Avyra Research Console
Avyra CUBSO
CUBSO
Catálogo CUBSO · read-only
Modelo de identificación de necesidad
Investigación para identificar el ítem CUBSO que representa con mayor precisión una necesidad de contratación pública.
Catálogo
CUBSO
Modo
Solo lectura
Fuente
Control Plane V3
Qué estamos viendo
28 experimentos nativos, 16 legacy, 10 cohorts y 1 cierre de línea.
Qué sabemos
CUBSO_A_SCALE_PROCRUSTES_BGE5K_RRF_DEV_FOLLOWUP_V1 was evaluated only on the preregistered DEV development cohort; TEST and TEMPORAL_TEST remain sealed.
Qué falta demostrar
Requiere replicación antes de ampliar la afirmación.
Research pulse
Volumen de registros científicos del programa. Esta comparación describe cobertura del Control Plane, no calidad relativa.
CUBSO_A_SCALE_PROCRUSTES_BGE5K_RRF_DEV_FOLLOWUP_V1Evidencia de decisión
CUBSO_A_SCALE_PROCRUSTES_BGE5K_RRF_DEV_FOLLOWUP_V1 was evaluated only on the preregistered DEV development cohort; TEST and TEMPORAL_TEST remain sealed.
CUBSO_A_SCALE_PROCRUSTES_BGE5K_RRF_DEV_FOLLOWUP_V1_DECISION_1Persisted lineage
Antecedentes directos registrados que informan al champion actual.
CUBSO_A_SCALE_BGE_M3_DENSE_FINETUNE_5K_V1
INFORMED_BYCUBSO_A_SCALE_PROCRUSTES_BGE5K_RRF_CALIBRATION_V1
INFORMED_BYCUBSO_A_SCALE_QUERY_PROJECTION_LC_V1
Decisiones científicas
Distribución de decisiones persistidas en experimentos native y legacy.
Recall@100
11 observaciones registradas.
No hay una serie comparable para este indicador.
Avyra sólo conecta puntos cuando coinciden la métrica, cohorte, versión de cohorte y política de evaluación. El Control Plane actual no contiene dos observaciones de Recall@100 con esa identidad exacta.
Decision-backed findings
Resúmenes de decisiones persistidas. No son inferencias nuevas del dashboard.
The preregistered operational gate retained frozen base BGE-M3 before any trained quality comparison because the measured full-training projection exceeded the meeting budget. Downstream reranking/selective-classification evidence is development-only; TEST/TEMPORAL_TEST remain sealed and the 2023 pilot is unlabeled pending blind review.
Both authorized executions failed before a complete target-free ranking. The terminal batch-8 attempt remained an execution-only amendment and also failed with CUDA OOM. No conclusion about mLateOn retrieval quality is authorized, and no further memory/search retuning is allowed inside this preregistered experiment.
This paired development ablation holds the frozen occurrence cohort, query embeddings, model revision, exhaustive cosine scorer and exact 85,711 candidate codes fixed while changing only candidate-side representation. Missing NEED-IR views remain misses in both arms. TEST and TEMPORAL_TEST remain sealed.
Lo que sabemos ahora
Conclusiones recientes, mostradas desde el rationale persistido de cada decisión. Avyra no genera una conclusión paralela.
The preregistered operational gate retained frozen base BGE-M3 before any trained quality comparison because the measured full-training projection exceeded the meeting budget. Downstream reranking/selective-classification evidence is development-only; TEST/TEMPORAL_TEST remain sealed and the 2023 pilot is unlabeled pending blind review.
CUBSO_ORDER_TEXT_SELECTIVE_2023_V2Both authorized executions failed before a complete target-free ranking. The terminal batch-8 attempt remained an execution-only amendment and also failed with CUDA OOM. No conclusion about mLateOn retrieval quality is authorized, and no further memory/search retuning is allowed inside this preregistered experiment.
CUBSO_NEED_IR_CURRENT_TECHNICAL_MLATEON_CHANNEL_SCREEN_ALIGNMENT_CLEAN_DEV_V1This paired development ablation holds the frozen occurrence cohort, query embeddings, model revision, exhaustive cosine scorer and exact 85,711 candidate codes fixed while changing only candidate-side representation. Missing NEED-IR views remain misses in both arms. TEST and TEMPORAL_TEST remain sealed.
CUBSO_NEED_IR_TECHNICAL_C0_VS_C1_ALIGNMENT_CLEAN_DEV_V1This paired development ablation holds the frozen occurrence cohort, query embeddings, model revision, exhaustive cosine scorer and exact 85,711 candidate codes fixed while changing only candidate-side representation. Missing NEED-IR views remain misses in both arms. TEST and TEMPORAL_TEST remain sealed.
CUBSO_NEED_IR_CORE_C0_VS_C1_ALIGNMENT_CLEAN_DEV_V1