docs(model-ports): proposed draft validation contracts for pending identities (issue #200)

.github/ci/reference/proposed_model_ports/afm45b.json

@@ -0,0 +1,42 @@
+{
+  "schema_version": 1,
+  "model": {
+    "repo": "arcee-ai/AFM-4.5B",
+    "revision": "e84d19711eca5817c1608cb344d931bf05c966dd",
+    "license": "apache-2.0"
+  },
+  "baseline_artifacts": {
+    "model": {
+      "sha256": "2de4692fc3404ba302701753f4110610d3d8f88cf49294d5ce0f176e01cc8871",
+      "quantization": "Q8_0",
+      "hf_repo": "arcee-ai/AFM-4.5B-GGUF"
+    },
+    "note": "Downloaded from a third-party community GGUF repo, not the original authors -- sha256 verified locally against the exact file used. The quantizer repo's own revision was not pinned at download time (no explicit commit reference captured); flagging honestly rather than guessing one."
+  },
+  "architecture": {
+    "general_architecture": "arcee"
+  },
+  "runtime": {
+    "submodule_path": "ported_models/llama_cpp_et/src/llama.cpp-et",
+    "required_device": "ET",
+    "required_gpu_layers": 99,
+    "require_full_offload": true
+  },
+  "quality": {
+    "perplexity": {
+      "corpus_artifact": "wikitext2_raw_test",
+      "corpus_sha256": "173c87a53759e0201f33e0ccf978e510c2042d7f2cb78229d9a50d79b9e7dd08",
+      "context_size": 128,
+      "batch_size": 128,
+      "ubatch_size": 128,
+      "chunks": 4,
+      "measurement_status": "pending_host_or_board_run",
+      "note": "No PPL number is populated here -- only local sysemu structural verification was performed this session (full ET layer offload, param count matching upstream, clean compute graph; see the recipe). A real first_run_perplexity and regression threshold should come from an actual host/board run, not be assumed."
+    }
+  },
+  "correctness": {
+    "host_reference": "CPU execution of the candidate artifact through the main-owned runtime harness",
+    "measurement_status": "pending_host_or_board_run"
+  },
+  "recipe": "ported_models/afm45b/docs/RECIPE.md"
+}

.github/ci/reference/proposed_model_ports/apertus_8b.json

@@ -0,0 +1,44 @@
+{
+  "schema_version": 1,
+  "model": {
+    "repo": "swiss-ai/Apertus-8B-Instruct-2509",
+    "revision": "b946d40447b2b597999b9c86d44bee0b452c919f",
+    "license": "apache-2.0"
+  },
+  "baseline_artifacts": {
+    "model": {
+      "sha256": "823a2ce1b89f8e31c197167e144fb4e5220f0adb44132b8e54ccdb2dd5bedae7",
+      "quantization": "Q8_0",
+      "hf_repo": "DevQuasar/swiss-ai.Apertus-8B-Instruct-2509-GGUF"
+    },
+    "note": "Downloaded from a third-party community GGUF repo, not the original authors -- sha256 verified locally against the exact file used."
+  },
+  "architecture": {
+    "general_architecture": "apertus"
+  },
+  "runtime": {
+    "submodule_path": "ported_models/llama_cpp_et/src/llama.cpp-et",
+    "required_device": "ET",
+    "required_gpu_layers": 99,
+    "require_full_offload": true
+  },
+  "quality": {
+    "perplexity": {
+      "corpus_artifact": "wikitext2_raw_test",
+      "corpus_sha256": "173c87a53759e0201f33e0ccf978e510c2042d7f2cb78229d9a50d79b9e7dd08",
+      "context_size": 128,
+      "batch_size": 128,
+      "ubatch_size": 128,
+      "chunks": 4,
+      "measurement_status": "measured_host_cpu_only",
+      "host_cpu_perplexity": 7.8747,
+      "host_cpu_perplexity_error": 1.46638,
+      "note": "Real host-CPU perplexity measured this session (plain GGML_ET=OFF build, board-pinned corpus, board's own ctx/batch/chunks params). ET-SoC1 board perplexity is NOT measured -- that needs the maintainer's trusted workflow."
+    }
+  },
+  "correctness": {
+    "host_reference": "CPU execution of the candidate artifact through the main-owned runtime harness",
+    "measurement_status": "measured_host_cpu_only"
+  },
+  "recipe": "ported_models/apertus_8b/docs/RECIPE.md"
+}

.github/ci/reference/proposed_model_ports/chatglm3_6b.json

@@ -0,0 +1,44 @@
+{
+  "schema_version": 1,
+  "model": {
+    "repo": "zai-org/chatglm3-6b",
+    "revision": "e9e0406d062cdb887444fe5bd546833920abd4ac",
+    "license": "apache-2.0"
+  },
+  "baseline_artifacts": {
+    "model": {
+      "sha256": "805e3761a997486eb08f45e0f74efa5c1cc0be8475afa2b170263b490a3ba8c8",
+      "quantization": "Q8_0",
+      "hf_repo": "hellork/chatglm3-6b-128k-Q8_0-GGUF"
+    },
+    "note": "Downloaded from a third-party community GGUF repo, not the original authors -- sha256 verified locally against the exact file used."
+  },
+  "architecture": {
+    "general_architecture": "chatglm"
+  },
+  "runtime": {
+    "submodule_path": "ported_models/llama_cpp_et/src/llama.cpp-et",
+    "required_device": "ET",
+    "required_gpu_layers": 99,
+    "require_full_offload": true
+  },
+  "quality": {
+    "perplexity": {
+      "corpus_artifact": "wikitext2_raw_test",
+      "corpus_sha256": "173c87a53759e0201f33e0ccf978e510c2042d7f2cb78229d9a50d79b9e7dd08",
+      "context_size": 128,
+      "batch_size": 128,
+      "ubatch_size": 128,
+      "chunks": 4,
+      "measurement_status": "measured_host_cpu_only",
+      "host_cpu_perplexity": 15.91,
+      "host_cpu_perplexity_error": 3.3745,
+      "note": "Real host-CPU perplexity measured this session (plain GGML_ET=OFF build, board-pinned corpus, board's own ctx/batch/chunks params). ET-SoC1 board perplexity is NOT measured -- that needs the maintainer's trusted workflow."
+    }
+  },
+  "correctness": {
+    "host_reference": "CPU execution of the candidate artifact through the main-owned runtime harness",
+    "measurement_status": "measured_host_cpu_only"
+  },
+  "recipe": "ported_models/chatglm3_6b/docs/RECIPE.md"
+}

.github/ci/reference/proposed_model_ports/codeshell_7b.json

@@ -0,0 +1,42 @@
+{
+  "schema_version": 1,
+  "model": {
+    "repo": "WisdomShell/CodeShell-7B-Chat",
+    "revision": "3cb06f589b7b1e2f8e728c77280b1114191d24de",
+    "license": "codeshell-license"
+  },
+  "baseline_artifacts": {
+    "model": {
+      "sha256": "207e1ee777455b17d3e58f3174dead52f1c8d007690c96679a7deba355284876",
+      "quantization": "Q8_0",
+      "hf_repo": "DevQuasar/WisdomShell.Shell-7B-Chat-GGUF"
+    },
+    "note": "Downloaded from a third-party community GGUF repo, not the original authors -- sha256 verified locally against the exact file used. The quantizer repo's own revision was not pinned at download time (no explicit commit reference captured); flagging honestly rather than guessing one."
+  },
+  "architecture": {
+    "general_architecture": "codeshell"
+  },
+  "runtime": {
+    "submodule_path": "ported_models/llama_cpp_et/src/llama.cpp-et",
+    "required_device": "ET",
+    "required_gpu_layers": 99,
+    "require_full_offload": true
+  },
+  "quality": {
+    "perplexity": {
+      "corpus_artifact": "wikitext2_raw_test",
+      "corpus_sha256": "173c87a53759e0201f33e0ccf978e510c2042d7f2cb78229d9a50d79b9e7dd08",
+      "context_size": 128,
+      "batch_size": 128,
+      "ubatch_size": 128,
+      "chunks": 4,
+      "measurement_status": "pending_host_or_board_run",
+      "note": "No PPL number is populated here -- only local sysemu structural verification was performed this session (full ET layer offload, param count matching upstream, clean compute graph; see the recipe). A real first_run_perplexity and regression threshold should come from an actual host/board run, not be assumed."
+    }
+  },
+  "correctness": {
+    "host_reference": "CPU execution of the candidate artifact through the main-owned runtime harness",
+    "measurement_status": "pending_host_or_board_run"
+  },
+  "recipe": "ported_models/codeshell_7b/docs/RECIPE.md"
+}

.github/ci/reference/proposed_model_ports/cohere2_7b.json

@@ -0,0 +1,42 @@
+{
+  "schema_version": 1,
+  "model": {
+    "repo": "CohereLabs/c4ai-command-r7b-12-2024",
+    "revision": "4f3d0aa6856e322f2f4480fe65420d5d53d297b8",
+    "license": "cc-by-nc-4.0"
+  },
+  "baseline_artifacts": {
+    "model": {
+      "sha256": "cd7281ec7974bb5810d8f6c922801d6fff3e8bd761073942757dcef990ca9b4d",
+      "quantization": "Q8_0",
+      "hf_repo": "bartowski/c4ai-command-r7b-12-2024-GGUF"
+    },
+    "note": "Downloaded from a third-party community GGUF repo, not the original authors -- sha256 verified locally against the exact file used. The quantizer repo's own revision was not pinned at download time (no explicit commit reference captured); flagging honestly rather than guessing one."
+  },
+  "architecture": {
+    "general_architecture": "cohere2"
+  },
+  "runtime": {
+    "submodule_path": "ported_models/llama_cpp_et/src/llama.cpp-et",
+    "required_device": "ET",
+    "required_gpu_layers": 99,
+    "require_full_offload": true
+  },
+  "quality": {
+    "perplexity": {
+      "corpus_artifact": "wikitext2_raw_test",
+      "corpus_sha256": "173c87a53759e0201f33e0ccf978e510c2042d7f2cb78229d9a50d79b9e7dd08",
+      "context_size": 128,
+      "batch_size": 128,
+      "ubatch_size": 128,
+      "chunks": 4,
+      "measurement_status": "pending_host_or_board_run",
+      "note": "No PPL number is populated here -- only local sysemu structural verification was performed this session (full ET layer offload, param count matching upstream, clean compute graph; see the recipe). A real first_run_perplexity and regression threshold should come from an actual host/board run, not be assumed."
+    }
+  },
+  "correctness": {
+    "host_reference": "CPU execution of the candidate artifact through the main-owned runtime harness",
+    "measurement_status": "pending_host_or_board_run"
+  },
+  "recipe": "ported_models/cohere2_7b/docs/RECIPE.md"
+}

.github/ci/reference/proposed_model_ports/decilm7b.json

@@ -0,0 +1,42 @@
+{
+  "schema_version": 1,
+  "model": {
+    "repo": "Deci/DeciLM-7B-instruct",
+    "revision": "4adc7aa9efe61b47b0a98b2cc94527d9c45c3b4f",
+    "license": "apache-2.0"
+  },
+  "baseline_artifacts": {
+    "model": {
+      "sha256": "1b60beae4c9d5d7131f846f7545cb386e20d8f2c613844a0142ee701537dbfe3",
+      "quantization": "Q8_0",
+      "hf_repo": "Deci/DeciLM-7B-instruct-GGUF"
+    },
+    "note": "Downloaded from a third-party community GGUF repo, not the original authors -- sha256 verified locally against the exact file used. The quantizer repo's own revision was not pinned at download time (no explicit commit reference captured); flagging honestly rather than guessing one."
+  },
+  "architecture": {
+    "general_architecture": "deci"
+  },
+  "runtime": {
+    "submodule_path": "ported_models/llama_cpp_et/src/llama.cpp-et",
+    "required_device": "ET",
+    "required_gpu_layers": 99,
+    "require_full_offload": true
+  },
+  "quality": {
+    "perplexity": {
+      "corpus_artifact": "wikitext2_raw_test",
+      "corpus_sha256": "173c87a53759e0201f33e0ccf978e510c2042d7f2cb78229d9a50d79b9e7dd08",
+      "context_size": 128,
+      "batch_size": 128,
+      "ubatch_size": 128,
+      "chunks": 4,
+      "measurement_status": "pending_host_or_board_run",
+      "note": "No PPL number is populated here -- only local sysemu structural verification was performed this session (full ET layer offload, param count matching upstream, clean compute graph; see the recipe). A real first_run_perplexity and regression threshold should come from an actual host/board run, not be assumed."
+    }
+  },
+  "correctness": {
+    "host_reference": "CPU execution of the candidate artifact through the main-owned runtime harness",
+    "measurement_status": "pending_host_or_board_run"
+  },
+  "recipe": "ported_models/decilm7b/docs/RECIPE.md"
+}

.github/ci/reference/proposed_model_ports/ernie45_03b.json

@@ -0,0 +1,42 @@
+{
+  "schema_version": 1,
+  "model": {
+    "repo": "baidu/ERNIE-4.5-0.3B-PT",
+    "revision": "b565cf6caebdb7a1eadf00100857b1ed5e044f12",
+    "license": "apache-2.0"
+  },
+  "baseline_artifacts": {
+    "model": {
+      "sha256": "022ce1cfc46a2818a96b8a683aec0944047cac6745f35aa5a2a48729a58caa74",
+      "quantization": "Q8_0",
+      "hf_repo": "bartowski/baidu_ERNIE-4.5-0.3B-PT-GGUF"
+    },
+    "note": "Downloaded from a third-party community GGUF repo, not the original authors -- sha256 verified locally against the exact file used. The quantizer repo's own revision was not pinned at download time (no explicit commit reference captured); flagging honestly rather than guessing one."
+  },
+  "architecture": {
+    "general_architecture": "ernie4_5"
+  },
+  "runtime": {
+    "submodule_path": "ported_models/llama_cpp_et/src/llama.cpp-et",
+    "required_device": "ET",
+    "required_gpu_layers": 99,
+    "require_full_offload": true
+  },
+  "quality": {
+    "perplexity": {
+      "corpus_artifact": "wikitext2_raw_test",
+      "corpus_sha256": "173c87a53759e0201f33e0ccf978e510c2042d7f2cb78229d9a50d79b9e7dd08",
+      "context_size": 128,
+      "batch_size": 128,
+      "ubatch_size": 128,
+      "chunks": 4,
+      "measurement_status": "pending_host_or_board_run",
+      "note": "No PPL number is populated here -- only local sysemu structural verification was performed this session (full ET layer offload, param count matching upstream, clean compute graph; see the recipe). A real first_run_perplexity and regression threshold should come from an actual host/board run, not be assumed."
+    }
+  },
+  "correctness": {
+    "host_reference": "CPU execution of the candidate artifact through the main-owned runtime harness",
+    "measurement_status": "pending_host_or_board_run"
+  },
+  "recipe": "ported_models/ernie45_03b/docs/RECIPE.md"
+}

.github/ci/reference/proposed_model_ports/falcon7b.json

@@ -0,0 +1,42 @@
+{
+  "schema_version": 1,
+  "model": {
+    "repo": "tiiuae/falcon-7b",
+    "revision": "ec89142b67d748a1865ea4451372db8313ada0d8",
+    "license": "apache-2.0"
+  },
+  "baseline_artifacts": {
+    "model": {
+      "sha256": "47e0a8ef8eb3d55765f7d412a1d9b816159fece3e9f87ce79caa7e6d9b2341a8",
+      "quantization": "Q8_0",
+      "hf_repo": "maddes8cht/tiiuae-falcon-7b-gguf"
+    },
+    "note": "Downloaded from a third-party community GGUF repo, not the original authors -- sha256 verified locally against the exact file used. The quantizer repo's own revision was not pinned at download time (no explicit commit reference captured); flagging honestly rather than guessing one."
+  },
+  "architecture": {
+    "general_architecture": "falcon"
+  },
+  "runtime": {
+    "submodule_path": "ported_models/llama_cpp_et/src/llama.cpp-et",
+    "required_device": "ET",
+    "required_gpu_layers": 99,
+    "require_full_offload": true
+  },
+  "quality": {
+    "perplexity": {
+      "corpus_artifact": "wikitext2_raw_test",
+      "corpus_sha256": "173c87a53759e0201f33e0ccf978e510c2042d7f2cb78229d9a50d79b9e7dd08",
+      "context_size": 128,
+      "batch_size": 128,
+      "ubatch_size": 128,
+      "chunks": 4,
+      "measurement_status": "pending_host_or_board_run",
+      "note": "No PPL number is populated here -- only local sysemu structural verification was performed this session (full ET layer offload, param count matching upstream, clean compute graph; see the recipe). A real first_run_perplexity and regression threshold should come from an actual host/board run, not be assumed."
+    }
+  },
+  "correctness": {
+    "host_reference": "CPU execution of the candidate artifact through the main-owned runtime harness",
+    "measurement_status": "pending_host_or_board_run"
+  },
+  "recipe": "ported_models/falcon7b/docs/RECIPE.md"
+}

.github/ci/reference/proposed_model_ports/falcon_h1_1_5b.json

@@ -0,0 +1,42 @@
+{
+  "schema_version": 1,
+  "model": {
+    "repo": "tiiuae/Falcon-H1-1.5B-Instruct",
+    "revision": "80ebc50d7799a440b96c93bb6686a3924a09b0cb",
+    "license": "falcon-llm-license"
+  },
+  "baseline_artifacts": {
+    "model": {
+      "sha256": "0044322ad5aaea5ccc2264f83f430f4ab2d20acc6db71a12d79b3f5acb0f69d5",
+      "quantization": "Q8_0",
+      "hf_repo": "tiiuae/Falcon-H1-1.5B-Instruct-GGUF"
+    },
+    "note": "Official quantization repo from the model's own authors."
+  },
+  "architecture": { "general_architecture": "falcon-h1" },
+  "runtime": {
+    "submodule_path": "ported_models/llama_cpp_et/src/llama.cpp-et",
+    "required_device": "ET",
+    "required_gpu_layers": 99,
+    "require_full_offload": true
+  },
+  "quality": {
+    "perplexity": {
+      "corpus_artifact": "wikitext2_raw_test",
+      "corpus_sha256": "173c87a53759e0201f33e0ccf978e510c2042d7f2cb78229d9a50d79b9e7dd08",
+      "context_size": 128,
+      "batch_size": 128,
+      "ubatch_size": 128,
+      "chunks": 4,
+      "measurement_status": "measured_host_cpu_only",
+      "host_cpu_perplexity": 11.6708,
+      "host_cpu_perplexity_error": 2.60807,
+      "note": "Real host-CPU perplexity measured this session (plain GGML_ET=OFF build, board-pinned corpus, board's own ctx/batch/chunks params). ET-SoC1 board perplexity is NOT measured -- that needs the maintainer's trusted workflow."
+    }
+  },
+  "correctness": {
+    "host_reference": "CPU execution of the candidate artifact through the main-owned runtime harness",
+    "measurement_status": "measured_host_cpu_only"
+  },
+  "recipe": "ported_models/falcon_h1_1_5b/docs/RECIPE.md"
+}