{
  "schema": "karti.model-recipe/v1",
  "model_name": "Karti-Small-Agent-3B",
  "serving_alias": "karti-3b",
  "status": "active-private-model-program",
  "weights_published": false,
  "private_training_rows_published": false,
  "base_model": {
    "repository": "HuggingFaceTB/SmolLM3-3B",
    "revision": "a07cc9a04f16550a088caea529712d1d335b0ac1",
    "license": "Apache-2.0"
  },
  "training": {
    "framework": "TRL",
    "first_stage": "BF16 LoRA SFT",
    "sequence_length": 2048,
    "seed": 115,
    "trainer_is_external": true,
    "later_stage": "optional verifier-driven GRPO after held-out improvement",
    "first_private_candidate": {
      "completed_on": "2026-08-26",
      "verified_training_rows": 512,
      "optimizer_steps": 64,
      "status": "first-cycle-complete"
    },
    "latest_candidate": {
      "status": "private-candidate-under-evaluation",
      "adapter": "LoRA",
      "epochs": 3,
      "optimizer_steps": 288,
      "learning_rate": 0.0001,
      "training_rows": 768,
      "held_out_rows": 224,
      "promoted": false
    },
    "serving": {
      "engine": "vLLM",
      "chat_template": "tool-call-aware Hermes rendering",
      "train_serve_token_parity": "required, verified byte-for-byte before a score is trusted"
    }
  },
  "public_bootstrap": [
    {
      "repository": "acon96/Home-Assistant-Requests",
      "revision": "75f5abe4a3fc4f30b6ccfeb239f4da30d22b3619",
      "config": "default",
      "license": "MIT",
      "use": "user-utterance seed only"
    },
    {
      "repository": "NousResearch/hermes-function-calling-v1",
      "revision": "dae3e1d28cfbcf4b915c04ea1e072030529b4bda",
      "config": "func_calling_singleturn",
      "license": "Apache-2.0",
      "use": "safe user-utterance seed only"
    },
    {
      "repository": "Team-ACE/ToolACE",
      "revision": "6bda777c88d21e5a204703c1ee45597a8fa4f734",
      "config": "default",
      "license": "Apache-2.0",
      "use": "safe user-utterance seed only"
    }
  ],
  "public_bootstrap_policy": {
    "training_eligible_as_downloaded": false,
    "original_labels_retained": false,
    "requires_private_relabel_and_verification": true,
    "gated_dataset_mirrors_forbidden": true
  },
  "private_data_policy": {
    "contents_disclosed": false,
    "required_gates": [
      "rights-or-consent",
      "redaction",
      "deduplication-before-split",
      "temporal-holdout",
      "provenance-digest"
    ],
    "forbidden": [
      "credentials",
      "controller-tokens",
      "precise-private-addresses",
      "unredacted-third-party-personal-data"
    ]
  },
  "evaluation": {
    "reward_schema": "karti.agent-tools.reward/v1",
    "components": [
      "exact-tool-sequence-and-arguments",
      "response-contract",
      "proposal-only-confirmation-boundary"
    ],
    "real_device_execution": false,
    "implementation_visibility": "private",
    "first_cycle": {
      "scope": "clean diagnostic evaluation",
      "status": "complete",
      "outcome": "continued private iteration",
      "next_focus": [
        "clarification coverage",
        "strict response contracts",
        "tool-call structure"
      ]
    },
    "promotion_policy": "held-out improvement plus owner review",
    "harness": {
      "framework": "Prime Intellect Verifiers",
      "runtime": "containerised agent, tools called over MCP",
      "held_out_tasks": 12,
      "decomposed_components": [
        "tool_selection",
        "argument_exactness",
        "routing",
        "policy_adherence",
        "efficiency",
        "task_success",
        "result_grounding"
      ],
      "hard_gates": [
        "critical policy violations"
      ]
    },
    "published_scores": "withheld until measured on a verified train/serve rendering path"
  },
  "release_formats": {
    "canonical": "private BF16 merged Safetensors plus the private PEFT adapter and pinned base revision",
    "primary_serving": "private GPTQ W4A16 group-128 compressed Safetensors for vLLM",
    "optional_fallback": "private Q4_K_M GGUF for llama.cpp after separate speed and tool-contract gates",
    "lineage": [
      "derive serving builds from the current canonical checkpoint",
      "never train from GGUF or quantized serving weights",
      "never requantize a previous serving release"
    ],
    "every_variant_requires_held_out_evaluation": true
  },
  "publication": {
    "recipe_and_model_card_only": true,
    "weights": "private; public release intended after a promotion gate passes",
    "adapters": "private",
    "training_rows": "private",
    "credentials": "never-published",
    "cadence": "public checkpoint on promotion, not on a fixed calendar"
  },
  "corpus": {
    "kind": "first-party synthetic, deterministic and regenerable from a manifest",
    "behaviour_classes": 14,
    "train_rows": 768,
    "eval_rows": 224,
    "train_eval_key_intersection": 0,
    "rows_with_no_tool_call": 0,
    "every_row_delivers_through_a_tool": true,
    "contains_real_world_content": false,
    "tool_surface": [
      "browser",
      "cron",
      "phone",
      "reply",
      "route",
      "tera"
    ],
    "tool_surface_note": "every row terminates by delivering through `reply`; no tool executes a confirmed action"
  }
}