Conjunto D1 sintético congelado e sanidade em CPU e GPU na OCI

- data/splits (d1-v1): dev 1183, calibração 1127, teste 1744 exemplos,
  43 rótulos, partição por lote de geração, SHA-256 em MANIFEST.json;
  cópias em dmb-entrada e dmb-logs-imutaveis.
- reports/datagen: relatórios de preparação e de anotação (3 LLMs em
  rodízio; kappa 0,98–0,99).
- reports/sanity/oci: sanidade dos 4 candidatos em CPU (us-chicago-1,
  sub-rede privada) e GPU A10 (sa-saopaulo-1).
- Plano: 47 de 78 tarefas concluídas; docs/02 com o resumo do conjunto.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

data/splits/MANIFEST.json

@@ -0,0 +1,54 @@
+{
+  "version": "d1-v1",
+  "taxonomy_version": "v1.0",
+  "built_at": "2026-10-07T17:13:53+00:00",
+  "source_run": "/data/runs/datagen-v1",
+  "seed": 20261007,
+  "splits": {
+    "dev": {
+      "file": "dev.jsonl",
+      "sha256": "964c58ca34054e7186b8544cbcddf8e89e317cdb51b43dc2ffd05c99ef757465",
+      "examples": 1183,
+      "groups": 99,
+      "labels": 43,
+      "oos": 178,
+      "gold_not_eligible": 98,
+      "n_options": {
+        "4": 320,
+        "8": 311,
+        "12": 274,
+        "16": 278
+      }
+    },
+    "calibration": {
+      "file": "calibration.jsonl",
+      "sha256": "c35d919324d5bbbcb2b9e7369a3554c5cfd77e5a50da3f7a4e6c4b971a03436c",
+      "examples": 1127,
+      "groups": 94,
+      "labels": 43,
+      "oos": 120,
+      "gold_not_eligible": 114,
+      "n_options": {
+        "4": 287,
+        "8": 293,
+        "12": 277,
+        "16": 270
+      }
+    },
+    "test": {
+      "file": "test.jsonl",
+      "sha256": "e3bd18c5ae6a19e996f0a81cece1b2afe1056e7969fa6b4ea962d3a1ca7df00b",
+      "examples": 1744,
+      "groups": 146,
+      "labels": 43,
+      "oos": 236,
+      "gold_not_eligible": 142,
+      "n_options": {
+        "4": 451,
+        "8": 431,
+        "12": 422,
+        "16": 440
+      }
+    }
+  }
+}

docs/02-desenho-experimental.md

@@ -94,6 +94,8 @@ Cada tarefa tem **pergunta**, **opções com descrição** e **regra de decisão
 | Anotação | `datagen.annotate`: três modelos (`openai.gpt-5.5`, `google.gemini-2.5-pro`, `xai.grok-4.3`) em rodízio por lote; dois anotam de forma independente, sem ver o rótulo de geração, e o terceiro desempata. Três rótulos distintos → exemplo excluído. Instruções: [06-guia-de-anotacao.md](06-guia-de-anotacao.md) | 29–32 |
 | Montagem | `datagen.build`: partição por lote de geração (3/2/2 lotes por intenção para teste/calibração/dev), conjuntos de 4, 8, 12 ou 16 opções simulando a elegibilidade (10% sem a intenção correta → `sem_correspondencia`), lint e congelamento com SHA-256 em `dmb-entrada` e `dmb-logs-imutaveis` | 33–36 |
 
+**Conjunto d1-v1 (congelado em 2026-10-07):** 4.068 gerados → 4.057 após a preparação → 4054 após a anotação (62 por desempate; 3 excluídos). Partições: dev 1183, calibração 1127, teste 1744, com 43 rótulos em todas. Kappa entre anotadores de 0.98 a 0.99: a concordância muito alta sugere que parte dos casos sintéticos é fácil. Para o relatório, os cortes por perfil (erros, regionalismo, emocional, indireto) e por `gold_eligible=false` devem receber atenção.
+
 Itens abaixo descrevem o desenho geral (válido também para dados reais em ciclos futuros).
 
 

docs/03-plano-de-tarefas.md

@@ -20,14 +20,14 @@ flowchart LR
 As fases F0–F1 (infra) e F2–F3 (dados) podem andar em paralelo.
 
 <!-- PROGRESSO:INICIO -->
-**Andamento em 2026-10-07:** 40 de 78 tarefas concluídas (51%), 8 em andamento.
+**Andamento em 2026-10-07:** 47 de 78 tarefas concluídas (60%), 2 em andamento.
 
 | Fase | Concluídas | Em andamento | Pendentes |
 |------|-----------|--------------|-----------|
 | F0 — Infraestrutura OCI | 6 | 2 | 2 |
 | F1 — Ambiente e artefatos | 9 | 0 | 0 |
 | F2 — Tarefas de decisão e taxonomia | 4 | 0 | 0 |
-| F3 — Dados e anotação | 5 | 6 | 1 |
+| F3 — Dados e anotação | 12 | 0 | 0 |
 | F4 — Contrato JEV e adaptadores | 16 | 0 | 0 |
 | F5 — Controle de execução | 0 | 0 | 3 |
 | F6 — Piloto | 0 | 0 | 4 |
@@ -35,7 +35,7 @@ As fases F0–F1 (infra) e F2–F3 (dados) podem andar em paralelo.
 | F8 — Teste final | 0 | 0 | 4 |
 | F9 — Métricas e análise | 0 | 0 | 8 |
 | F10 — Relatório e encerramento | 0 | 0 | 4 |
-| **Total** | **40** | **8** | **30** |
+| **Total** | **47** | **2** | **29** |
 
 Legenda: `[x]` concluída · `[~]` em andamento ou parcial · `[ ]` pendente.
 <!-- PROGRESSO:FIM -->
@@ -111,7 +111,7 @@ Legenda: `[x]` concluída · `[~]` em andamento ou parcial · `[ ]` pendente.
   - ↳ **100% sintéticos** (`datagen.generate`, na VM): 4.068 enunciados em 339 lotes (42 intenções × 7 perfis + 5 categorias fora de escopo × 9 lotes), geradores `llama-4-maverick` e `command-a` alternados.
   - ↳ A unidade é a **conversa com vários turnos**, não o enunciado isolado. Cada turno guarda o estado (agente ativo e sessões suspensas).
 - [x] **25.** Remover ou mascarar dados pessoais, segredos e identificadores que o benchmark não precisa.
-- [~] **26.** Separar os textos por domínio, comprimento, grau de ambiguidade e classe esperada.
+- [x] **26.** Separar os textos por domínio, comprimento, grau de ambiguidade e classe esperada.
   - ↳ Acrescentar a posição do turno e a presença de agente ativo ou de sessões suspensas.
   - ↳ Primeiro ciclo (só D1): metadados `domain`, `length_bucket`, `ambiguity`, `profile`, `sentiment`, `kind` e `n_options` gravados por `datagen.build`.
 - [x] **27.** Criar exemplos adversariais com abreviações, erros de ortografia, regionalismos, textos curtos e várias intenções.
@@ -122,16 +122,19 @@ Legenda: `[x]` concluída · `[~]` em andamento ou parcial · `[ ]` pendente.
 - [x] **29.** Produzir o guia de anotação humana.
   - ↳ [06-guia-de-anotacao.md](06-guia-de-anotacao.md), usado literalmente pelos anotadores.
   - ↳ O guia cobre D1–D4 com exemplos-limite (por exemplo, quando um enunciado curto continua a tarefa do agente ativo e quando inicia um novo agente).
-- [~] **30.** Anotar cada exemplo com dois avaliadores independentes.
+- [x] **30.** Anotar cada exemplo com dois avaliadores independentes.
   - ↳ Avaliadores = modelos generativos em rodízio (`openai.gpt-5.5`, `google.gemini-2.5-pro`, `xai.grok-4.3`). *(Q15)*
-- [~] **31.** Resolver as discordâncias com a revisão de um terceiro avaliador.
+  - ↳ Resultado: 4057 anotados; 3992 consensos (98.4%); kappa por par entre 0.982 e 0.986; concordância com o rótulo de geração 97.9%. Relatório em `reports/datagen/annotation-report.json`.
-- [~] **32.** Registrar, para cada exemplo, a resposta de referência, a justificativa e a versão da taxonomia.
+- [x] **31.** Resolver as discordâncias com a revisão de um terceiro avaliador.
+- [x] **32.** Registrar, para cada exemplo, a resposta de referência, a justificativa e a versão da taxonomia.
   - ↳ Reportar também a concordância entre anotadores.
-- [~] **33.** Separar os dados em desenvolvimento, calibração e teste final por grupo de origem. *(Q12)*
+- [x] **33.** Separar os dados em desenvolvimento, calibração e teste final por grupo de origem. *(Q12)*
-  - ↳ O grupo é `group_id`, ou seja, a conversa.
+  - ↳ O grupo é `group_id`, ou seja, a conversa. No ciclo sintético, o grupo é o lote de geração.
-- [~] **34.** Impedir que textos derivados do mesmo caso apareçam em mais de uma partição.
+  - ↳ `data/splits/` (d1-v1): dev 1183 · calibração 1127 · teste 1744 exemplos (teste ≥ 1.300, Q12).
-- [ ] **35.** Congelar o conjunto de teste final antes de qualquer ajuste de parâmetros.
+- [x] **34.** Impedir que textos derivados do mesmo caso apareçam em mais de uma partição.
+- [x] **35.** Congelar o conjunto de teste final antes de qualquer ajuste de parâmetros.
   - ↳ Registrar o hash do arquivo congelado e guardar uma cópia em `logs-imutaveis`.
+  - ↳ Congelado em 2026-10-07: SHA-256 em `data/splits/MANIFEST.json`, cópias em `dmb-entrada` e `dmb-logs-imutaveis` (`splits/d1-v1/`), junto com os dados brutos da geração e da anotação.
 
 ## F4 — Contrato JEV e adaptadores
 
@@ -171,7 +174,7 @@ Legenda: `[x]` concluída · `[~]` em andamento ou parcial · `[ ]` pendente.
 - [ ] **55.** Executar um piloto de 50 exemplos por candidato em CPU.
   - ↳ Sanidade em CPU (us-chicago-1, sub-rede privada) aprovada nos 4 candidatos: Laya 125 ms, GLiNER 194 ms, Rizzo Flow 7,3 s, SemIf 16,4 s (latência mediana). *(Q11)*
 - [ ] **56.** Executar um piloto de 50 exemplos por candidato em GPU.
-  - ↳ GPU em **sa-saopaulo-1, A10.1**, uma por vez *(Q14)*. Sanidade em GPU: Laya 21 ms, GLiNER 16 ms, SemIf 66 ms (latência mediana). Rizzo Flow exigiu compilar o llama.cpp com CUDA para a glibc da imagem dos jobs (`infra/ds/build_llama_cuda.sh`).
+  - ↳ GPU em **sa-saopaulo-1, A10.1**, uma por vez *(Q14)*. Sanidade em GPU: Laya 21 ms, GLiNER 16 ms, SemIf 66 ms (latência mediana). Rizzo Flow 62 ms, depois de compilar o llama.cpp com CUDA para a glibc da imagem dos jobs (`infra/ds/build_llama_cuda.sh`).
 - [ ] **57.** Medir memória de GPU, memória RAM, tempo de carga, tempo de aquecimento e falhas do piloto.
 - [ ] **58.** Ajustar apenas os parâmetros operacionais necessários para eliminar falhas de execução.
 

reports/datagen/annotation-report.json

@@ -0,0 +1,66 @@
+{
+  "annotators": [
+    "openai.gpt-5.5",
+    "google.gemini-2.5-pro",
+    "xai.grok-4.3"
+  ],
+  "items": 4057,
+  "status": {
+    "consenso": 3992,
+    "desempate": 62,
+    "excluido_tres_rotulos": 2,
+    "excluido_sem_desempate": 1
+  },
+  "cohen_kappa": {
+    "google.gemini-2.5-pro × openai.gpt-5.5": 0.9862,
+    "google.gemini-2.5-pro × xai.grok-4.3": 0.9816,
+    "openai.gpt-5.5 × xai.grok-4.3": 0.9823
+  },
+  "raw_agreement": 0.984,
+  "agreement_with_generation": 0.9793,
+  "gold_distribution": {
+    "sem_correspondencia": 533,
+    "atendimento_humano": 94,
+    "conta_explicacao": 92,
+    "recarga_nao_creditada": 91,
+    "plano_troca": 88,
+    "conta_contestacao": 87,
+    "internet_consumo": 86,
+    "conta_segunda_via": 86,
+    "roaming_duvida": 86,
+    "negociacao_religacao": 85,
+    "internet_pacote_adicional": 85,
+    "negociacao_divida": 85,
+    "desconto_retencao": 85,
+    "vas_consulta": 85,
+    "vas_cancelamento": 85,
+    "marketplace_pedido": 85,
+    "roaming_cobranca": 84,
+    "seguranca_perda_roubo": 84,
+    "suporte_configuracao": 84,
+    "beneficios_consulta": 84,
+    "recarga_realizar": 84,
+    "troca_numero": 84,
+    "cadastro_atualizacao": 84,
+    "conta_vencimento": 84,
+    "fidelidade_pontos": 83,
+    "marketplace_compra": 83,
+    "fidelidade_resgate": 83,
+    "conta_debito_automatico_digital": 83,
+    "roaming_pacote": 83,
+    "suporte_sem_sinal": 83,
+    "conta_pagamento_confirmacao": 83,
+    "cancelamento_linha": 83,
+    "internet_problema_conexao": 83,
+    "suporte_chip_esim": 83,
+    "bonus_nao_recebido": 83,
+    "portabilidade": 82,
+    "vas_contratacao": 81,
+    "plano_consulta": 81,
+    "parceiros_ofertas": 80,
+    "plano_nova_linha": 79,
+    "seguranca_fraude": 78,
+    "reclamacao_atendimento": 75,
+    "recarga_saldo": 75
+  }
+}

reports/datagen/prepare-report.json

@@ -0,0 +1,7 @@
+{
+  "pii_masked": 4,
+  "excluded_near_dup": 2,
+  "excluded_exact_dup": 9,
+  "generated": 4068,
+  "kept": 4057
+}

reports/sanity/oci/gliner-cpu.json

@@ -0,0 +1,222 @@
+{
+  "candidate": {
+    "name": "gliner",
+    "model_id": "gliner@a35a0cd3b7a0",
+    "revision": "a35a0cd3b7a0f00f2effc576f454cd48fa98aa5f",
+    "adapter_version": "0.1.0",
+    "device": "cpu",
+    "context_limit": 4096,
+    "probabilities_source": "native"
+  },
+  "device": "cpu",
+  "load_seconds": 19.49,
+  "conformance": [
+    {
+      "name": "valid_choice_noul",
+      "expected": "ok",
+      "got": "ok"
+    },
+    {
+      "name": "valid_score",
+      "expected": "ok",
+      "got": "ok"
+    },
+    {
+      "name": "valid_state_object",
+      "expected": "ok",
+      "got": "ok"
+    },
+    {
+      "name": "invalid_17_options",
+      "expected": "invalid_request",
+      "got": "invalid_request"
+    },
+    {
+      "name": "invalid_single_option",
+      "expected": "invalid_request",
+      "got": "invalid_request"
+    },
+    {
+      "name": "invalid_noul_missing_false",
+      "expected": "invalid_request",
+      "got": "invalid_request"
+    },
+    {
+      "name": "invalid_unknown_type",
+      "expected": "invalid_request",
+      "got": "invalid_request"
+    },
+    {
+      "name": "invalid_empty_state",
+      "expected": "invalid_request",
+      "got": "invalid_request"
+    },
+    {
+      "name": "invalid_no_questions",
+      "expected": "invalid_request",
+      "got": "invalid_request"
+    },
+    {
+      "name": "invalid_extra_field",
+      "expected": "invalid_request",
+      "got": "invalid_request"
+    }
+  ],
+  "sanity": [
+    {
+      "name": "choice:pagamento",
+      "gold": "pagamento",
+      "pred": "pagamento",
+      "latency_ms": 207.2,
+      "answer": {
+        "type": "choice",
+        "choice": "pagamento",
+        "confidence": 0.6344144486132514,
+        "probabilities": {
+          "pagamento": 0.6344144486132514,
+          "cancelamento": 0.016834581647096163,
+          "suporte_tecnico": 0.022163255635572723,
+          "sem_correspondencia": 0.3265877141040798
+        }
+      },
+      "pred_permuted": "pagamento",
+      "permutation_invariant": true,
+      "correct": true
+    },
+    {
+      "name": "choice:cancelamento",
+      "gold": "cancelamento",
+      "pred": "cancelamento",
+      "latency_ms": 206.5,
+      "answer": {
+        "type": "choice",
+        "choice": "cancelamento",
+        "confidence": 0.8718590947504495,
+        "probabilities": {
+          "pagamento": 0.010419044895931304,
+          "cancelamento": 0.8718590947504495,
+          "suporte_tecnico": 0.012619423054256475,
+          "sem_correspondencia": 0.10510243729936272
+        }
+      },
+      "pred_permuted": "cancelamento",
+      "permutation_invariant": true,
+      "correct": true
+    },
+    {
+      "name": "choice:suporte_tecnico",
+      "gold": "suporte_tecnico",
+      "pred": "sem_correspondencia",
+      "latency_ms": 204.3,
+      "answer": {
+        "type": "choice",
+        "choice": "sem_correspondencia",
+        "confidence": 0.634299124256301,
+        "probabilities": {
+          "pagamento": 0.02020097044151347,
+          "cancelamento": 0.04278694633119902,
+          "suporte_tecnico": 0.30271295897098655,
+          "sem_correspondencia": 0.634299124256301
+        }
+      },
+      "pred_permuted": "sem_correspondencia",
+      "permutation_invariant": true,
+      "correct": false
+    },
+    {
+      "name": "choice:sem_correspondencia",
+      "gold": "sem_correspondencia",
+      "pred": "sem_correspondencia",
+      "latency_ms": 194.3,
+      "answer": {
+        "type": "choice",
+        "choice": "sem_correspondencia",
+        "confidence": 0.8604282044571466,
+        "probabilities": {
+          "pagamento": 0.02875952763006619,
+          "cancelamento": 0.0316480618158572,
+          "suporte_tecnico": 0.07916420609692999,
+          "sem_correspondencia": 0.8604282044571466
+        }
+      },
+      "pred_permuted": "sem_correspondencia",
+      "permutation_invariant": true,
+      "correct": true
+    },
+    {
+      "name": "noul:clara",
+      "gold": "true",
+      "pred": "false",
+      "latency_ms": 173.0,
+      "answer": {
+        "type": "noul",
+        "noul": 0.2368284577603868
+      },
+      "correct": false
+    },
+    {
+      "name": "noul:vaga",
+      "gold": "false",
+      "pred": "false",
+      "latency_ms": 166.4,
+      "answer": {
+        "type": "noul",
+        "noul": 0.20341389158881357
+      },
+      "correct": true
+    },
+    {
+      "name": "score:alta",
+      "gold": "2",
+      "pred": "0",
+      "latency_ms": 163.5,
+      "answer": {
+        "type": "score",
+        "score": 0.9812680775305056,
+        "confidence": 0.3614072907390383,
+        "probabilities": {
+          "0": 0.3614072907390383,
+          "1": 0.2959173409914178,
+          "2": 0.34267536826954387
+        },
+        "legend": {
+          "0": "Baixa",
+          "1": "Média",
+          "2": "Alta"
+        }
+      },
+      "correct": false
+    },
+    {
+      "name": "score:baixa",
+      "gold": "0",
+      "pred": "0",
+      "latency_ms": 155.1,
+      "answer": {
+        "type": "score",
+        "score": 0.4816716677697748,
+        "confidence": 0.6750852767416213,
+        "probabilities": {
+          "0": 0.6750852767416213,
+          "1": 0.16815777874698246,
+          "2": 0.15675694451139618
+        },
+        "legend": {
+          "0": "Baixa",
+          "1": "Média",
+          "2": "Alta"
+        }
+      },
+      "correct": true
+    }
+  ],
+  "summary": {
+    "contract_violations": [],
+    "sanity_correct": "5/8",
+    "permutation_invariant": "4/4"
+  },
+  "oci": {
+    "region": "us-chicago-1",
+    "run_id": "20261007T153410Z-gliner-sanity"
+  }
+}

reports/sanity/oci/gliner-cuda.json

@@ -0,0 +1,223 @@
+{
+  "candidate": {
+    "name": "gliner",
+    "model_id": "gliner@a35a0cd3b7a0",
+    "revision": "a35a0cd3b7a0f00f2effc576f454cd48fa98aa5f",
+    "adapter_version": "0.1.0",
+    "device": "cuda",
+    "context_limit": 4096,
+    "probabilities_source": "native"
+  },
+  "device": "cuda",
+  "load_seconds": 21.32,
+  "process_guard": "ok",
+  "conformance": [
+    {
+      "name": "valid_choice_noul",
+      "expected": "ok",
+      "got": "ok"
+    },
+    {
+      "name": "valid_score",
+      "expected": "ok",
+      "got": "ok"
+    },
+    {
+      "name": "valid_state_object",
+      "expected": "ok",
+      "got": "ok"
+    },
+    {
+      "name": "invalid_17_options",
+      "expected": "invalid_request",
+      "got": "invalid_request"
+    },
+    {
+      "name": "invalid_single_option",
+      "expected": "invalid_request",
+      "got": "invalid_request"
+    },
+    {
+      "name": "invalid_noul_missing_false",
+      "expected": "invalid_request",
+      "got": "invalid_request"
+    },
+    {
+      "name": "invalid_unknown_type",
+      "expected": "invalid_request",
+      "got": "invalid_request"
+    },
+    {
+      "name": "invalid_empty_state",
+      "expected": "invalid_request",
+      "got": "invalid_request"
+    },
+    {
+      "name": "invalid_no_questions",
+      "expected": "invalid_request",
+      "got": "invalid_request"
+    },
+    {
+      "name": "invalid_extra_field",
+      "expected": "invalid_request",
+      "got": "invalid_request"
+    }
+  ],
+  "sanity": [
+    {
+      "name": "choice:pagamento",
+      "gold": "pagamento",
+      "pred": "pagamento",
+      "latency_ms": 16.4,
+      "answer": {
+        "type": "choice",
+        "choice": "pagamento",
+        "confidence": 0.6344155822832185,
+        "probabilities": {
+          "pagamento": 0.6344155822832185,
+          "cancelamento": 0.016834572365227046,
+          "suporte_tecnico": 0.022163173720466455,
+          "sem_correspondencia": 0.32658667163108795
+        }
+      },
+      "pred_permuted": "pagamento",
+      "permutation_invariant": true,
+      "correct": true
+    },
+    {
+      "name": "choice:cancelamento",
+      "gold": "cancelamento",
+      "pred": "cancelamento",
+      "latency_ms": 16.1,
+      "answer": {
+        "type": "choice",
+        "choice": "cancelamento",
+        "confidence": 0.8718585964717693,
+        "probabilities": {
+          "pagamento": 0.010419094712095255,
+          "cancelamento": 0.8718585964717693,
+          "suporte_tecnico": 0.012619434782507587,
+          "sem_correspondencia": 0.10510287403362785
+        }
+      },
+      "pred_permuted": "cancelamento",
+      "permutation_invariant": true,
+      "correct": true
+    },
+    {
+      "name": "choice:suporte_tecnico",
+      "gold": "suporte_tecnico",
+      "pred": "sem_correspondencia",
+      "latency_ms": 16.9,
+      "answer": {
+        "type": "choice",
+        "choice": "sem_correspondencia",
+        "confidence": 0.6342994187351121,
+        "probabilities": {
+          "pagamento": 0.02020120315928321,
+          "cancelamento": 0.04278698707030298,
+          "suporte_tecnico": 0.3027123910353017,
+          "sem_correspondencia": 0.6342994187351121
+        }
+      },
+      "pred_permuted": "sem_correspondencia",
+      "permutation_invariant": true,
+      "correct": false
+    },
+    {
+      "name": "choice:sem_correspondencia",
+      "gold": "sem_correspondencia",
+      "pred": "sem_correspondencia",
+      "latency_ms": 15.5,
+      "answer": {
+        "type": "choice",
+        "choice": "sem_correspondencia",
+        "confidence": 0.8604289616881081,
+        "probabilities": {
+          "pagamento": 0.028759510277003655,
+          "cancelamento": 0.0316479121558098,
+          "suporte_tecnico": 0.07916361587907846,
+          "sem_correspondencia": 0.8604289616881081
+        }
+      },
+      "pred_permuted": "sem_correspondencia",
+      "permutation_invariant": true,
+      "correct": true
+    },
+    {
+      "name": "noul:clara",
+      "gold": "true",
+      "pred": "false",
+      "latency_ms": 16.1,
+      "answer": {
+        "type": "noul",
+        "noul": 0.2368276601557085
+      },
+      "correct": false
+    },
+    {
+      "name": "noul:vaga",
+      "gold": "false",
+      "pred": "false",
+      "latency_ms": 16.4,
+      "answer": {
+        "type": "noul",
+        "noul": 0.20341337939580556
+      },
+      "correct": true
+    },
+    {
+      "name": "score:alta",
+      "gold": "2",
+      "pred": "0",
+      "latency_ms": 16.5,
+      "answer": {
+        "type": "score",
+        "score": 0.9812681382516582,
+        "confidence": 0.3614074480114148,
+        "probabilities": {
+          "0": 0.3614074480114148,
+          "1": 0.29591696572551235,
+          "2": 0.3426755862630729
+        },
+        "legend": {
+          "0": "Baixa",
+          "1": "Média",
+          "2": "Alta"
+        }
+      },
+      "correct": false
+    },
+    {
+      "name": "score:baixa",
+      "gold": "0",
+      "pred": "0",
+      "latency_ms": 15.9,
+      "answer": {
+        "type": "score",
+        "score": 0.48167205020737197,
+        "confidence": 0.6750848881824415,
+        "probabilities": {
+          "0": 0.6750848881824415,
+          "1": 0.16815817342774508,
+          "2": 0.15675693838981344
+        },
+        "legend": {
+          "0": "Baixa",
+          "1": "Média",
+          "2": "Alta"
+        }
+      },
+      "correct": true
+    }
+  ],
+  "summary": {
+    "contract_violations": [],
+    "sanity_correct": "5/8",
+    "permutation_invariant": "4/4"
+  },
+  "oci": {
+    "region": "sa-saopaulo-1",
+    "run_id": "20261007T162549Z-gliner-sanity"
+  }
+}

reports/sanity/oci/laya-cpu.json

@@ -0,0 +1,222 @@
+{
+  "candidate": {
+    "name": "laya",
+    "model_id": "laya@1720e3e3357c",
+    "revision": "1720e3e3357cfe1e281542e223f8273b0890ca34",
+    "adapter_version": "0.1.0",
+    "device": "cpu",
+    "context_limit": 1024,
+    "probabilities_source": "native"
+  },
+  "device": "cpu",
+  "load_seconds": 12.29,
+  "conformance": [
+    {
+      "name": "valid_choice_noul",
+      "expected": "ok",
+      "got": "ok"
+    },
+    {
+      "name": "valid_score",
+      "expected": "ok",
+      "got": "ok"
+    },
+    {
+      "name": "valid_state_object",
+      "expected": "ok",
+      "got": "ok"
+    },
+    {
+      "name": "invalid_17_options",
+      "expected": "invalid_request",
+      "got": "invalid_request"
+    },
+    {
+      "name": "invalid_single_option",
+      "expected": "invalid_request",
+      "got": "invalid_request"
+    },
+    {
+      "name": "invalid_noul_missing_false",
+      "expected": "invalid_request",
+      "got": "invalid_request"
+    },
+    {
+      "name": "invalid_unknown_type",
+      "expected": "invalid_request",
+      "got": "invalid_request"
+    },
+    {
+      "name": "invalid_empty_state",
+      "expected": "invalid_request",
+      "got": "invalid_request"
+    },
+    {
+      "name": "invalid_no_questions",
+      "expected": "invalid_request",
+      "got": "invalid_request"
+    },
+    {
+      "name": "invalid_extra_field",
+      "expected": "invalid_request",
+      "got": "invalid_request"
+    }
+  ],
+  "sanity": [
+    {
+      "name": "choice:pagamento",
+      "gold": "pagamento",
+      "pred": "pagamento",
+      "latency_ms": 140.1,
+      "answer": {
+        "type": "choice",
+        "choice": "pagamento",
+        "confidence": 0.9977,
+        "probabilities": {
+          "pagamento": 0.9997979802020598,
+          "cancelamento": 1.0000980096049412e-06,
+          "suporte_tecnico": 1.0000980096049412e-06,
+          "sem_correspondencia": 0.00020001960192098824
+        }
+      },
+      "pred_permuted": "pagamento",
+      "permutation_invariant": true,
+      "correct": true
+    },
+    {
+      "name": "choice:cancelamento",
+      "gold": "cancelamento",
+      "pred": "cancelamento",
+      "latency_ms": 124.9,
+      "answer": {
+        "type": "choice",
+        "choice": "cancelamento",
+        "confidence": 1.0,
+        "probabilities": {
+          "pagamento": 9.999970000090002e-07,
+          "cancelamento": 0.9999970000090003,
+          "suporte_tecnico": 9.999970000090002e-07,
+          "sem_correspondencia": 9.999970000090002e-07
+        }
+      },
+      "pred_permuted": "cancelamento",
+      "permutation_invariant": true,
+      "correct": true
+    },
+    {
+      "name": "choice:suporte_tecnico",
+      "gold": "suporte_tecnico",
+      "pred": "suporte_tecnico",
+      "latency_ms": 142.7,
+      "answer": {
+        "type": "choice",
+        "choice": "suporte_tecnico",
+        "confidence": 0.2935,
+        "probabilities": {
+          "pagamento": 0.0496049604960496,
+          "cancelamento": 0.047204720472047206,
+          "suporte_tecnico": 0.5706570657065706,
+          "sem_correspondencia": 0.33253325332533257
+        }
+      },
+      "pred_permuted": "suporte_tecnico",
+      "permutation_invariant": true,
+      "correct": true
+    },
+    {
+      "name": "choice:sem_correspondencia",
+      "gold": "sem_correspondencia",
+      "pred": "sem_correspondencia",
+      "latency_ms": 137.6,
+      "answer": {
+        "type": "choice",
+        "choice": "sem_correspondencia",
+        "confidence": 0.7909,
+        "probabilities": {
+          "pagamento": 0.0034996500349965005,
+          "cancelamento": 0.0044995500449955,
+          "suporte_tecnico": 0.0642935706429357,
+          "sem_correspondencia": 0.9277072292770723
+        }
+      },
+      "pred_permuted": "sem_correspondencia",
+      "permutation_invariant": true,
+      "correct": true
+    },
+    {
+      "name": "noul:clara",
+      "gold": "true",
+      "pred": "true",
+      "latency_ms": 107.2,
+      "answer": {
+        "type": "noul",
+        "noul": 0.9982
+      },
+      "correct": true
+    },
+    {
+      "name": "noul:vaga",
+      "gold": "false",
+      "pred": "false",
+      "latency_ms": 106.4,
+      "answer": {
+        "type": "noul",
+        "noul": 0.1563
+      },
+      "correct": true
+    },
+    {
+      "name": "score:alta",
+      "gold": "2",
+      "pred": "2",
+      "latency_ms": 113.8,
+      "answer": {
+        "type": "score",
+        "score": 1.1202,
+        "confidence": 0.01,
+        "probabilities": {
+          "0": 0.2718,
+          "1": 0.3362,
+          "2": 0.392
+        },
+        "legend": {
+          "0": "Baixa",
+          "1": "Média",
+          "2": "Alta"
+        }
+      },
+      "correct": true
+    },
+    {
+      "name": "score:baixa",
+      "gold": "0",
+      "pred": "1",
+      "latency_ms": 116.9,
+      "answer": {
+        "type": "score",
+        "score": 1.0998999999999999,
+        "confidence": 0.1514,
+        "probabilities": {
+          "0": 0.1477,
+          "1": 0.6047,
+          "2": 0.2476
+        },
+        "legend": {
+          "0": "Baixa",
+          "1": "Média",
+          "2": "Alta"
+        }
+      },
+      "correct": false
+    }
+  ],
+  "summary": {
+    "contract_violations": [],
+    "sanity_correct": "7/8",
+    "permutation_invariant": "4/4"
+  },
+  "oci": {
+    "region": "us-chicago-1",
+    "run_id": "20261007T154451Z-laya-sanity"
+  }
+}

reports/sanity/oci/laya-cuda.json

@@ -0,0 +1,223 @@
+{
+  "candidate": {
+    "name": "laya",
+    "model_id": "laya@1720e3e3357c",
+    "revision": "1720e3e3357cfe1e281542e223f8273b0890ca34",
+    "adapter_version": "0.1.0",
+    "device": "cuda",
+    "context_limit": 1024,
+    "probabilities_source": "native"
+  },
+  "device": "cuda",
+  "load_seconds": 8.01,
+  "process_guard": "ok",
+  "conformance": [
+    {
+      "name": "valid_choice_noul",
+      "expected": "ok",
+      "got": "ok"
+    },
+    {
+      "name": "valid_score",
+      "expected": "ok",
+      "got": "ok"
+    },
+    {
+      "name": "valid_state_object",
+      "expected": "ok",
+      "got": "ok"
+    },
+    {
+      "name": "invalid_17_options",
+      "expected": "invalid_request",
+      "got": "invalid_request"
+    },
+    {
+      "name": "invalid_single_option",
+      "expected": "invalid_request",
+      "got": "invalid_request"
+    },
+    {
+      "name": "invalid_noul_missing_false",
+      "expected": "invalid_request",
+      "got": "invalid_request"
+    },
+    {
+      "name": "invalid_unknown_type",
+      "expected": "invalid_request",
+      "got": "invalid_request"
+    },
+    {
+      "name": "invalid_empty_state",
+      "expected": "invalid_request",
+      "got": "invalid_request"
+    },
+    {
+      "name": "invalid_no_questions",
+      "expected": "invalid_request",
+      "got": "invalid_request"
+    },
+    {
+      "name": "invalid_extra_field",
+      "expected": "invalid_request",
+      "got": "invalid_request"
+    }
+  ],
+  "sanity": [
+    {
+      "name": "choice:pagamento",
+      "gold": "pagamento",
+      "pred": "pagamento",
+      "latency_ms": 20.8,
+      "answer": {
+        "type": "choice",
+        "choice": "pagamento",
+        "confidence": 0.9977,
+        "probabilities": {
+          "pagamento": 0.9997979802020598,
+          "cancelamento": 1.0000980096049412e-06,
+          "suporte_tecnico": 1.0000980096049412e-06,
+          "sem_correspondencia": 0.00020001960192098824
+        }
+      },
+      "pred_permuted": "pagamento",
+      "permutation_invariant": true,
+      "correct": true
+    },
+    {
+      "name": "choice:cancelamento",
+      "gold": "cancelamento",
+      "pred": "cancelamento",
+      "latency_ms": 20.1,
+      "answer": {
+        "type": "choice",
+        "choice": "cancelamento",
+        "confidence": 1.0,
+        "probabilities": {
+          "pagamento": 9.999970000090002e-07,
+          "cancelamento": 0.9999970000090003,
+          "suporte_tecnico": 9.999970000090002e-07,
+          "sem_correspondencia": 9.999970000090002e-07
+        }
+      },
+      "pred_permuted": "cancelamento",
+      "permutation_invariant": true,
+      "correct": true
+    },
+    {
+      "name": "choice:suporte_tecnico",
+      "gold": "suporte_tecnico",
+      "pred": "suporte_tecnico",
+      "latency_ms": 21.0,
+      "answer": {
+        "type": "choice",
+        "choice": "suporte_tecnico",
+        "confidence": 0.292,
+        "probabilities": {
+          "pagamento": 0.0487,
+          "cancelamento": 0.0457,
+          "suporte_tecnico": 0.557,
+          "sem_correspondencia": 0.3486
+        }
+      },
+      "pred_permuted": "suporte_tecnico",
+      "permutation_invariant": true,
+      "correct": true
+    },
+    {
+      "name": "choice:sem_correspondencia",
+      "gold": "sem_correspondencia",
+      "pred": "sem_correspondencia",
+      "latency_ms": 20.5,
+      "answer": {
+        "type": "choice",
+        "choice": "sem_correspondencia",
+        "confidence": 0.7962,
+        "probabilities": {
+          "pagamento": 0.0032,
+          "cancelamento": 0.0044,
+          "suporte_tecnico": 0.0623,
+          "sem_correspondencia": 0.9301
+        }
+      },
+      "pred_permuted": "sem_correspondencia",
+      "permutation_invariant": true,
+      "correct": true
+    },
+    {
+      "name": "noul:clara",
+      "gold": "true",
+      "pred": "true",
+      "latency_ms": 20.8,
+      "answer": {
+        "type": "noul",
+        "noul": 0.9981
+      },
+      "correct": true
+    },
+    {
+      "name": "noul:vaga",
+      "gold": "false",
+      "pred": "false",
+      "latency_ms": 20.1,
+      "answer": {
+        "type": "noul",
+        "noul": 0.1546
+      },
+      "correct": true
+    },
+    {
+      "name": "score:alta",
+      "gold": "2",
+      "pred": "2",
+      "latency_ms": 20.9,
+      "answer": {
+        "type": "score",
+        "score": 1.123,
+        "confidence": 0.0105,
+        "probabilities": {
+          "0": 0.2703,
+          "1": 0.3364,
+          "2": 0.3933
+        },
+        "legend": {
+          "0": "Baixa",
+          "1": "Média",
+          "2": "Alta"
+        }
+      },
+      "correct": true
+    },
+    {
+      "name": "score:baixa",
+      "gold": "0",
+      "pred": "1",
+      "latency_ms": 21.0,
+      "answer": {
+        "type": "score",
+        "score": 1.1033,
+        "confidence": 0.1552,
+        "probabilities": {
+          "0": 0.1446,
+          "1": 0.6075,
+          "2": 0.2479
+        },
+        "legend": {
+          "0": "Baixa",
+          "1": "Média",
+          "2": "Alta"
+        }
+      },
+      "correct": false
+    }
+  ],
+  "summary": {
+    "contract_violations": [],
+    "sanity_correct": "7/8",
+    "permutation_invariant": "4/4"
+  },
+  "oci": {
+    "region": "sa-saopaulo-1",
+    "run_id": "20261007T162118Z-laya-sanity"
+  }
+}

reports/sanity/oci/rizzo_flow-cpu.json

@@ -0,0 +1,222 @@
+{
+  "candidate": {
+    "name": "rizzo_flow",
+    "model_id": "rizzo_flow@55633c8cbd2b",
+    "revision": "55633c8cbd2b826bd3eefdeb05310450996649df",
+    "adapter_version": "0.1.0",
+    "device": "cpu",
+    "context_limit": 2048,
+    "probabilities_source": "native"
+  },
+  "device": "cpu",
+  "load_seconds": 4.32,
+  "conformance": [
+    {
+      "name": "valid_choice_noul",
+      "expected": "ok",
+      "got": "ok"
+    },
+    {
+      "name": "valid_score",
+      "expected": "ok",
+      "got": "ok"
+    },
+    {
+      "name": "valid_state_object",
+      "expected": "ok",
+      "got": "ok"
+    },
+    {
+      "name": "invalid_17_options",
+      "expected": "invalid_request",
+      "got": "invalid_request"
+    },
+    {
+      "name": "invalid_single_option",
+      "expected": "invalid_request",
+      "got": "invalid_request"
+    },
+    {
+      "name": "invalid_noul_missing_false",
+      "expected": "invalid_request",
+      "got": "invalid_request"
+    },
+    {
+      "name": "invalid_unknown_type",
+      "expected": "invalid_request",
+      "got": "invalid_request"
+    },
+    {
+      "name": "invalid_empty_state",
+      "expected": "invalid_request",
+      "got": "invalid_request"
+    },
+    {
+      "name": "invalid_no_questions",
+      "expected": "invalid_request",
+      "got": "invalid_request"
+    },
+    {
+      "name": "invalid_extra_field",
+      "expected": "invalid_request",
+      "got": "invalid_request"
+    }
+  ],
+  "sanity": [
+    {
+      "name": "choice:pagamento",
+      "gold": "pagamento",
+      "pred": "pagamento",
+      "latency_ms": 7381.4,
+      "answer": {
+        "type": "choice",
+        "choice": "pagamento",
+        "confidence": 0.9939581449644033,
+        "probabilities": {
+          "pagamento": 0.9954686087233022,
+          "cancelamento": 0.0014294878384867884,
+          "suporte_tecnico": 0.0006663578744809516,
+          "sem_correspondencia": 0.002435545563729951
+        }
+      },
+      "pred_permuted": "pagamento",
+      "permutation_invariant": true,
+      "correct": true
+    },
+    {
+      "name": "choice:cancelamento",
+      "gold": "cancelamento",
+      "pred": "cancelamento",
+      "latency_ms": 7903.2,
+      "answer": {
+        "type": "choice",
+        "choice": "cancelamento",
+        "confidence": 0.9959894668208951,
+        "probabilities": {
+          "pagamento": 0.0006611097629735292,
+          "cancelamento": 0.9969921001156713,
+          "suporte_tecnico": 0.0005694308917635679,
+          "sem_correspondencia": 0.0017773592295916517
+        }
+      },
+      "pred_permuted": "cancelamento",
+      "permutation_invariant": true,
+      "correct": true
+    },
+    {
+      "name": "choice:suporte_tecnico",
+      "gold": "suporte_tecnico",
+      "pred": "suporte_tecnico",
+      "latency_ms": 7312.8,
+      "answer": {
+        "type": "choice",
+        "choice": "suporte_tecnico",
+        "confidence": 0.9550325369120003,
+        "probabilities": {
+          "pagamento": 0.0009793333603067293,
+          "cancelamento": 0.002014121845135763,
+          "suporte_tecnico": 0.9662744026840004,
+          "sem_correspondencia": 0.030732142110557085
+        }
+      },
+      "pred_permuted": "suporte_tecnico",
+      "permutation_invariant": true,
+      "correct": true
+    },
+    {
+      "name": "choice:sem_correspondencia",
+      "gold": "sem_correspondencia",
+      "pred": "sem_correspondencia",
+      "latency_ms": 7599.3,
+      "answer": {
+        "type": "choice",
+        "choice": "sem_correspondencia",
+        "confidence": 0.9885549303273677,
+        "probabilities": {
+          "pagamento": 0.0036670988032871665,
+          "cancelamento": 0.0022098453219148453,
+          "suporte_tecnico": 0.002706858129272241,
+          "sem_correspondencia": 0.9914161977455258
+        }
+      },
+      "pred_permuted": "sem_correspondencia",
+      "permutation_invariant": true,
+      "correct": true
+    },
+    {
+      "name": "noul:clara",
+      "gold": "true",
+      "pred": "true",
+      "latency_ms": 6481.8,
+      "answer": {
+        "type": "noul",
+        "noul": 0.9479776743420092
+      },
+      "correct": true
+    },
+    {
+      "name": "noul:vaga",
+      "gold": "false",
+      "pred": "true",
+      "latency_ms": 6144.2,
+      "answer": {
+        "type": "noul",
+        "noul": 0.5076455176696537
+      },
+      "correct": false
+    },
+    {
+      "name": "score:alta",
+      "gold": "2",
+      "pred": "2",
+      "latency_ms": 5982.9,
+      "answer": {
+        "type": "score",
+        "score": 1.9453121647986211,
+        "confidence": 0.9510302861376827,
+        "probabilities": {
+          "0": 0.02204135929316713,
+          "1": 0.01060511661504444,
+          "2": 0.9673535240917883
+        },
+        "legend": {
+          "0": "Baixa",
+          "1": "Média",
+          "2": "Alta"
+        }
+      },
+      "correct": true
+    },
+    {
+      "name": "score:baixa",
+      "gold": "0",
+      "pred": "0",
+      "latency_ms": 6279.1,
+      "answer": {
+        "type": "score",
+        "score": 0.26957479262097783,
+        "confidence": 0.6046525975421342,
+        "probabilities": {
+          "0": 0.7364350650280896,
+          "1": 0.2575550773228431,
+          "2": 0.0060098576490673645
+        },
+        "legend": {
+          "0": "Baixa",
+          "1": "Média",
+          "2": "Alta"
+        }
+      },
+      "correct": true
+    }
+  ],
+  "summary": {
+    "contract_violations": [],
+    "sanity_correct": "7/8",
+    "permutation_invariant": "4/4"
+  },
+  "oci": {
+    "region": "us-chicago-1",
+    "run_id": "20261007T152414Z-rizzo_flow-sanity"
+  }
+}

reports/sanity/oci/rizzo_flow-cuda.json

@@ -0,0 +1,223 @@
+{
+  "candidate": {
+    "name": "rizzo_flow",
+    "model_id": "rizzo_flow@55633c8cbd2b",
+    "revision": "55633c8cbd2b826bd3eefdeb05310450996649df",
+    "adapter_version": "0.1.0",
+    "device": "cuda",
+    "context_limit": 2048,
+    "probabilities_source": "native"
+  },
+  "device": "cuda",
+  "load_seconds": 5.3,
+  "process_guard": "ok",
+  "conformance": [
+    {
+      "name": "valid_choice_noul",
+      "expected": "ok",
+      "got": "ok"
+    },
+    {
+      "name": "valid_score",
+      "expected": "ok",
+      "got": "ok"
+    },
+    {
+      "name": "valid_state_object",
+      "expected": "ok",
+      "got": "ok"
+    },
+    {
+      "name": "invalid_17_options",
+      "expected": "invalid_request",
+      "got": "invalid_request"
+    },
+    {
+      "name": "invalid_single_option",
+      "expected": "invalid_request",
+      "got": "invalid_request"
+    },
+    {
+      "name": "invalid_noul_missing_false",
+      "expected": "invalid_request",
+      "got": "invalid_request"
+    },
+    {
+      "name": "invalid_unknown_type",
+      "expected": "invalid_request",
+      "got": "invalid_request"
+    },
+    {
+      "name": "invalid_empty_state",
+      "expected": "invalid_request",
+      "got": "invalid_request"
+    },
+    {
+      "name": "invalid_no_questions",
+      "expected": "invalid_request",
+      "got": "invalid_request"
+    },
+    {
+      "name": "invalid_extra_field",
+      "expected": "invalid_request",
+      "got": "invalid_request"
+    }
+  ],
+  "sanity": [
+    {
+      "name": "choice:pagamento",
+      "gold": "pagamento",
+      "pred": "pagamento",
+      "latency_ms": 62.3,
+      "answer": {
+        "type": "choice",
+        "choice": "pagamento",
+        "confidence": 0.9937313892294425,
+        "probabilities": {
+          "pagamento": 0.9952985419220819,
+          "cancelamento": 0.0014555718031770129,
+          "suporte_tecnico": 0.0007138147545963065,
+          "sem_correspondencia": 0.0025320715201448793
+        }
+      },
+      "pred_permuted": "pagamento",
+      "permutation_invariant": true,
+      "correct": true
+    },
+    {
+      "name": "choice:cancelamento",
+      "gold": "cancelamento",
+      "pred": "cancelamento",
+      "latency_ms": 66.4,
+      "answer": {
+        "type": "choice",
+        "choice": "cancelamento",
+        "confidence": 0.9958236126678967,
+        "probabilities": {
+          "pagamento": 0.000736589116875824,
+          "cancelamento": 0.9968677095009225,
+          "suporte_tecnico": 0.000579527206419017,
+          "sem_correspondencia": 0.0018161741757827666
+        }
+      },
+      "pred_permuted": "cancelamento",
+      "permutation_invariant": true,
+      "correct": true
+    },
+    {
+      "name": "choice:suporte_tecnico",
+      "gold": "suporte_tecnico",
+      "pred": "suporte_tecnico",
+      "latency_ms": 64.4,
+      "answer": {
+        "type": "choice",
+        "choice": "suporte_tecnico",
+        "confidence": 0.9535000077378192,
+        "probabilities": {
+          "pagamento": 0.001133195848223361,
+          "cancelamento": 0.0021639955558770034,
+          "suporte_tecnico": 0.9651250058033644,
+          "sem_correspondencia": 0.031577802792535394
+        }
+      },
+      "pred_permuted": "suporte_tecnico",
+      "permutation_invariant": true,
+      "correct": true
+    },
+    {
+      "name": "choice:sem_correspondencia",
+      "gold": "sem_correspondencia",
+      "pred": "sem_correspondencia",
+      "latency_ms": 66.7,
+      "answer": {
+        "type": "choice",
+        "choice": "sem_correspondencia",
+        "confidence": 0.9874519299503838,
+        "probabilities": {
+          "pagamento": 0.004033476792992324,
+          "cancelamento": 0.0025021007457173418,
+          "suporte_tecnico": 0.0028754749985024732,
+          "sem_correspondencia": 0.9905889474627878
+        }
+      },
+      "pred_permuted": "sem_correspondencia",
+      "permutation_invariant": true,
+      "correct": true
+    },
+    {
+      "name": "noul:clara",
+      "gold": "true",
+      "pred": "true",
+      "latency_ms": 61.8,
+      "answer": {
+        "type": "noul",
+        "noul": 0.947542960858974
+      },
+      "correct": true
+    },
+    {
+      "name": "noul:vaga",
+      "gold": "false",
+      "pred": "false",
+      "latency_ms": 55.3,
+      "answer": {
+        "type": "noul",
+        "noul": 0.4973615950856545
+      },
+      "correct": true
+    },
+    {
+      "name": "score:alta",
+      "gold": "2",
+      "pred": "2",
+      "latency_ms": 55.9,
+      "answer": {
+        "type": "score",
+        "score": 1.9423142016746011,
+        "confidence": 0.9482293944199722,
+        "probabilities": {
+          "0": 0.023172061272047177,
+          "1": 0.011341675781304721,
+          "2": 0.9654862629466482
+        },
+        "legend": {
+          "0": "Baixa",
+          "1": "Média",
+          "2": "Alta"
+        }
+      },
+      "correct": true
+    },
+    {
+      "name": "score:baixa",
+      "gold": "0",
+      "pred": "0",
+      "latency_ms": 55.1,
+      "answer": {
+        "type": "score",
+        "score": 0.2562808099480609,
+        "confidence": 0.624089907903691,
+        "probabilities": {
+          "0": 0.7493932719357941,
+          "1": 0.244932646180351,
+          "2": 0.005674081883854951
+        },
+        "legend": {
+          "0": "Baixa",
+          "1": "Média",
+          "2": "Alta"
+        }
+      },
+      "correct": true
+    }
+  ],
+  "summary": {
+    "contract_violations": [],
+    "sanity_correct": "8/8",
+    "permutation_invariant": "4/4"
+  },
+  "oci": {
+    "region": "sa-saopaulo-1",
+    "run_id": "20261007T171106Z-rizzo_flow-sanity"
+  }
+}

reports/sanity/oci/semif-cpu.json

@@ -0,0 +1,222 @@
+{
+  "candidate": {
+    "name": "semif",
+    "model_id": "semif@851bf6e806ef",
+    "revision": "851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a",
+    "adapter_version": "0.1.0",
+    "device": "cpu",
+    "context_limit": 8192,
+    "probabilities_source": "native"
+  },
+  "device": "cpu",
+  "load_seconds": 6.05,
+  "conformance": [
+    {
+      "name": "valid_choice_noul",
+      "expected": "ok",
+      "got": "ok"
+    },
+    {
+      "name": "valid_score",
+      "expected": "ok",
+      "got": "ok"
+    },
+    {
+      "name": "valid_state_object",
+      "expected": "ok",
+      "got": "ok"
+    },
+    {
+      "name": "invalid_17_options",
+      "expected": "invalid_request",
+      "got": "invalid_request"
+    },
+    {
+      "name": "invalid_single_option",
+      "expected": "invalid_request",
+      "got": "invalid_request"
+    },
+    {
+      "name": "invalid_noul_missing_false",
+      "expected": "invalid_request",
+      "got": "invalid_request"
+    },
+    {
+      "name": "invalid_unknown_type",
+      "expected": "invalid_request",
+      "got": "invalid_request"
+    },
+    {
+      "name": "invalid_empty_state",
+      "expected": "invalid_request",
+      "got": "invalid_request"
+    },
+    {
+      "name": "invalid_no_questions",
+      "expected": "invalid_request",
+      "got": "invalid_request"
+    },
+    {
+      "name": "invalid_extra_field",
+      "expected": "invalid_request",
+      "got": "invalid_request"
+    }
+  ],
+  "sanity": [
+    {
+      "name": "choice:pagamento",
+      "gold": "pagamento",
+      "pred": "pagamento",
+      "latency_ms": 16411.3,
+      "answer": {
+        "type": "choice",
+        "choice": "pagamento",
+        "confidence": 0.9999580355069111,
+        "probabilities": {
+          "pagamento": 0.9999580355069111,
+          "cancelamento": 2.7535293796608728e-05,
+          "suporte_tecnico": 6.540205536604591e-06,
+          "sem_correspondencia": 7.888993755683035e-06
+        }
+      },
+      "pred_permuted": "pagamento",
+      "permutation_invariant": true,
+      "correct": true
+    },
+    {
+      "name": "choice:cancelamento",
+      "gold": "cancelamento",
+      "pred": "cancelamento",
+      "latency_ms": 17233.2,
+      "answer": {
+        "type": "choice",
+        "choice": "cancelamento",
+        "confidence": 0.9982468109808258,
+        "probabilities": {
+          "pagamento": 0.001700633028283725,
+          "cancelamento": 0.9982468109808258,
+          "suporte_tecnico": 3.114818042829917e-05,
+          "sem_correspondencia": 2.1407810462216812e-05
+        }
+      },
+      "pred_permuted": "cancelamento",
+      "permutation_invariant": true,
+      "correct": true
+    },
+    {
+      "name": "choice:suporte_tecnico",
+      "gold": "suporte_tecnico",
+      "pred": "suporte_tecnico",
+      "latency_ms": 16699.8,
+      "answer": {
+        "type": "choice",
+        "choice": "suporte_tecnico",
+        "confidence": 0.9987593932138263,
+        "probabilities": {
+          "pagamento": 0.0011694270195878285,
+          "cancelamento": 3.1164174472760134e-05,
+          "suporte_tecnico": 0.9987593932138263,
+          "sem_correspondencia": 4.001559211311531e-05
+        }
+      },
+      "pred_permuted": "suporte_tecnico",
+      "permutation_invariant": true,
+      "correct": true
+    },
+    {
+      "name": "choice:sem_correspondencia",
+      "gold": "sem_correspondencia",
+      "pred": "sem_correspondencia",
+      "latency_ms": 16903.6,
+      "answer": {
+        "type": "choice",
+        "choice": "sem_correspondencia",
+        "confidence": 0.9850240302415921,
+        "probabilities": {
+          "pagamento": 0.010942628542039331,
+          "cancelamento": 0.0031351155675534968,
+          "suporte_tecnico": 0.0008982256488151342,
+          "sem_correspondencia": 0.9850240302415921
+        }
+      },
+      "pred_permuted": "sem_correspondencia",
+      "permutation_invariant": true,
+      "correct": true
+    },
+    {
+      "name": "noul:clara",
+      "gold": "true",
+      "pred": "true",
+      "latency_ms": 13812.9,
+      "answer": {
+        "type": "noul",
+        "noul": 0.9465966702001758
+      },
+      "correct": true
+    },
+    {
+      "name": "noul:vaga",
+      "gold": "false",
+      "pred": "false",
+      "latency_ms": 13046.2,
+      "answer": {
+        "type": "noul",
+        "noul": 0.014063627043245475
+      },
+      "correct": true
+    },
+    {
+      "name": "score:alta",
+      "gold": "2",
+      "pred": "2",
+      "latency_ms": 13712.6,
+      "answer": {
+        "type": "score",
+        "score": 1.9927177259683637,
+        "confidence": 0.9963109844368413,
+        "probabilities": {
+          "0": 0.0035932584684774537,
+          "1": 9.575709468113931e-05,
+          "2": 0.9963109844368413
+        },
+        "legend": {
+          "0": "Baixa",
+          "1": "Média",
+          "2": "Alta"
+        }
+      },
+      "correct": true
+    },
+    {
+      "name": "score:baixa",
+      "gold": "0",
+      "pred": "1",
+      "latency_ms": 15147.9,
+      "answer": {
+        "type": "score",
+        "score": 0.7051250202229055,
+        "confidence": 0.46831026433271183,
+        "probabilities": {
+          "0": 0.41328235772219135,
+          "1": 0.46831026433271183,
+          "2": 0.11840737794509683
+        },
+        "legend": {
+          "0": "Baixa",
+          "1": "Média",
+          "2": "Alta"
+        }
+      },
+      "correct": false
+    }
+  ],
+  "summary": {
+    "contract_violations": [],
+    "sanity_correct": "7/8",
+    "permutation_invariant": "4/4"
+  },
+  "oci": {
+    "region": "us-chicago-1",
+    "run_id": "20261007T153216Z-semif-sanity"
+  }
+}

reports/sanity/oci/semif-cuda.json

@@ -0,0 +1,223 @@
+{
+  "candidate": {
+    "name": "semif",
+    "model_id": "semif@851bf6e806ef",
+    "revision": "851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a",
+    "adapter_version": "0.1.0",
+    "device": "cuda",
+    "context_limit": 8192,
+    "probabilities_source": "native"
+  },
+  "device": "cuda",
+  "load_seconds": 6.47,
+  "process_guard": "ok",
+  "conformance": [
+    {
+      "name": "valid_choice_noul",
+      "expected": "ok",
+      "got": "ok"
+    },
+    {
+      "name": "valid_score",
+      "expected": "ok",
+      "got": "ok"
+    },
+    {
+      "name": "valid_state_object",
+      "expected": "ok",
+      "got": "ok"
+    },
+    {
+      "name": "invalid_17_options",
+      "expected": "invalid_request",
+      "got": "invalid_request"
+    },
+    {
+      "name": "invalid_single_option",
+      "expected": "invalid_request",
+      "got": "invalid_request"
+    },
+    {
+      "name": "invalid_noul_missing_false",
+      "expected": "invalid_request",
+      "got": "invalid_request"
+    },
+    {
+      "name": "invalid_unknown_type",
+      "expected": "invalid_request",
+      "got": "invalid_request"
+    },
+    {
+      "name": "invalid_empty_state",
+      "expected": "invalid_request",
+      "got": "invalid_request"
+    },
+    {
+      "name": "invalid_no_questions",
+      "expected": "invalid_request",
+      "got": "invalid_request"
+    },
+    {
+      "name": "invalid_extra_field",
+      "expected": "invalid_request",
+      "got": "invalid_request"
+    }
+  ],
+  "sanity": [
+    {
+      "name": "choice:pagamento",
+      "gold": "pagamento",
+      "pred": "pagamento",
+      "latency_ms": 66.5,
+      "answer": {
+        "type": "choice",
+        "choice": "pagamento",
+        "confidence": 0.9999580355069111,
+        "probabilities": {
+          "pagamento": 0.9999580355069111,
+          "cancelamento": 2.7535293796608728e-05,
+          "suporte_tecnico": 6.540205536604591e-06,
+          "sem_correspondencia": 7.888993755683035e-06
+        }
+      },
+      "pred_permuted": "pagamento",
+      "permutation_invariant": true,
+      "correct": true
+    },
+    {
+      "name": "choice:cancelamento",
+      "gold": "cancelamento",
+      "pred": "cancelamento",
+      "latency_ms": 65.9,
+      "answer": {
+        "type": "choice",
+        "choice": "cancelamento",
+        "confidence": 0.9984434836109674,
+        "probabilities": {
+          "pagamento": 0.0015010990652337887,
+          "cancelamento": 0.9984434836109674,
+          "suporte_tecnico": 3.11543171817569e-05,
+          "sem_correspondencia": 2.4263006617207197e-05
+        }
+      },
+      "pred_permuted": "cancelamento",
+      "permutation_invariant": true,
+      "correct": true
+    },
+    {
+      "name": "choice:suporte_tecnico",
+      "gold": "suporte_tecnico",
+      "pred": "suporte_tecnico",
+      "latency_ms": 64.8,
+      "answer": {
+        "type": "choice",
+        "choice": "suporte_tecnico",
+        "confidence": 0.9987677467505007,
+        "probabilities": {
+          "pagamento": 0.001169436800573682,
+          "cancelamento": 2.7502517470556258e-05,
+          "suporte_tecnico": 0.9987677467505007,
+          "sem_correspondencia": 3.531393145509289e-05
+        }
+      },
+      "pred_permuted": "suporte_tecnico",
+      "permutation_invariant": true,
+      "correct": true
+    },
+    {
+      "name": "choice:sem_correspondencia",
+      "gold": "sem_correspondencia",
+      "pred": "sem_correspondencia",
+      "latency_ms": 65.7,
+      "answer": {
+        "type": "choice",
+        "choice": "sem_correspondencia",
+        "confidence": 0.9850240302415921,
+        "probabilities": {
+          "pagamento": 0.010942628542039331,
+          "cancelamento": 0.0031351155675534968,
+          "suporte_tecnico": 0.0008982256488151342,
+          "sem_correspondencia": 0.9850240302415921
+        }
+      },
+      "pred_permuted": "sem_correspondencia",
+      "permutation_invariant": true,
+      "correct": true
+    },
+    {
+      "name": "noul:clara",
+      "gold": "true",
+      "pred": "true",
+      "latency_ms": 68.7,
+      "answer": {
+        "type": "noul",
+        "noul": 0.9465966702001758
+      },
+      "correct": true
+    },
+    {
+      "name": "noul:vaga",
+      "gold": "false",
+      "pred": "false",
+      "latency_ms": 65.0,
+      "answer": {
+        "type": "noul",
+        "noul": 0.01590639171181471
+      },
+      "correct": true
+    },
+    {
+      "name": "score:alta",
+      "gold": "2",
+      "pred": "2",
+      "latency_ms": 63.8,
+      "answer": {
+        "type": "score",
+        "score": 1.991764792381704,
+        "confidence": 0.9958345405394164,
+        "probabilities": {
+          "0": 0.004069748157712474,
+          "1": 9.571130287104325e-05,
+          "2": 0.9958345405394164
+        },
+        "legend": {
+          "0": "Baixa",
+          "1": "Média",
+          "2": "Alta"
+        }
+      },
+      "correct": true
+    },
+    {
+      "name": "score:baixa",
+      "gold": "0",
+      "pred": "0",
+      "latency_ms": 64.7,
+      "answer": {
+        "type": "score",
+        "score": 0.6683472691146671,
+        "confidence": 0.44388424362844425,
+        "probabilities": {
+          "0": 0.44388424362844425,
+          "1": 0.44388424362844425,
+          "2": 0.11223151274311144
+        },
+        "legend": {
+          "0": "Baixa",
+          "1": "Média",
+          "2": "Alta"
+        }
+      },
+      "correct": true
+    }
+  ],
+  "summary": {
+    "contract_violations": [],
+    "sanity_correct": "8/8",
+    "permutation_invariant": "4/4"
+  },
+  "oci": {
+    "region": "sa-saopaulo-1",
+    "run_id": "20261007T163132Z-semif-sanity"
+  }
+}