Initial public ModelForge release

2026-09-01 21:30:16 +02:00
commit 7082ab955a
490 changed files with 104252 additions and 0 deletions
@@ -0,0 +1,14 @@
+suite:
+  id: examplerag-retrieval-v1
+  capability: rag.embedding
+  version: 1
+  purpose: Compare embedding deployments against representative ExampleRAG retrieval cases.
+  dataset_revision: pending-curated-fixture
+  runnable: false
+metrics:
+  required: [recall_at_5, recall_at_10, mrr, ndcg_at_10]
+performance:
+  capture: [latency_p50, latency_p95, peak_vram_mb, throughput_items_per_second]
+release_policy:
+  forbid_unexplained_regression_percent: 2.0
+  requires_project_owner_review: true
@@ -0,0 +1,13 @@
+suite:
+  id: examplevision-card-retrieval-v1
+  capability: vision.embedding
+  version: 1
+  purpose: Rank the correct Pokémon card among visually similar catalog candidates.
+  dataset_revision: pending-curated-fixture
+  runnable: false
+metrics:
+  required: [recall_at_1, recall_at_5, recall_at_10, mrr]
+performance:
+  capture: [latency_p50, latency_p95, peak_vram_mb]
+release_policy:
+  requires_project_owner_review: true
@@ -0,0 +1,35 @@
+capability: assistant.general
+version: 1
+description: General local text and multimodal assistant workloads.
+input_schema:
+  type: object
+  required: [messages]
+  properties:
+    messages: {type: array, minItems: 1}
+    response_schema: {type: object}
+output_schema:
+  type: object
+  required: [content]
+  properties:
+    content: {type: string}
+    structured_output: {type: object}
+    tool_calls: {type: array}
+modalities: {input: [text, image], output: [text, json, tool_call]}
+languages: [nl, en]
+streaming: true
+structured_output: true
+quality_metrics: [task_success, instruction_following, structured_output_validity]
+upgrade_class: behavioral
+fallback: {allowed: false, mode: hard_fail}
+privacy: {classification: confidential, allow_persistence: false, allow_logging_payloads: false, allow_network_egress: false}
+resources: {accelerator_required: true, cpu_fallback_allowed: false}
+production_priority: interactive
+default_residency: keep_warm
+estate:
+  category: ASSISTANTS
+  purpose: Future local general assistant contract; not an M9 executable baseline.
+  stability: planned
+  resource_class: EXCLUSIVE_GPU
+  evaluation_type: llm
+  consumers: [example-ops]
+  payload_limits: {max_bytes: 262144, max_batch_count: 1}
@@ -0,0 +1,24 @@
+capability: assistant.vision
+version: 1
+description: Visual question answering and structured image reasoning.
+input_schema: {type: object, required: [image, prompt], properties: {image: {type: string}, prompt: {type: string}}}
+output_schema: {type: object, required: [content], properties: {content: {type: string}, structured_output: {type: object}}}
+modalities: {input: [image, text], output: [text, json]}
+languages: [nl, en]
+streaming: false
+structured_output: true
+quality_metrics: [task_success, structured_output_validity]
+upgrade_class: behavioral
+fallback: {allowed: false, mode: hard_fail}
+privacy: {classification: confidential, allow_persistence: false, allow_logging_payloads: false, allow_network_egress: false}
+resources: {accelerator_required: true, cpu_fallback_allowed: false}
+production_priority: interactive
+default_residency: load_on_demand
+estate:
+  category: ASSISTANTS
+  purpose: Future local image reasoning contract; not an M9 executable baseline.
+  stability: planned
+  resource_class: EXCLUSIVE_GPU
+  evaluation_type: llm
+  consumers: [examplevision, example-ops]
+  payload_limits: {max_bytes: 8388608, max_batch_count: 1, max_width: 4096, max_height: 4096}
@@ -0,0 +1,38 @@
+capability: document.ocr
+version: 1
+description: OCR and structured document extraction.
+input_schema:
+  type: object
+  additionalProperties: false
+  required: [content_base64, media_type]
+  properties:
+    content_base64: {type: string, minLength: 4, maxLength: 11184812}
+    media_type: {enum: [image/png, image/jpeg]}
+    language_hint: {enum: [nl, en, auto]}
+output_schema:
+  type: object
+  additionalProperties: false
+  required: [text, pages]
+  properties:
+    text: {type: string}
+    pages: {type: array, minItems: 1, maxItems: 1}
+    blocks: {type: array}
+    confidence: {type: [number, 'null']}
+modalities: {input: [image, document], output: [text, json]}
+languages: [nl, en]
+structured_output: true
+quality_metrics: [cer, wer, field_accuracy, layout_accuracy]
+upgrade_class: schema_breaking
+fallback: {allowed: false, mode: hard_fail}
+privacy: {classification: restricted, allow_persistence: false, allow_logging_payloads: false, allow_network_egress: false}
+resources: {accelerator_required: true, cpu_fallback_allowed: false}
+production_priority: background
+default_residency: load_on_demand
+estate:
+  category: DOCUMENT
+  purpose: Offline text extraction from bounded local document images.
+  stability: experimental
+  resource_class: MEDIUM
+  evaluation_type: ocr
+  consumers: [examplerag, examplevision, example-ops]
+  payload_limits: {max_bytes: 8388608, max_batch_count: 1, max_width: 4096, max_height: 4096, max_pages: 1}
@@ -0,0 +1,40 @@
+capability: rag.embedding
+version: 1
+description: Multilingual text embeddings for semantic retrieval.
+input_schema:
+  type: object
+  additionalProperties: false
+  required: [input]
+  properties:
+    input:
+      oneOf:
+        - {type: string, minLength: 1, maxLength: 8192}
+        - {type: array, items: {type: string, minLength: 1, maxLength: 8192}, minItems: 1, maxItems: 8}
+output_schema:
+  type: object
+  additionalProperties: false
+  required: [vectors, dimension, normalized, embedding_space_id]
+  properties:
+    vectors: {type: array, minItems: 1, maxItems: 8, items: {type: array, minItems: 1024, maxItems: 1024, items: {type: number}}}
+    dimension: {const: 1024}
+    normalized: {const: true}
+    embedding_space_id: {type: string, format: uuid}
+modalities: {input: [text], output: [vector]}
+languages: [nl, en]
+vector: {dimensionality: 1024, normalized: true, cross_deployment_compatible: false}
+slo: {latency_p95_ms: 15000, availability_percent: 99.0, max_concurrency: 1}
+quality_metrics: [recall_at_5, recall_at_10, mrr, ndcg_at_10]
+upgrade_class: requires_reindex
+fallback: {allowed: false, mode: hard_fail}
+privacy: {classification: confidential, allow_persistence: false, allow_logging_payloads: false, allow_network_egress: false}
+resources: {accelerator_required: true, cpu_fallback_allowed: false}
+production_priority: production
+default_residency: keep_warm
+estate:
+  category: RAG
+  purpose: Multilingual semantic retrieval for local knowledge projects.
+  stability: stable
+  resource_class: MEDIUM
+  evaluation_type: retrieval
+  consumers: [examplerag, portfolio-assistant]
+  payload_limits: {max_bytes: 65536, max_batch_count: 8}
@@ -0,0 +1,54 @@
+capability: rag.reranking
+version: 1
+description: Second-stage ranking of retrieved text candidates.
+input_schema:
+  type: object
+  additionalProperties: false
+  required: [query, documents, top_n]
+  properties:
+    query: {type: string, minLength: 1, maxLength: 4000}
+    documents:
+      type: array
+      minItems: 1
+      maxItems: 40
+      items:
+        type: object
+        additionalProperties: false
+        required: [id, text]
+        properties:
+          id: {type: string, minLength: 1, maxLength: 255}
+          text: {type: string, minLength: 1, maxLength: 16384}
+    top_n: {type: integer, minimum: 1, maximum: 10}
+output_schema:
+  type: object
+  additionalProperties: false
+  required: [results]
+  properties:
+    results:
+      type: array
+      maxItems: 10
+      items:
+        type: object
+        additionalProperties: false
+        required: [id, score, rank]
+        properties:
+          id: {type: string}
+          score: {type: number}
+          rank: {type: integer, minimum: 1, maximum: 10}
+modalities: {input: [text], output: [score]}
+languages: [nl, en]
+quality_metrics: [mrr, ndcg_at_10, ranking_accuracy]
+upgrade_class: behavioral
+fallback: {allowed: false, mode: hard_fail}
+privacy: {classification: confidential, allow_persistence: false, allow_logging_payloads: false, allow_network_egress: false}
+resources: {accelerator_required: true, cpu_fallback_allowed: false}
+production_priority: production
+default_residency: load_on_demand
+estate:
+  category: RAG
+  purpose: Evidence-gated second-stage ranking for retrieval pipelines.
+  stability: experimental
+  resource_class: MEDIUM
+  evaluation_type: retrieval
+  consumers: [examplerag]
+  payload_limits: {max_bytes: 262144, max_batch_count: 40}
@@ -0,0 +1,39 @@
+capability: speech.transcription
+version: 1
+description: Offline bounded speech-to-text transcription.
+input_schema:
+  type: object
+  additionalProperties: false
+  required: [audio_base64, media_type]
+  properties:
+    audio_base64: {type: string, minLength: 4, maxLength: 22369624}
+    media_type: {const: audio/wav}
+    language: {enum: [nl, en, fr, auto]}
+output_schema:
+  type: object
+  additionalProperties: false
+  required: [text, language, duration_seconds]
+  properties:
+    text: {type: string}
+    language: {type: string}
+    duration_seconds: {type: number}
+    segments: {type: array}
+modalities: {input: [audio], output: [text, json]}
+languages: [nl, en, fr]
+streaming: false
+structured_output: true
+quality_metrics: [wer, real_time_factor]
+upgrade_class: behavioral
+fallback: {allowed: false, mode: hard_fail}
+privacy: {classification: restricted, allow_persistence: false, allow_logging_payloads: false, allow_network_egress: false}
+resources: {accelerator_required: true, cpu_fallback_allowed: false}
+production_priority: background
+default_residency: load_on_demand
+estate:
+  category: AUDIO
+  purpose: Private local transcription of bounded WAV audio without cloud fallback.
+  stability: experimental
+  resource_class: HEAVY
+  evaluation_type: asr
+  consumers: []
+  payload_limits: {max_bytes: 16777216, max_batch_count: 1, max_duration_seconds: 120}
@@ -0,0 +1,39 @@
+capability: vision.embedding
+version: 1
+description: Image/text embeddings for visual retrieval.
+input_schema:
+  type: object
+  additionalProperties: false
+  required: [items]
+  properties:
+    items:
+      type: array
+      minItems: 1
+      maxItems: 4
+      items:
+        oneOf:
+          - {type: object, additionalProperties: false, required: [image_base64, media_type], properties: {image_base64: {type: string}, media_type: {enum: [image/png, image/jpeg]}}}
+          - {type: object, additionalProperties: false, required: [text], properties: {text: {type: string, minLength: 1, maxLength: 4096}}}
+output_schema:
+  type: object
+  required: [vectors]
+  properties:
+    vectors: {type: array, items: {type: array, items: {type: number}}}
+modalities: {input: [image, text], output: [vector]}
+languages: [nl, en]
+vector: {dimensionality: null, normalized: null, cross_deployment_compatible: false}
+quality_metrics: [recall_at_1, recall_at_5, recall_at_10, mrr]
+upgrade_class: requires_reindex
+fallback: {allowed: false, mode: hard_fail}
+privacy: {classification: confidential, allow_persistence: false, allow_logging_payloads: false, allow_network_egress: false}
+resources: {accelerator_required: true, cpu_fallback_allowed: false}
+production_priority: interactive
+default_residency: load_on_demand
+estate:
+  category: VISION
+  purpose: Shared image and text vector space for local visual retrieval.
+  stability: experimental
+  resource_class: HEAVY
+  evaluation_type: visual-retrieval
+  consumers: [examplevision]
+  payload_limits: {max_bytes: 8388608, max_batch_count: 4, max_width: 4096, max_height: 4096}
@@ -0,0 +1,97 @@
+# Seed discovery list only. Nothing in this file is implicitly production-approved.
+# Exact upstream IDs, revisions, licenses, runtimes and hardware compatibility must be verified.
+
+candidates:
+  - id: qwen-general
+    display_name: Qwen3.5-9B
+    source: Qwen/Qwen3.5-9B
+    intended_capabilities: [assistant.general, assistant.vision]
+    proposed_role: General and vision assistant candidate
+    preferred_runtime: vllm
+
+  - id: gpt-oss-reasoning
+    display_name: gpt-oss-20b
+    source: openai/gpt-oss-20b
+    intended_capabilities: [assistant.reasoning]
+    proposed_role: Reasoning and tool-use candidate
+    preferred_runtime: vllm
+
+  - id: qwen-coder
+    display_name: Qwen2.5-Coder-14B-Instruct
+    source: Qwen/Qwen2.5-Coder-14B-Instruct
+    intended_capabilities: [assistant.coding]
+    proposed_role: Coding assistant candidate
+    preferred_runtime: vllm
+
+  - id: qwen-embed-fast
+    display_name: Qwen3-Embedding-0.6B
+    source: Qwen/Qwen3-Embedding-0.6B
+    intended_capabilities: [rag.embedding.fast]
+    proposed_role: Latency-oriented embedding candidate
+
+  - id: qwen-embed-quality
+    display_name: Qwen3-Embedding-4B
+    source: Qwen/Qwen3-Embedding-4B
+    intended_capabilities: [rag.embedding.quality]
+    proposed_role: Quality-oriented embedding candidate
+
+  - id: qwen-reranker-fast
+    display_name: Qwen3-Reranker-0.6B
+    source: Qwen/Qwen3-Reranker-0.6B
+    intended_capabilities: [rag.reranking.fast]
+    proposed_role: Latency-oriented reranking candidate
+
+  - id: qwen-reranker-quality
+    display_name: Qwen3-Reranker-4B
+    source: Qwen/Qwen3-Reranker-4B
+    intended_capabilities: [rag.reranking.quality]
+    proposed_role: Quality-oriented reranking candidate
+
+  - id: bge-m3
+    display_name: BGE-M3
+    source: BAAI/bge-m3
+    intended_capabilities: [rag.hybrid_embedding]
+    proposed_role: Hybrid retrieval reference candidate
+
+  - id: paddleocr-vl
+    display_name: PaddleOCR-VL
+    source: PaddlePaddle/PaddleOCR-VL-1.5
+    intended_capabilities: [document.ocr, document.layout]
+    proposed_role: OCR and document-layout candidate
+
+  - id: siglip2
+    display_name: SigLIP2 So400m
+    source: google/siglip2-so400m-patch14-384
+    intended_capabilities: [vision.embedding]
+    proposed_role: Visual retrieval candidate
+
+  - id: sam21
+    display_name: SAM 2.1 Large
+    source: facebook/sam2.1-hiera-large
+    intended_capabilities: [vision.segmentation]
+    proposed_role: Image segmentation candidate
+
+  - id: dinov3
+    display_name: DINOv3 ViT-L
+    source: facebook/dinov3-vitl16-pretrain-lvd1689m
+    intended_capabilities: [vision.features]
+    proposed_role: Visual feature extraction candidate
+
+  - id: whisper-turbo
+    display_name: Whisper large-v3-turbo
+    source: openai/whisper-large-v3-turbo
+    intended_capabilities: [speech.transcription]
+    proposed_role: Speech transcription candidate
+
+  - id: chatterbox
+    display_name: Chatterbox Multilingual
+    source: ResembleAI/chatterbox
+    intended_capabilities: [speech.synthesis]
+    proposed_role: Speech synthesis candidate
+
+  - id: flux-klein
+    display_name: FLUX.2 klein 4B
+    source: black-forest-labs/FLUX.2-klein-4B
+    intended_capabilities: [image.generation, image.editing]
+    proposed_role: Image generation and editing candidate
+    preferred_runtime: diffusers
@@ -0,0 +1,24 @@
+security:
+  trust_remote_code: false
+  prefer_safetensors: true
+  require_exact_revision_for_approval: true
+  require_artifact_digest: true
+  inference_network_egress: false
+  expose_runtime_workers: false
+
+lifecycle:
+  automatic_production_promotion: false
+  require_local_benchmark_for_stable: true
+  retain_previous_stable_as_rollback_target: true
+  destructive_cleanup_requires_explicit_approval: true
+
+scheduler:
+  priority_order: [production, interactive, background, benchmark, maintenance]
+  reserve_vram_mb: 1024
+  default_warm_ttl_seconds: 900
+  benchmark_may_disrupt_production: false
+
+upgrades:
+  default_behavioral_change_class: behavioral
+  vector_producer_default_change_class: requires_reindex
+  schema_change_requires_contract_version_bump: true
@@ -0,0 +1,15 @@
+project:
+  id: example-ops
+  name: ExampleOps
+  description: Deterministic network configuration generation from structured requirements.
+
+bindings:
+  assistant.general:
+    contract_version: 1
+    channel: stable
+    priority: interactive
+    benchmark_suites: [netops-requirements-extraction-v1]
+
+notes:
+  - LLM output must be normalized into a validated intermediate representation.
+  - Final device CLI must be generated by deterministic logic, not direct unconstrained LLM output.
@@ -0,0 +1,24 @@
+project:
+  id: examplerag
+  name: ExampleRAG
+  description: Central retrieval-augmented-generation platform.
+
+bindings:
+  rag.embedding:
+    contract_version: 1
+    channel: stable
+    priority: production
+    migration_support: reindex
+    benchmark_suites: [examplerag-retrieval-v1]
+
+  rag.reranking:
+    contract_version: 1
+    channel: stable
+    priority: production
+    benchmark_suites: [examplerag-reranking-v1]
+
+  document.ocr:
+    contract_version: 1
+    channel: stable
+    priority: background
+    optional: true
@@ -0,0 +1,34 @@
+project:
+  id: examplevision
+  name: ExampleVision
+  description: Pokémon card collection, recognition and value tracking.
+
+bindings:
+  vision.embedding:
+    contract_version: 1
+    channel: experimental
+    priority: interactive
+    migration_support: reindex
+    benchmark_suites: [examplevision-card-retrieval-v1]
+
+  document.ocr:
+    contract_version: 1
+    channel: experimental
+    priority: interactive
+    benchmark_suites: [examplevision-card-ocr-v1]
+
+  speech.transcription:
+    contract_version: 1
+    channel: experimental
+    priority: interactive
+    benchmark_suites: [examplevision-spoken-card-search-v1]
+
+  assistant.vision:
+    contract_version: 1
+    channel: stable
+    priority: interactive
+    benchmark_suites: [examplevision-card-reasoning-v1]
+
+notes:
+  - Vision embedding and speech transcription are shadow integrations; neither changes the production recognizer.
+  - Document OCR remains experimental until a safe candidate passes the project quality gate.
@@ -0,0 +1,23 @@
+{
+  "schema_version": 1,
+  "reviewed_at": "2026-09-01",
+  "review_reference": "docs/quality/V1_2_1_RELEASE_VERIFICATION.md",
+  "vulnerabilities": [
+    {"vulnerability_id": "CVE-2026-41992", "package": "gzip", "installed_version": "1.13-1", "severity": "HIGH"},
+    {"vulnerability_id": "CVE-2026-54369", "package": "libacl1", "installed_version": "2.3.2-2+b1", "severity": "HIGH"},
+    {"vulnerability_id": "CVE-2025-69720", "package": "libncursesw6", "installed_version": "6.5+20250216-2", "severity": "HIGH"},
+    {"vulnerability_id": "CVE-2026-11822", "package": "libsqlite3-0", "installed_version": "3.46.1-7+deb13u1", "severity": "HIGH"},
+    {"vulnerability_id": "CVE-2026-11824", "package": "libsqlite3-0", "installed_version": "3.46.1-7+deb13u1", "severity": "HIGH"},
+    {"vulnerability_id": "CVE-2025-69720", "package": "libtinfo6", "installed_version": "6.5+20250216-2", "severity": "HIGH"},
+    {"vulnerability_id": "CVE-2025-69720", "package": "ncurses-base", "installed_version": "6.5+20250216-2", "severity": "HIGH"},
+    {"vulnerability_id": "CVE-2025-69720", "package": "ncurses-bin", "installed_version": "6.5+20250216-2", "severity": "HIGH"},
+    {"vulnerability_id": "CVE-2026-13221", "package": "perl-base", "installed_version": "5.40.1-6", "severity": "CRITICAL"},
+    {"vulnerability_id": "CVE-2026-42496", "package": "perl-base", "installed_version": "5.40.1-6", "severity": "CRITICAL"},
+    {"vulnerability_id": "CVE-2026-8376", "package": "perl-base", "installed_version": "5.40.1-6", "severity": "CRITICAL"},
+    {"vulnerability_id": "CVE-2026-42497", "package": "perl-base", "installed_version": "5.40.1-6", "severity": "HIGH"},
+    {"vulnerability_id": "CVE-2026-48962", "package": "perl-base", "installed_version": "5.40.1-6", "severity": "HIGH"},
+    {"vulnerability_id": "CVE-2026-57432", "package": "perl-base", "installed_version": "5.40.1-6", "severity": "HIGH"},
+    {"vulnerability_id": "CVE-2026-57433", "package": "perl-base", "installed_version": "5.40.1-6", "severity": "HIGH"},
+    {"vulnerability_id": "CVE-2026-9538", "package": "perl-base", "installed_version": "5.40.1-6", "severity": "HIGH"}
+  ]
+}