NATIVE · CANONICAL_COMPLETE
CUBSO_A_SCALE_QUERY_PROJECTION_LC_V1
Track CUBSO_A_ONLY_SUPERVISED_RETRIEVAL
Pregunta científica
Does a supervised query-only orthogonal alignment improve frozen A/C0 exact retrieval on held-out hard-group DEV while A, C0 and BGE-M3 remain unchanged?
An orthogonal Procrustes transform learned only from TRAIN improves DEV Recall@100 by at least 2 percentage points without reducing Recall@20 or Recall@50.
Cohort
CUBSO_A_SCALE_DEVDEVELOPMENT
Exposure: INFORMED_HYPOTHESIS · Sealed: SEALED
212409444c17fb058878658b886f2557a46e91d5ea93396437604d25c7f645beBaseline
Frozen normalized A/BGE-M3 query embeddings against unchanged frozen C0.
Treatment
Frozen A/BGE-M3 query embeddings transformed by one TRAIN-fitted 1024x1024 orthogonal Procrustes matrix; candidate vectors remain unchanged.
Variable causal
supervised query-side orthogonal Procrustes transform only
Modelo y ejecución
BAAI/bge-m3 + orthogonal Procrustes query transform
5617a9fPASS_REMOTE_HASHES_CANONICAL_DEV_JOIN_9730Runtime: Kaggle Python 3.12 / torch 2.10.0+cu128 / CUDA 12.8 / 2x Tesla T4 · Provider: kaggle:fredericklied/sicoti-cubso-a-scale-procrustes-lc-v1-320c8b7ddb0d
e2595e128a11973bb1c61f59868020a8c9863c18Decisión científica
KEEP on preregistered DEV gate; strong seen-target/service lift, but target-unseen remains the limiting slice and prevents any claim of product readiness.
Regla: KEEP_QUERY_PROJECTION
Resultado: matched preregistered full-data DEV decision gate
External validation: no · Replication required: sí
Métricas
No hay dos métricas primarias con semántica de visualización compatible; los valores exactos permanecen disponibles en la tabla.
| Métrica | Valor exacto |
|---|---|
| DeltaMRR | 0.0779442561684 |
| DeltaRecall@100 | 0.0726618705036 |
| DeltaRecall@20 | 0.0921891058582 |
| DeltaRecall@50 | 0.0821171634121 |
| DeltaTop1 | 0.0695786228160 |
| MRR | 0.392461447202 |
| Recall@100 | 0.748920863309 |
| Recall@20 | 0.630421377184 |
| Recall@50 | 0.704213771840 |
| Top1 | 0.304008221994 |
Tests estadísticos
- cluster_bootstrap_repetitions
- 2000
- cluster_bootstrap_seed
- 20260823
- cluster_count
- 5038
- recall_at_100_losses
- 117
- recall_at_100_mcnemar_exact_p
- 1.378829575001173e-131
- recall_at_100_wins
- 824
- recall_at_20_losses
- 141
- recall_at_20_mcnemar_exact_p
- 2.921761163857805e-169
- recall_at_20_wins
- 1038
- recall_at_50_losses
- 128
- recall_at_50_mcnemar_exact_p
- 4.779730442353226e-150
- recall_at_50_wins
- 927
Incertidumbre
- DeltaMRR
- [0.07112454739012941,0.08515411716572528]
- DeltaRecall@100
- [0.06497392186724642,0.08094869572564607]
- DeltaRecall@20
- [0.08339363462917225,0.10190670160237579]
- DeltaRecall@50
- [0.07366017294265712,0.09164499162235011]
- DeltaTop1
- [0.06178641720999159,0.07749303151237484]
Lineage
Informed by
CUBSO_A_SCALE_FROZEN_DENSE_REPLICATION_V1Same population as
Code bindings
workspaces/conosce/conosce/cubso/experiments/retrieval/a_scale/projection_evaluation.py
PRESERVED_ARTIFACT
c5bfac0ecbadd5f8cd0144f53cc10ac78cb4a325d4a906b6ed37eb5de884c54bworkspaces/conosce/conosce/cubso/experiments/retrieval/a_scale/projection_worker.py
GIT_BLOB
e2595e128a11973bb1c61f59868020a8c9863c187f286db62de0b70fef6483b2ba48b3a7590d47029979bd0408ce4a13a9e69747Artefactos y evidencia
HISTORICAL_CODE_SNAPSHOT
artifacts/derived/procurement_need/registry/control-plane/code-snapshots/c5bfac0ecbadd5f8cd0144f53cc10ac78cb4a325d4a906b6ed37eb5de884c54b.py
c5bfac0ecbadd5f8cd0144f53cc10ac78cb4a325d4a906b6ed37eb5de884c54bLEARNING_CURVE
artifacts/derived/procurement_need/tasks/cubso/experiments/a_scale_query_projection_lc_v1_20260823/providers/kaggle-train/320c8b7ddb0d/results/learning_curve.json
45f246345ba02790c998c4816c684755a6b4fd2817a305d7296ac5e5e5f0c275ORTHOGONAL_PROJECTION
artifacts/derived/procurement_need/tasks/cubso/experiments/a_scale_query_projection_lc_v1_20260823/providers/kaggle-train/320c8b7ddb0d/results/full_projection.parquet
18d8b990a6b072c6b48ed0b34fec183ab527d7435950938f0d52103ad93b6495PAIRED_DEV_EVALUATION
artifacts/derived/procurement_need/tasks/cubso/experiments/a_scale_query_projection_lc_v1_20260823/evaluation.json
e62e8311a0accac219b3665988ac1a8285b2a27c3cc99ac7ce9547eed3b15821PREREGISTRATION
docs/research/cubso/preregistrations/CUBSO_A_SCALE_QUERY_PROJECTION_LC_V1.json
59e03667b9382f4cac2dd3716ae246871fb073e7388195c91fa75dc5cb5ca40cPROVIDER_JOB_SPEC
artifacts/derived/procurement_need/tasks/cubso/experiments/a_scale_query_projection_lc_v1_20260823/providers/kaggle-train/320c8b7ddb0d/dataset/job-spec.json
daacf7b88d4844bed416c38e067078d1a6ef20d631a9b0d7401a5ab3403b478aPROVIDER_JOB_SUMMARY
artifacts/derived/procurement_need/tasks/cubso/experiments/a_scale_query_projection_lc_v1_20260823/providers/kaggle-train/320c8b7ddb0d/results/job_summary.json
43ae78105cc1d5ba71489f8a1fe269726e32f9f52670112f933eaecfa0076177REMOTE_DEV_RANKINGS
artifacts/derived/procurement_need/tasks/cubso/experiments/a_scale_query_projection_lc_v1_20260823/providers/kaggle-train/320c8b7ddb0d/results/full_dev_rankings.parquet
8ba37c7e922649f0ddbce8c8f5176bbfcf2c72aa5a828671f45a3e3e58c2f187TRAINING_INPUT
artifacts/derived/procurement_need/tasks/cubso/experiments/a_scale_query_projection_lc_v1_20260823/candidate_embeddings.parquet
6e663819fb44c92759603a557d8900da04bfae9b75b656fc3bfab048609801c5TRAINING_INPUT
artifacts/derived/procurement_need/tasks/cubso/experiments/a_scale_query_projection_lc_v1_20260823/dev_mapping.parquet
30c25f47791c1758fd738470d19715df27cb7eda931eb17a19d4c3587a6e4356TRAINING_INPUT
artifacts/derived/procurement_need/tasks/cubso/experiments/a_scale_query_projection_lc_v1_20260823/dev_query_embeddings.parquet
0b94e69f3dcb99e9417b477f485d69f0cb4443c3b89049f62a8ad90427f5d398TRAINING_INPUT
artifacts/derived/procurement_need/tasks/cubso/experiments/a_scale_query_projection_lc_v1_20260823/train_mapping.parquet
758f96536f3685fdf12a866cfd81819ec7e47484d52ca15b05b46a45573d1735TRAINING_INPUT
artifacts/derived/procurement_need/tasks/cubso/experiments/a_scale_query_projection_lc_v1_20260823/train_query_embeddings.parquet
acfe4c8e2f50da34e73626c463668d43f5d4bff39d98928fa574d8f42e2b4f83VALIDATED_DEV_RANKINGS
artifacts/derived/procurement_need/tasks/cubso/experiments/a_scale_query_projection_lc_v1_20260823/full_dev_rankings_validated.parquet
5c2e94ee5cf629488c48160d59f78e4ee047ca044d5837e5edcea730110f8ced
Provenance relacionado
Entities
- a-scale:projection:preregistration
- PREREGISTRATION
- a-scale:projection:train-membership
- TRAIN_MEMBERSHIP
Activities
- a-scale:projection:preregistration-activity
- PREREGISTRATION_GIT_BOUND · 6e3fdcbf614d7f3085af677139dec954b60f8d09 · 2026-08-23T22:54:52-05:00 · 2026-08-23T22:54:52-05:00
Edges
- a-scale:projection:edge:prereg-generated
- WAS_GENERATED_BY: a-scale:projection:preregistration → a-scale:projection:preregistration-activity
- a-scale:projection:edge:prereg-used-train
- USED: a-scale:projection:preregistration-activity → a-scale:projection:train-membership