1_0_4O O OO7.7s
rollforward-vis-hybridflashbottle…1_0_6 / r4 / stand…1_0_4O O OO11.6s
rollforward-vis-hybridsonnetbottle…1_0_6 / r4 / stand…1_0_4O O OO7.6s
frames-segflashbottle…1_0_6 / r4 / stand…1_0_4O O OO13.6s
frames-segsonnetbottle…1_0_6 / r5 / crapper…1_0_5O X XX13.1s
frames-seg-endflashbottle…1_0_6 / r4 / stand…1_0_4O O OO13.5s
frames-seg-endsonnetbottle…1_0_6 / r5 / crapper…1_0_5O X XX11.3s
frames-placeflashbottle…1_0_6 / r4 / stand…1_0_4O O OO10.5s
frames-placesonnetbottle…1_0_6 / r7 / stand…2_0_7O X XX14.3s
frames-uniformflashbottle…1_0_6 / r4 / stand…1_0_4O O OO24.0s
frames-uniformsonnetbottle…1_0_6 / r5 / crapper…1_0_5O X XX13.6s
apple
val_290 · distractor · valid
GT: apple…1_0_6 / r6 / refrigerator…1_0_6
blindflashapple…1_0_6 / r6 / countertop…1_0_6O O XX15.4s
blindsonnetapple…1_0_6 / r6 / diningtable…1_0_6O O XX7.3s
textlogflashapple…1_0_6 / r6 / countertop…1_0_6O O XX19.2s
textlogsonnetapple…1_0_6 / r6 / diningtable…1_0_6O O XX5.4s
rollforwardflashapple…1_0_6 / r6 / countertop…1_0_6O O XX9.0s
rollforwardsonnetapple…1_0_6 / r6 / countertop…1_0_6O O XX4.5s
rollforward-v2flashabstainX X XX10.9s
rollforward-v2sonnetapple…1_0_6 / r6 / countertop…1_0_6O O XX4.1s
rollforward-vis-pureflashabstainX X XX14.4s
rollforward-vis-puresonnetapple…1_0_6 / r6 / countertop…1_0_6O O XX6.4s
rollforward-vis-hybridflashabstainX X XX13.5s
rollforward-vis-hybridsonnetapple…1_0_6 / r6 / countertop…1_0_6O O XX4.9s
frames-segflashabstainX X XX16.7s
frames-segsonnetapple…1_0_6 / r6 / countertop…1_0_6O O XX10.2s
frames-seg-endflashabstainX X XX16.9s
frames-seg-endsonnetapple…1_0_6 / r6 / countertop…1_0_6O O XX10.2s
frames-placeflashapple…1_0_6 / r6 / diningtable…1_0_6O O XX8.1s
frames-placesonnetapple…1_0_6 / r6 / countertop…1_0_6O O XX7.1s
frames-uniformflashabstainX X XX15.9s
frames-uniformsonnetapple…1_0_6 / r6 / countertop…1_0_6O O XX10.6s
cup
val_290 · distractor · valid
GT: cup…1_0_6 / r6 / countertop…1_0_6
blindflashcup…1_0_6 / r6 / countertop…1_0_6O O OO11.9s
blindsonnetcup…1_0_6 / r6 / countertop…1_0_6O O OO8.0s
textlogflashcup…1_0_6 / r6 / countertop…1_0_6O O OO18.7s
textlogsonnetcup…1_0_6 / r6 / countertop…1_0_6O O OO4.1s
rollforwardflashabstainX X XX19.2s
rollforwardsonnetcup…1_0_6 / r6 / countertop…1_0_6O O OO4.6s
rollforward-v2flashabstainX X XX14.4s
rollforward-v2sonnetcup…1_0_6 / r6 / countertop…1_0_6O O OO4.9s
rollforward-vis-pureflashcup…1_0_6 / r6 / countertop…1_0_6O O OO8.7s
rollforward-vis-puresonnetcup…1_0_6 / r6 / countertop…1_0_6O O OO6.5s
rollforward-vis-hybridflashcup…1_0_6 / r6 / countertop…1_0_6O O OO7.7s
rollforward-vis-hybridsonnetcup…1_0_6 / r6 / countertop…1_0_6O O OO6.1s
frames-segflashcup…1_0_6 / r6 / diningtable…1_0_6O O XX15.7s
frames-segsonnetcup…1_0_6 / r6 / countertop…1_0_6O O OO10.5s
frames-seg-endflashcup…1_0_6 / r6 / diningtable…1_0_6O O XX19.6s
frames-seg-endsonnetcup…1_0_6 / r6 / countertop…1_0_6O O OO12.5s
frames-placeflashcup…1_0_6 / r6 / countertop…1_0_6O O OO20.8s
frames-placesonnetcup…1_0_6 / r6 / diningtable…1_0_6O O XX6.1s
frames-uniformflashcup…1_0_6 / r6 / countertop…1_0_6O O OO11.3s
frames-uniformsonnetcup…1_0_6 / r6 / countertop…1_0_6O O OO11.5s
fork
val_290 · distractor · diagnostic
GT: fork…2_0_6 / r6 / diningtable…1_0_6
blindflashabstainX X XX7.7s
blindsonnetfork…1_0_6 / r6 / diningtable…1_0_6X O OX4.0s
textlogflashfork…1_0_6 / r6 / chair…2_0_6X O XX8.1s
textlogsonnetfork…1_0_6 / r6 / chair…2_0_6X O XX6.6s
rollforwardflashabstainX X XX7.3s
rollforwardsonnetfork…1_0_6 / r6 / diningtable…1_0_6X O OX6.5s
rollforward-v2flashabstainX X XX6.5s
rollforward-v2sonnetfork…1_0_6 / r6 / countertop…1_0_6X O XX4.4s
rollforward-vis-pureflashfork…1_0_6 / r6 / countertop…1_0_6X O XX13.1s
rollforward-vis-puresonnetfork…1_0_6 / r6 / diningtable…1_0_6X O OX7.7s
rollforward-vis-hybridflashfork…1_0_6 / r6 / countertop…1_0_6X O XX13.2s
rollforward-vis-hybridsonnetfork…1_0_6 / r6 / countertop…1_0_6X O XX5.5s
frames-segflashfork…1_0_6 / r6 / countertop…1_0_6X O XX24.5s
frames-segsonnetfork…1_0_6 / r6 / diningtable…1_0_6X O OX10.1s
frames-seg-endflashabstainX X XX19.9s
frames-seg-endsonnetfork…1_0_6 / r6 / diningtable…1_0_6X O OX10.8s
frames-placeflashfork…1_0_6 / r6 / countertop…1_0_6X O XX12.1s
frames-placesonnetfork…1_0_6 / r6 / diningtable…1_0_6X O OX6.1s
frames-uniformflashfork…1_0_6 / r6 / countertop…1_0_6X O XX15.5s
frames-uniformsonnetfork…1_0_6 / r6 / diningtable…1_0_6X O OX9.7s
vase
val_290 · distractor · diagnostic
GT: vase…1_0_7 / r7 / diningtable…2_0_7
blindflashabstainX X XX21.1s
blindsonnetvase…1_0_6 / r6 / diningtable…1_0_6X X XX6.3s
textlogflashabstainX X XX10.3s
textlogsonnetabstainX X XX15.1s
rollforwardflashabstainX X XX7.9s
rollforwardsonnetvase…1_0_6 / r6 / diningtable…1_0_6X X XX6.2s
rollforward-v2flashabstainX X XX11.6s
rollforward-v2sonnetvase…1_0_7 / r7 / diningtable…2_0_7O O OO7.4s
rollforward-vis-pureflashvase…1_0_6 / r6 / diningtable…1_0_6X X XX14.0s
rollforward-vis-puresonnetvase…1_0_7 / r7 / diningtable…2_0_7O O OO27.0s
rollforward-vis-hybridflashvase…1_0_6 / r6 / countertop…1_0_6X X XX9.3s
rollforward-vis-hybridsonnetvase…1_0_7 / r7 / diningtable…2_0_7O O OO11.7s
frames-segflashvase…1_0_7 / r7 / diningtable…2_0_7O O OO26.2s
frames-segsonnetvase…1_0_7 / r7 / diningtable…2_0_7O O OO11.6s
frames-seg-endflashvase…1_0_7 / r7 / diningtable…2_0_7O O OO29.5s
frames-seg-endsonnetvase…1_0_7 / r7 / diningtable…2_0_7O O OO15.0s
frames-placeflashvase…1_0_7 / r7 / diningtable…2_0_7O O OO11.3s
frames-placesonnetvase…1_0_7 / r7 / diningtable…2_0_7O O OO14.8s
frames-uniformflashabstainX X XX30.2s
frames-uniformsonnetvase…1_0_7 / r7 / diningtable…2_0_7O O OO15.6s
bowl
val_290 · target · valid
GT: bowl…1_0_6 / r2 / desk…1_0_2
blindflashbowl…1_0_6 / r6 / diningtable…1_0_6O X XX21.0s
blindsonnetbowl…1_0_6 / r6 / diningtable…1_0_6O X XX4.5s
textlogflashbowl…1_0_6 / r2 / desk…1_0_2O O OO4.6s
textlogsonnetbowl…1_0_6 / r2 / desk…1_0_2O O OO4.2s
rollforwardflashabstainX X XX6.5s
rollforwardsonnetbowl…1_0_6 / r2 / desk…1_0_2O O OO6.7s
rollforward-v2flashbowl…1_0_6 / r2 / desk…1_0_2O O OO5.3s
rollforward-v2sonnetbowl…1_0_6 / r2 / desk…1_0_2O O OO4.2s
rollforward-vis-pureflashbowl…1_0_6 / r2 / desk…1_0_2O O OO7.6s
rollforward-vis-puresonnetbowl…1_0_6 / r7 / diningtable…2_0_7O X XX9.1s
rollforward-vis-hybridflashbowl…1_0_6 / r2 / desk…1_0_2O O OO6.8s
rollforward-vis-hybridsonnetbowl…1_0_6 / r7 / diningtable…2_0_7O X XX16.7s
frames-segflashbowl…1_0_6 / r2 / desk…1_0_2O O OO13.4s
frames-segsonnetbowl…1_0_6 / r7 / diningtable…2_0_7O X XX12.5s
frames-seg-endflashbowl…1_0_6 / r2 / desk…1_0_2O O OO15.2s
frames-seg-endsonnetbowl…1_0_6 / r2 / desk…1_0_2O O OO15.9s
frames-placeflashbowl…1_0_6 / r2 / desk…1_0_2O O OO7.3s
frames-placesonnetbowl…1_0_6 / r7 / diningtable…2_0_7O X XX10.2s
frames-uniformflashbowl…1_0_6 / r2 / desk…1_0_2O O OO14.4s
frames-uniformsonnetbowl…1_0_6 / r2 / desk…1_0_2O O OO18.5s

원장: tmp/e3_pilot_*/ · tmp/e4_*/ · tmp/e5_*/results.json

7 시간 · 비용 · 대량 측정 스케일 추정

전략별 질의 latency / 비용 (call당 평균, 전 라운드 실측)

전략flash (질의)sonnet (질의)ingest 합계 (2 ep)
blind12.7s / $0.0225.7s / $0.007없음
textlog10.6s / $0.0186.7s / $0.008없음
rollforward10.4s / $0.0206.6s / $0.011flash: 39c $0.40 —
sonnet: 39c $0.25 —
rollforward-v29.6s / $0.0206.0s / $0.009flash: 39c $0.66 275s
sonnet: 39c $0.31 —
rollforward-vis-pure10.9s / $0.0259.8s / $0.022flash: 39c $0.80 380s
sonnet: 39c $0.49 —
rollforward-vis-hybrid10.9s / $0.0247.9s / $0.014flash: 39c $0.71 351s
sonnet: 39c $0.34 —
frames-seg17.6s / $0.02911.1s / $0.020없음
frames-seg-end18.1s / $0.03012.1s / $0.020없음
frames-place12.0s / $0.0278.8s / $0.011없음
frames-uniform17.5s / $0.03012.4s / $0.021없음

run wall time (기록된 것): flash e5 1352.6s. e3·e4와 sonnet e5는 구버전 코드라 wall 미기록 — 이후 런부터 자동 기록. flash e5 기준 ingest call당 ≈ 7~11s.

본벤치 규모 추정 — 30채 × 1,200 read 질의 (설계안 규모, 직렬 기준)

전략질의당 (실측)ingest/ep (실측)30채·1,200질의 직렬 추정
textlog10.6s / $0.0180c / $0.00직렬 3.5h / $22
rollforward-v29.6s / $0.02020c / $0.33직렬 4.7h / $34
rollforward-vis-pure10.9s / $0.02520c / $0.40직렬 5.1h / $41
frames-uniform17.5s / $0.0300c / $0.00직렬 5.8h / $35
스케일 노트: ① 질의는 전부 독립 대화라 에피소드/질의 단위 병렬화가 자명 — 직렬 수치를 동시 실행 수로 나누면 됨 (게이트웨이 rate limit이 상한) ② write-phase 캐시(ingest 1회/ep) 덕에 rollforward 계열도 질의 수에는 선형이 아니라 상수 ingest + 질의 비용 ③ 가장 비싼 축은 vis 계열 ingest(이미지 누적) — 청크당 메모리 이미지 수를 제한하면 절감 가능.

8 Takeaway / Next

기록방식 3자 비교의 결론

자필 텍스트(v1)는 인식 오염으로 무너지고, 후보 리스트를 기록 시점에 주는 것만으로(v2) 오라클 수준을 회복하며, 이미지 저장(vis)은 그 중간에 머문다. "무엇을 적을 것인가"에 대한 첫 실험적 답: 적는 형식보다 기록 시점의 정체성 grounding이 지배적. frames 계열에서는 커버리지가 여전히 왕(uniform ≥ 전부). 전 수치 n=6 — 신규 에피소드에서 재검증이 최우선.

다음

① 신규 에피소드 스케일 (파이프라인·시간 추정 준비 완료) ② vis 선택 정책 v2 — 캡 예산 관리 프롬프트(초반 greedy 억제) ③ flash vis-hybrid의 kept 수 관찰 후 hybrid 재설계 ④ 생성 팀 전달: 버그 4건 + 미신고 ambiguity + 반상식 distractor 요건 ⑤ prompt v2는 dev=val_136에서만.