Add machine-readable serving profile catalog

ARTIFACT_MANIFEST.json

@@ -20,7 +20,7 @@
       "warning_checks": 0
     }
   },
-  "generated_at": "2026-06-28T00:52:18.227594+00:00",
+  "generated_at": "2026-06-28T00:58:55.373853+00:00",
   "known_remaining_gaps": [
     "No release-candidate SFT/post-training adapter exists because SFT was stopped and checkpoint-8000 failed behavior gates.",
     "The recommended strong repair blend and resume preflight are prepared CPU-side, but no resumed GPU training has been run from them.",
@@ -146,18 +146,18 @@
     }
   ],
   "release_readiness": {
-    "latest_report_generated_at": "2026-06-28T00:51:35.069342+00:00",
+    "latest_report_generated_at": "2026-06-28T00:58:25.555064+00:00",
     "markdown_report": "probes/release-readiness-report.md",
     "note": "Serving GGUF release readiness is separated from full project completion because the SFT adapter remains unreleased.",
     "report": "probes/release-readiness-report.json",
     "script": "scripts/build_release_readiness_report.py",
     "summary": {
-      "checks": 22,
+      "checks": 25,
       "failed_checks": 0,
       "full_project_status": "incomplete",
       "serving_release_status": "pass",
       "status_counts": {
-        "pass": 17,
+        "pass": 20,
         "warn": 5
       },
       "warning_checks": 5
@@ -183,7 +183,7 @@
     "script": "scripts/verify_hf_release.py"
   },
   "remote_file_count_before_upload": 241,
-  "remote_head_before_upload": "7dcf02d87d07e4b173d7b375e9fc747605d80652",
+  "remote_head_before_upload": "16abb00acc4a1c89dbc5c357252fe677205931cd",
   "repair_artifacts": {
     "code_only_extract_format": {
       "assistant_code_fence": 0.0,
@@ -293,37 +293,54 @@
     "sft_status": "stopped; no GPU training is running; no LoRA adapter released"
   },
   "repo_id": "LordNeel/Ornith-1.0-35B-GGUF-llamacpp-tp1",
-  "staged_file_count_excluding_generated_audit_files": 213,
+  "serving_profiles": {
+    "config": "configs/serving_profiles.yaml",
+    "latest_report_generated_at": "2026-06-28T00:56:49.528851+00:00",
+    "markdown_report": "probes/serving-profile-validation.md",
+    "note": "Machine-readable single-GPU llama.cpp serving profiles with Q4 default, Q3 lowest-memory, IQ4_XS middle-footprint, and adaptive MTP entries.",
+    "report": "probes/serving-profile-validation.json",
+    "script": "scripts/validate_serving_profiles.py",
+    "summary": {
+      "checks": 43,
+      "failed_checks": 0,
+      "status": "pass",
+      "status_counts": {
+        "pass": 43
+      },
+      "warnings": 0
+    }
+  },
+  "staged_file_count_excluding_generated_audit_files": 217,
   "staged_files": [
     {
-      "bytes": 12198,
+      "bytes": 12795,
       "path": "MODEL_INVESTIGATION_HARNESS.md",
       "remote_before_upload": {
         "blob_id": "11d19ea8434b8a6b792692c62612340ab47f22c4",
         "bytes": 12198,
         "exists": true
       },
-      "sha256": "5de5c74adce1e5594cc45066be1982a0d2165b453620bbf90e42f0403f55dccf"
+      "sha256": "0275706ac68b747d6ab6b2e65131f8d621b109937939779aa354227effeb367f"
     },
     {
-      "bytes": 8575,
+      "bytes": 9338,
       "path": "PROJECT_STATUS_AUDIT.md",
       "remote_before_upload": {
         "blob_id": "12abd37fee2e13cdc59ac09f34ad41047c37532c",
         "bytes": 8575,
         "exists": true
       },
-      "sha256": "6fb7495e6102142f15e7f828a63afd5191364876783cbb60b556873ed47cf764"
+      "sha256": "1f38f8f0ef4085cfbc0a7e708af86f2ea200130e8df84316e3a5f1258e36f3e9"
     },
     {
-      "bytes": 10523,
+      "bytes": 11157,
       "path": "README.md",
       "remote_before_upload": {
-        "blob_id": "e95d5292315da9d193e383e8830dc8e9fb41b046",
-        "bytes": 10499,
+        "blob_id": "fcbe763783bc5975c28adec65acd1c66a9b7cbdf",
+        "bytes": 10523,
         "exists": true
       },
-      "sha256": "33910886eb0e18d0b41441c9a515841d30e10d3a1ee15479db54968816141566"
+      "sha256": "f67e32ef769791b92ea113ccc825d783edae9e514ed4304f114d80da551a365d"
     },
     {
       "bytes": 12029,
@@ -1626,14 +1643,22 @@
       "sha256": "84efb88c5d571d9942e118cd9cde458fda9ce327b5e71b671c55c9c2836a6a43"
     },
     {
-      "bytes": 2287,
+      "bytes": 2630,
       "path": "configs/serving_matrix.yaml",
       "remote_before_upload": {
         "blob_id": "63d4c96c5aaace66eff3a86ce72c379c948422bc",
         "bytes": 2287,
         "exists": true
       },
-      "sha256": "0c8f6d4f79ca9867c417121320020e8d5f3fb08fc6644d3ffc694ad2b671f812"
+      "sha256": "580b74248e913b8208bbe040267b8bc40e440c7b8643b7621884dfa1a49f86c1"
+    },
+    {
+      "bytes": 5432,
+      "path": "configs/serving_profiles.yaml",
+      "remote_before_upload": {
+        "exists": false
+      },
+      "sha256": "f3a8638b6aae46ac552984f655a39a000f3ebf188524cc24a28ae6e97cbdeee6"
     },
     {
       "bytes": 5232,
@@ -1936,24 +1961,40 @@
       "sha256": "fa01e4d58c37d581786ad78b0b773547acefb1dbce0a08373ca36eb6c7229eb7"
     },
     {
-      "bytes": 8562,
+      "bytes": 9923,
       "path": "probes/release-readiness-report.json",
       "remote_before_upload": {
-        "blob_id": "8447a31f6294be99561ce4de7bfac79736299419",
+        "blob_id": "6b35d98f285cb24ab5b918bc78143860a85d1d97",
         "bytes": 8562,
         "exists": true
       },
-      "sha256": "2634236a47b00ed8dd8f279c3f890771b766432d2e952f6482232eb5ee983085"
+      "sha256": "7cb6f2ca33e64d751a833a16cbed4206321fcf031b5f09fcb85c09e88d423e4d"
     },
     {
-      "bytes": 3017,
+      "bytes": 3287,
       "path": "probes/release-readiness-report.md",
       "remote_before_upload": {
-        "blob_id": "c5d5658958875c290da1aa18753ef4c58969869a",
+        "blob_id": "d0cc1a5f7a22ef1656f0b17850127d0cbb20d325",
         "bytes": 3017,
         "exists": true
       },
-      "sha256": "50c326ab01c6cd617c84e163c7cb4328dac6b37a413410c6c5c1183b4d302652"
+      "sha256": "92ea357c21517714878c1024543ca2100da8b80f0c23a0590dbeaa22d76b68e0"
+    },
+    {
+      "bytes": 17664,
+      "path": "probes/serving-profile-validation.json",
+      "remote_before_upload": {
+        "exists": false
+      },
+      "sha256": "1e272a261e35d60c710890471a3adb2dbc03e4685a3ee179fadf6bed48c1dac0"
+    },
+    {
+      "bytes": 5737,
+      "path": "probes/serving-profile-validation.md",
+      "remote_before_upload": {
+        "exists": false
+      },
+      "sha256": "0e4d13af8d1f1c7d7694361e16e0d121f04c45700b89fe773d01c8534e6b7824"
     },
     {
       "bytes": 1267,
@@ -2186,14 +2227,14 @@
       "sha256": "cf2d52ac14ac23b9c84aa35a8a4c965afed5573f711221411eb93acd634ef526"
     },
     {
-      "bytes": 15496,
+      "bytes": 16856,
       "path": "scripts/build_release_readiness_report.py",
       "remote_before_upload": {
         "blob_id": "a7dc67e2fce48ab06d6a9afe0c24f8e992b4c064",
         "bytes": 15496,
         "exists": true
       },
-      "sha256": "fac895b30e79c25037428bbf7118de4b4cd6d8f6ee7ef90f68d9dbcf28f3861c"
+      "sha256": "05190a4ce370af14c88853a566b52410d62abd562b49179a55bb2fe3bc6c42d0"
     },
     {
       "bytes": 4467,
@@ -2405,6 +2446,14 @@
       },
       "sha256": "340edc3ee4e104e9bf8145195b05ce9f106c63f3b128c32d181ed0512140868a"
     },
+    {
+      "bytes": 9335,
+      "path": "scripts/validate_serving_profiles.py",
+      "remote_before_upload": {
+        "exists": false
+      },
+      "sha256": "7cd47c2fb7b723daa2f1b493e6cc8642499eac0f50057463786ed4dac26cc9f2"
+    },
     {
       "bytes": 15231,
       "path": "scripts/validate_sft_jsonl.py",
@@ -2416,14 +2465,14 @@
       "sha256": "b1db75a43b9b3415cba337b868638a496f960cf492bbdfa4c31ad978c277908c"
     },
     {
-      "bytes": 19576,
+      "bytes": 20402,
       "path": "scripts/verify_hf_release.py",
       "remote_before_upload": {
         "blob_id": "8aa8b46de86b2f23902e09b03ecd0d76d6079759",
         "bytes": 19576,
         "exists": true
       },
-      "sha256": "5d7e71481492ee07056e035be32ed72b7e46d949b159eeef0cb9b492f820e5e5"
+      "sha256": "a4afcad099491fa52fbaffe10927966c82ac9c4c8d532144432e441a8f31bdd0"
     }
   ]
 }

MODEL_INVESTIGATION_HARNESS.md

@@ -100,6 +100,22 @@ For Ornith this reports `serving_release_status=pass` and
 serving release is usable, while the fine-tuning/post-training deliverable
 remains unreleased after the stopped SFT run.
 
+Validate the machine-readable serving profile catalog:
+
+```bash
+.venv-lab/bin/python scripts/validate_serving_profiles.py \
+  --catalog configs/serving_profiles.yaml \
+  --manifest hf_upload/ARTIFACT_MANIFEST.json \
+  --hf-root hf_upload \
+  --output runs/serving-profile-validation.json \
+  --markdown-output runs/serving-profile-validation.md \
+  --fail-on-error
+```
+
+The current catalog validates 5 single-GPU llama.cpp profiles with 43/43 checks
+passing: Q4_K_M short-context speed, Q4_K_M coding-context default, Q3_K_M
+lowest memory, IQ4_XS middle footprint, and adaptive IQ4_XS MTP graft.
+
 ## Serving And Benchmark Flow
 
 Start a single-GPU llama.cpp server:

PROJECT_STATUS_AUDIT.md

@@ -18,7 +18,7 @@ release-candidate adapter.
 | Requirement | Status | Evidence | Notes |
 |---|---|---|---|
 | Single-GPU serving only, `tp=1` | Complete | `hf_upload/ARTIFACT_MANIFEST.json`, `hf_upload/RELEASE_AUDIT.md`, `hf_upload/README.md` | Multi-GPU serving is explicitly out of scope. Both GPUs are used only as independent experiment lanes. |
-| Fast serving profile | Complete for GGUF | `benchmarks/llamacpp-quant-benchmarks.md`, `scripts/serve_llamacpp_gpu0.sh`, `scripts/serve_llamacpp_gpu1.sh` | Recommended speed profile is llama.cpp `Q4_K_M` with `REASONING=off`. |
+| Fast serving profile | Complete for GGUF | `configs/serving_profiles.yaml`, `runs/serving-profile-validation.json`, `benchmarks/llamacpp-quant-benchmarks.md`, `scripts/serve_llamacpp_gpu0.sh`, `scripts/serve_llamacpp_gpu1.sh` | Recommended default is llama.cpp `Q4_K_M` coding context with `REASONING=off`. The serving catalog validates 5 single-GPU profiles with 43/43 checks passing. |
 | Integrated MTP serving profile | Complete with adaptive policy | `benchmarks/llamacpp-iq4-xs-mtp-graft-adaptive-profile.md`, `patches/llamacpp-mtp-adaptive-serving-and-recurrent-rollback.patch` | Low concurrency benefits from MTP; saturated c16 should use adaptive throttling via `LLAMA_SPEC_MAX_DRAFTING_SLOTS=1`. Always-on MTP is not recommended for high-concurrency batches. |
 | Find/fix model or serving issues | Complete for discovered issues | `probes/investigation-ornith-q4-live.json`, `benchmarks/llamacpp-q4-reasoning-off-fix.md`, llama.cpp patches under `patches/` | Found reasoning-content empty-final issue, vLLM GGUF corruption, upstream config-only MTP gap, and llama.cpp MTP recurrent/scheduler issues. Documented/fixed where locally controllable. |
 | Reusable bug-discovery harness | Complete | `MODEL_INVESTIGATION_HARNESS.md`, `configs/model_behavior_suite.yaml`, `scripts/run_model_investigation.py`, `scripts/probe_model_behavior.py`, `scripts/summarize_behavior_matrix.py`, `runs/behavior-regression-matrix.json` | Published to HF with dependencies and gate scripts. Metadata-only smoke passed with 13 checks, 0 failures, 2 low-severity notes. The behavior matrix compares 7 artifacts across 14 reusable checks and separates 5 passing artifacts from 2 known-bad artifacts. |
@@ -32,6 +32,7 @@ release-candidate adapter.
 | Reusable HF release verifier | Complete | `scripts/verify_hf_release.py`, `runs/hf-release-verification-final.json` | Verifies model-card metadata, required support files, staged file hashes, generated release-file hashes, all published GGUF artifacts, and SFT gate summaries. Latest local run passed 21/21 checks with 0 warnings. |
 | Reusable behavior regression matrix | Complete | `scripts/summarize_behavior_matrix.py`, `runs/behavior-regression-matrix.json`, `runs/behavior-regression-matrix.md` | CPU-only aggregation over existing behavior/investigation reports. Current run: 7 artifacts, 98 checks, 5 passing artifacts, 2 failing artifacts, 9 failed checks, and 6 blocking issues in known-bad profiles. |
 | Reusable release-readiness gate | Complete | `scripts/build_release_readiness_report.py`, `runs/release-readiness-report.json`, `runs/release-readiness-report.md` | CPU-only gate over HF readback, manifest, behavior matrix, quant/KLD docs, MTP policy, required docs, and SFT state. Current run: serving release `pass`, full project `incomplete`, 22 checks, 0 failures, 5 warnings. |
+| Machine-readable serving profile catalog | Complete | `configs/serving_profiles.yaml`, `scripts/validate_serving_profiles.py`, `runs/serving-profile-validation.json`, `runs/serving-profile-validation.md` | Catalog captures Q4_K_M speed/default, Q3_K_M lowest-memory, IQ4_XS middle-footprint, and adaptive IQ4_XS MTP graft profiles. Validation passes 43/43 checks against manifest, scripts, evidence docs, and tp=1 policy. |
 | Keep GPUs idle after SFT stop | Complete at audit time | `nvidia-smi`: GPU0 10 MiB / 0%, GPU1 2 MiB / 0%; process scan empty | No `train_sft_lora`, watcher, eval, server, quantize, or convert processes were running. |
 
 ## Current Release State
@@ -40,6 +41,8 @@ release-candidate adapter.
 - Last verified remote commit before this status-audit refresh: `04a18958f05ac002c2ff846d7b5ced6361f9af30`
 - Latest local HF verifier report: `runs/hf-release-verification-final.json` (`21/21` checks passed, `0` warnings)
 - Published HF verifier entrypoint: `scripts/verify_hf_release.py`
+- Serving profile catalog: `configs/serving_profiles.yaml`
+- Latest serving profile validation: `runs/serving-profile-validation.json` and `runs/serving-profile-validation.md`
 - Latest behavior regression matrix: `runs/behavior-regression-matrix.json` and `runs/behavior-regression-matrix.md`
 - Latest release-readiness report: `runs/release-readiness-report.json` and `runs/release-readiness-report.md`
 - Release audit: `RELEASE_AUDIT.md`

README.md

@@ -37,6 +37,8 @@ This workspace tracks optimization work for `deepreinforce-ai/Ornith-1.0-35B`.
 Current reusable workflow: [MODEL_INVESTIGATION_HARNESS.md](MODEL_INVESTIGATION_HARNESS.md).
 Current requirement audit: [PROJECT_STATUS_AUDIT.md](PROJECT_STATUS_AUDIT.md).
 SFT remediation plan: [SFT_REMEDIATION_PLAN.md](SFT_REMEDIATION_PLAN.md).
+Serving profile catalog: [configs/serving_profiles.yaml](configs/serving_profiles.yaml)
+with validation report [probes/serving-profile-validation.md](probes/serving-profile-validation.md).
 Behavior regression matrix: [probes/behavior-regression-matrix.md](probes/behavior-regression-matrix.md)
 via [scripts/summarize_behavior_matrix.py](scripts/summarize_behavior_matrix.py).
 Release readiness report: [probes/release-readiness-report.md](probes/release-readiness-report.md)
@@ -125,6 +127,7 @@ Build the CPU-only release-readiness report:
 ```bash
 .venv-lab/bin/python scripts/build_release_readiness_report.py \
   --allow-unreleased-sft \
+  --serving-profile-validation hf_upload/probes/serving-profile-validation.json \
   --required-doc hf_upload/README.md \
   --required-doc hf_upload/RELEASE_AUDIT.md \
   --required-doc hf_upload/PROJECT_STATUS_AUDIT.md \
@@ -138,6 +141,18 @@ Build the CPU-only release-readiness report:
 The current readiness report passes the GGUF serving release and marks the full
 project incomplete because no SFT/post-training adapter is released.
 
+Validate the machine-readable serving profiles:
+
+```bash
+.venv-lab/bin/python scripts/validate_serving_profiles.py \
+  --catalog configs/serving_profiles.yaml \
+  --manifest hf_upload/ARTIFACT_MANIFEST.json \
+  --hf-root hf_upload \
+  --output runs/serving-profile-validation.json \
+  --markdown-output runs/serving-profile-validation.md \
+  --fail-on-error
+```
+
 Run reusable behavior probes against any OpenAI-compatible backend:
 
 ```bash

RELEASE_AUDIT.md

@@ -1,13 +1,14 @@
 # Release Audit
 
-Generated: 2026-06-28T00:52:18.227594+00:00
+Generated: 2026-06-28T00:58:55.373853+00:00
 
 - HF repo: `LordNeel/Ornith-1.0-35B-GGUF-llamacpp-tp1`
-- Remote head before this upload: `7dcf02d87d07e4b173d7b375e9fc747605d80652`
+- Remote head before this upload: `16abb00acc4a1c89dbc5c357252fe677205931cd`
 - Remote files before this upload: 241
-- Staged support files excluding generated audit files: 213
+- Staged support files excluding generated audit files: 217
 - Overall status: `pass for serving/quant/MTP support files; SFT adapter remains unreleased`
 - Latest HF release verifier: `pass` (`25` checks, `0` failures, `1` expected SFT warning)
+- Latest serving profile validation: `pass` (`43/43` checks)
 - Latest behavior regression matrix: `5` passing artifacts, `2` known-bad failing artifacts
 - Latest release readiness: serving release `pass`, full project `incomplete`
 
@@ -72,12 +73,22 @@ Generated: 2026-06-28T00:52:18.227594+00:00
 - Known-bad artifacts are intentionally included: the old reasoning-on Q4
   profile and the stopped checkpoint-8000 LoRA adapter.
 
+## Serving Profiles
+
+- Catalog: `configs/serving_profiles.yaml`
+- Validator: `scripts/validate_serving_profiles.py`
+- Latest report: `probes/serving-profile-validation.json`
+- Markdown summary: `probes/serving-profile-validation.md`
+- Current run: `43/43` checks passed, `0` failures, `0` warnings.
+- Profiles: Q4_K_M speed/default, Q3_K_M lowest-memory, IQ4_XS
+  middle-footprint, and adaptive IQ4_XS MTP graft.
+
 ## Release Readiness
 
 - Script: `scripts/build_release_readiness_report.py`
 - Latest report: `probes/release-readiness-report.json`
 - Markdown summary: `probes/release-readiness-report.md`
-- Current run: `22` checks, `0` failures, `5` warnings.
+- Current run: `25` checks, `0` failures, `5` warnings.
 - Serving GGUF release status: `pass`.
 - Full project status: `incomplete` because no SFT/post-training adapter is
   released.

configs/serving_matrix.yaml

@@ -13,16 +13,20 @@ policy:
     - multi_gpu_serving
 
 current_baseline:
-  date: "2026-06-26"
-  winner: vllm_bf16_tp1
-  reason: higher aggregate output throughput and lower p95 TTFT than the current SGLang profile at concurrency 4+
+  date: "2026-06-28"
+  winner: llamacpp_gguf_q4_k_m_tp1
+  reason: verified GGUF release path; Q4_K_M is the fastest validated llama.cpp quant, vLLM GGUF remains unsupported for this model family, and BF16 baselines are retained only as historical notes
   constraints:
     tensor_parallel_size: 1
     max_model_len: 131072
-    dtype: bfloat16
+    quant: Q4_K_M
+    backend: llama.cpp
+    reasoning: off
   benchmark_notes:
-    vllm_profile: runs/vllm-initial-profile.md
-    sglang_profile: runs/sglang-initial-profile.md
+    serving_profiles: configs/serving_profiles.yaml
+    quant_benchmarks: benchmarks/llamacpp-quant-benchmarks.md
+    behavior_matrix: probes/behavior-regression-matrix.md
+    release_readiness: probes/release-readiness-report.md
 
 serving_candidates:
   - name: sglang_bf16_tp1
@@ -57,6 +61,8 @@ serving_candidates:
     expected_strengths:
       - fast_gguf_loading
       - small_memory_footprint
+      - verified_hf_release_artifact
+      - reasoning_off_final_answer_mode
 
 benchmarks:
   smoke:

configs/serving_profiles.yaml

@@ -0,0 +1,173 @@
+model:
+  hf_repo: deepreinforce-ai/Ornith-1.0-35B
+  release_repo: LordNeel/Ornith-1.0-35B-GGUF-llamacpp-tp1
+  served_name: Ornith-1.0-35B
+
+policy:
+  serving_parallelism: tp1_only
+  tensor_parallel_size: 1
+  backend: llama.cpp
+  reasoning: "off"
+  gpu_usage_note: Use one independent model copy per GPU. Do not split one serving replica across GPUs.
+  forbidden:
+    - tensor_parallel_size_gt_1
+    - multi_gpu_serving
+
+default_profile: llamacpp_q4_k_m_coding_context
+lowest_memory_profile: llamacpp_q3_k_m_short_context
+middle_footprint_profile: llamacpp_iq4_xs_short_context
+adaptive_mtp_profile: llamacpp_iq4_xs_mtp_graft_adaptive
+
+profiles:
+  - id: llamacpp_q4_k_m_short_context
+    status: recommended_speed_benchmark
+    backend: llama.cpp
+    tensor_parallel_size: 1
+    artifact: ornith-1.0-35b-Q4_K_M.gguf
+    quant: Q4_K_M
+    command:
+      script: scripts/serve_llamacpp_gpu0.sh
+      env:
+        QUANT: Q4_K_M
+        PORT: "8002"
+        CTX_SIZE: "8192"
+        PARALLEL: "16"
+        BATCH_SIZE: "4096"
+        UBATCH_SIZE: "512"
+        REASONING: "off"
+    measured:
+      loaded_vram_gib: 21.31
+      c1_output_tps: 243.3
+      c16_output_tps: 655.6
+      c16_p95_ttft_ms: 650.0
+      behavior_gate: pass
+    evidence:
+      benchmark: benchmarks/llamacpp-quant-benchmarks.md
+      behavior: probes/behavior-llamacpp-q4-script-default.json
+    notes:
+      - Short-context benchmark profile exposes about 512 tokens per slot because CTX_SIZE is split across PARALLEL slots.
+
+  - id: llamacpp_q4_k_m_coding_context
+    status: recommended_default
+    backend: llama.cpp
+    tensor_parallel_size: 1
+    artifact: ornith-1.0-35b-Q4_K_M.gguf
+    quant: Q4_K_M
+    command:
+      script: scripts/serve_llamacpp_gpu0.sh
+      env:
+        QUANT: Q4_K_M
+        PORT: "8002"
+        CTX_SIZE: "131072"
+        PARALLEL: "16"
+        BATCH_SIZE: "4096"
+        UBATCH_SIZE: "512"
+        REASONING: "off"
+    measured:
+      loaded_vram_gib_short_context: 21.31
+      c16_output_tps_short_context: 655.6
+      behavior_gate: pass
+    evidence:
+      benchmark: benchmarks/llamacpp-quant-benchmarks.md
+      behavior: probes/behavior-llamacpp-q4-script-default.json
+    notes:
+      - Practical coding-serving profile with 16 slots and about 8k context per slot.
+      - Throughput numbers come from the matched short-context sweep.
+
+  - id: llamacpp_q3_k_m_short_context
+    status: lowest_memory_validated
+    backend: llama.cpp
+    tensor_parallel_size: 1
+    artifact: ornith-1.0-35b-Q3_K_M.gguf
+    quant: Q3_K_M
+    command:
+      script: scripts/serve_llamacpp_gpu0.sh
+      env:
+        QUANT: Q3_K_M
+        PORT: "8002"
+        CTX_SIZE: "8192"
+        PARALLEL: "16"
+        BATCH_SIZE: "4096"
+        UBATCH_SIZE: "512"
+        REASONING: "off"
+    measured:
+      gguf_gib: 15.61
+      loaded_vram_gib: 17.27
+      c1_output_tps: 240.5
+      c16_output_tps: 493.0
+      mean_kld_nats_vs_bf16_top64: 0.3620
+      behavior_gate: pass
+    evidence:
+      benchmark: benchmarks/llamacpp-quant-benchmarks.md
+      kld: benchmarks/kld-quant-vs-bf16-top64.md
+      behavior: probes/behavior-llamacpp-q3-k-m-parallel16.json
+
+  - id: llamacpp_iq4_xs_short_context
+    status: middle_footprint_validated
+    backend: llama.cpp
+    tensor_parallel_size: 1
+    artifact: ornith-1.0-35b-IQ4_XS.gguf
+    quant: IQ4_XS
+    command:
+      script: scripts/serve_llamacpp_gpu0.sh
+      env:
+        QUANT: IQ4_XS
+        PORT: "8002"
+        CTX_SIZE: "8192"
+        PARALLEL: "16"
+        BATCH_SIZE: "4096"
+        UBATCH_SIZE: "512"
+        REASONING: "off"
+    measured:
+      gguf_gib: 17.64
+      loaded_vram_gib: 19.34
+      c1_output_tps: 234.1
+      c16_output_tps: 476.0
+      mean_kld_nats_vs_bf16_top64: 0.1426
+      behavior_gate: pass
+    evidence:
+      benchmark: benchmarks/llamacpp-quant-benchmarks.md
+      kld: benchmarks/kld-quant-vs-bf16-top64.md
+      behavior: probes/behavior-llamacpp-iq4-xs-parallel16.json
+
+  - id: llamacpp_iq4_xs_mtp_graft_adaptive
+    status: experimental_adaptive_mtp
+    backend: llama.cpp
+    tensor_parallel_size: 1
+    artifact: ornith-1.0-35b-IQ4_XS-MTP-graft-headQ6.gguf
+    quant: IQ4_XS_with_Q6_MTP_head
+    command:
+      script: scripts/serve_llamacpp_gpu0.sh
+      env:
+        MODEL_PATH: ornith-1.0-35b-IQ4_XS-MTP-graft-headQ6.gguf
+        ALIAS: Ornith-1.0-35B-GGUF-IQ4_XS-MTP
+        PORT: "8002"
+        CTX_SIZE: "8192"
+        PARALLEL: "16"
+        BATCH_SIZE: "4096"
+        UBATCH_SIZE: "512"
+        REASONING: "off"
+        SPEC_TYPE: draft-mtp
+        SPEC_DRAFT_N_MAX: "2"
+        SPEC_DRAFT_N_MIN: "0"
+        SPEC_DRAFT_BACKEND_SAMPLING: "1"
+        LLAMA_SPEC_MAX_DRAFTING_SLOTS: "1"
+        LLAMA_MTP_FAST_BACKEND_SAMPLE: "1"
+        LLAMA_MTP_DRAFT_TOP_K: "1"
+        LLAMA_MTP_DRAFT_TOP_P: "1"
+        LLAMA_MTP_DRAFT_TEMP: "1"
+    measured:
+      c1_ar_tps: 221.80
+      c1_mtp_adaptive_tps: 319.53
+      c1_acceptance_by_position:
+        - 0.953
+        - 0.865
+      c16_ar_tps: 568.57
+      c16_adaptive_tps: 564.48
+      c16_threshold1_tps: 592.45
+    evidence:
+      benchmark: benchmarks/llamacpp-iq4-xs-mtp-graft-adaptive-profile.md
+      patch: patches/llamacpp-mtp-adaptive-serving-and-recurrent-rollback.patch
+    caveats:
+      - Always-on MTP is not recommended at saturated c16 because it was slower than target-only decoding.
+      - Adaptive throttling uses MTP for low-concurrency requests and disables it under saturated batches.

probes/release-readiness-report.json

@@ -180,6 +180,52 @@
       "recommendation": "Keep known-bad profiles labeled and do not promote them as release candidates.",
       "status": "warn"
     },
+    {
+      "detail": "serving profile validation report exists",
+      "evidence": [
+        "hf_upload/probes/serving-profile-validation.json"
+      ],
+      "metrics": {},
+      "name": "serving_profiles:validation_exists",
+      "recommendation": null,
+      "status": "pass"
+    },
+    {
+      "detail": "serving profile validation status is 'pass'",
+      "evidence": [
+        "hf_upload/probes/serving-profile-validation.json"
+      ],
+      "metrics": {
+        "checks": 43,
+        "failed_checks": 0,
+        "status": "pass",
+        "status_counts": {
+          "pass": 43
+        },
+        "warnings": 0
+      },
+      "name": "serving_profiles:validation_status",
+      "recommendation": null,
+      "status": "pass"
+    },
+    {
+      "detail": "5 serving profiles are validated",
+      "evidence": [
+        "hf_upload/probes/serving-profile-validation.json"
+      ],
+      "metrics": {
+        "profiles": [
+          "llamacpp_q4_k_m_short_context",
+          "llamacpp_q4_k_m_coding_context",
+          "llamacpp_q3_k_m_short_context",
+          "llamacpp_iq4_xs_short_context",
+          "llamacpp_iq4_xs_mtp_graft_adaptive"
+        ]
+      },
+      "name": "serving_profiles:profile_count",
+      "recommendation": null,
+      "status": "pass"
+    },
     {
       "detail": "stopped; no GPU training is running; no LoRA adapter released",
       "evidence": [
@@ -280,22 +326,23 @@
       "status": "pass"
     }
   ],
-  "generated_at": "2026-06-28T00:51:35.069342+00:00",
+  "generated_at": "2026-06-28T00:58:25.555064+00:00",
   "inputs": {
     "behavior_matrix": "hf_upload/probes/behavior-regression-matrix.json",
     "hf_verification": "hf_upload/probes/hf-release-verification.json",
     "kld_report": "hf_upload/benchmarks/kld-quant-vs-bf16-top64.md",
     "manifest": "hf_upload/ARTIFACT_MANIFEST.json",
     "mtp_profile": "hf_upload/benchmarks/llamacpp-iq4-xs-mtp-graft-adaptive-profile.md",
-    "quant_benchmarks": "hf_upload/benchmarks/llamacpp-quant-benchmarks.md"
+    "quant_benchmarks": "hf_upload/benchmarks/llamacpp-quant-benchmarks.md",
+    "serving_profile_validation": "hf_upload/probes/serving-profile-validation.json"
   },
   "summary": {
-    "checks": 22,
+    "checks": 25,
     "failed_checks": 0,
     "full_project_status": "incomplete",
     "serving_release_status": "pass",
     "status_counts": {
-      "pass": 17,
+      "pass": 20,
       "warn": 5
     },
     "warning_checks": 5

probes/release-readiness-report.md

@@ -1,12 +1,12 @@
 # Release Readiness Report
 
-Generated: 2026-06-28T00:51:35.069342+00:00
+Generated: 2026-06-28T00:58:25.555064+00:00
 
 ## Summary
 
 - Serving release status: `pass`
 - Full project status: `incomplete`
-- Checks: 22
+- Checks: 25
 - Failures: 0
 - Warnings: 5
 
@@ -27,6 +27,9 @@ Generated: 2026-06-28T00:51:35.069342+00:00
 | `behavior_matrix:exists` | `pass` | behavior regression matrix exists |
 | `behavior_matrix:passing_artifacts` | `pass` | 5 artifacts pass the behavior matrix |
 | `behavior_matrix:known_bad_artifacts_disclosed` | `warn` | 2 failing artifacts are present in the matrix for regression visibility |
+| `serving_profiles:validation_exists` | `pass` | serving profile validation report exists |
+| `serving_profiles:validation_status` | `pass` | serving profile validation status is 'pass' |
+| `serving_profiles:profile_count` | `pass` | 5 serving profiles are validated |
 | `sft:release_adapter_state` | `warn` | stopped; no GPU training is running; no LoRA adapter released |
 | `docs:README.md` | `pass` | required evidence doc README.md exists |
 | `docs:RELEASE_AUDIT.md` | `pass` | required evidence doc RELEASE_AUDIT.md exists |

probes/serving-profile-validation.json

@@ -0,0 +1,566 @@
+{
+  "checks": [
+    {
+      "detail": "catalog exists: configs/serving_profiles.yaml",
+      "name": "catalog:exists",
+      "status": "pass"
+    },
+    {
+      "detail": "serving policy is single-GPU tp=1",
+      "metrics": {
+        "backend": "llama.cpp",
+        "forbidden": [
+          "tensor_parallel_size_gt_1",
+          "multi_gpu_serving"
+        ],
+        "gpu_usage_note": "Use one independent model copy per GPU. Do not split one serving replica across GPUs.",
+        "reasoning": "off",
+        "serving_parallelism": "tp1_only",
+        "tensor_parallel_size": 1
+      },
+      "name": "policy:tp1_only",
+      "status": "pass"
+    },
+    {
+      "detail": "multi-GPU serving and tensor_parallel_size>1 are explicitly forbidden",
+      "metrics": {
+        "forbidden": [
+          "multi_gpu_serving",
+          "tensor_parallel_size_gt_1"
+        ]
+      },
+      "name": "policy:multi_gpu_forbidden",
+      "status": "pass"
+    },
+    {
+      "detail": "default serving policy disables reasoning final-answer leakage",
+      "metrics": {
+        "reasoning": "off"
+      },
+      "name": "policy:reasoning_off",
+      "status": "pass"
+    },
+    {
+      "detail": "5 serving profiles are listed",
+      "metrics": {
+        "profiles": [
+          "llamacpp_q4_k_m_short_context",
+          "llamacpp_q4_k_m_coding_context",
+          "llamacpp_q3_k_m_short_context",
+          "llamacpp_iq4_xs_short_context",
+          "llamacpp_iq4_xs_mtp_graft_adaptive"
+        ]
+      },
+      "name": "profiles:count",
+      "status": "pass"
+    },
+    {
+      "detail": "default_profile points to a listed profile",
+      "metrics": {
+        "value": "llamacpp_q4_k_m_coding_context"
+      },
+      "name": "profiles:default_profile",
+      "status": "pass"
+    },
+    {
+      "detail": "lowest_memory_profile points to a listed profile",
+      "metrics": {
+        "value": "llamacpp_q3_k_m_short_context"
+      },
+      "name": "profiles:lowest_memory_profile",
+      "status": "pass"
+    },
+    {
+      "detail": "middle_footprint_profile points to a listed profile",
+      "metrics": {
+        "value": "llamacpp_iq4_xs_short_context"
+      },
+      "name": "profiles:middle_footprint_profile",
+      "status": "pass"
+    },
+    {
+      "detail": "adaptive_mtp_profile points to a listed profile",
+      "metrics": {
+        "value": "llamacpp_iq4_xs_mtp_graft_adaptive"
+      },
+      "name": "profiles:adaptive_mtp_profile",
+      "status": "pass"
+    },
+    {
+      "detail": "profile uses tensor_parallel_size=1",
+      "metrics": {
+        "tensor_parallel_size": 1
+      },
+      "name": "profile:llamacpp_q4_k_m_short_context:tp1",
+      "status": "pass"
+    },
+    {
+      "detail": "serving script exists: scripts/serve_llamacpp_gpu0.sh",
+      "evidence": [
+        "hf_upload/scripts/serve_llamacpp_gpu0.sh"
+      ],
+      "name": "profile:llamacpp_q4_k_m_short_context:script",
+      "status": "pass"
+    },
+    {
+      "detail": "profile sets REASONING=off",
+      "metrics": {
+        "REASONING": "off"
+      },
+      "name": "profile:llamacpp_q4_k_m_short_context:reasoning_off",
+      "status": "pass"
+    },
+    {
+      "detail": "artifact is published: ornith-1.0-35b-Q4_K_M.gguf",
+      "metrics": {
+        "artifact": "ornith-1.0-35b-Q4_K_M.gguf"
+      },
+      "name": "profile:llamacpp_q4_k_m_short_context:artifact_published",
+      "status": "pass"
+    },
+    {
+      "detail": "evidence exists: benchmarks/llamacpp-quant-benchmarks.md",
+      "evidence": [
+        "hf_upload/benchmarks/llamacpp-quant-benchmarks.md"
+      ],
+      "name": "profile:llamacpp_q4_k_m_short_context:evidence:benchmark",
+      "status": "pass"
+    },
+    {
+      "detail": "evidence exists: probes/behavior-llamacpp-q4-script-default.json",
+      "evidence": [
+        "hf_upload/probes/behavior-llamacpp-q4-script-default.json"
+      ],
+      "name": "profile:llamacpp_q4_k_m_short_context:evidence:behavior",
+      "status": "pass"
+    },
+    {
+      "detail": "profile uses tensor_parallel_size=1",
+      "metrics": {
+        "tensor_parallel_size": 1
+      },
+      "name": "profile:llamacpp_q4_k_m_coding_context:tp1",
+      "status": "pass"
+    },
+    {
+      "detail": "serving script exists: scripts/serve_llamacpp_gpu0.sh",
+      "evidence": [
+        "hf_upload/scripts/serve_llamacpp_gpu0.sh"
+      ],
+      "name": "profile:llamacpp_q4_k_m_coding_context:script",
+      "status": "pass"
+    },
+    {
+      "detail": "profile sets REASONING=off",
+      "metrics": {
+        "REASONING": "off"
+      },
+      "name": "profile:llamacpp_q4_k_m_coding_context:reasoning_off",
+      "status": "pass"
+    },
+    {
+      "detail": "artifact is published: ornith-1.0-35b-Q4_K_M.gguf",
+      "metrics": {
+        "artifact": "ornith-1.0-35b-Q4_K_M.gguf"
+      },
+      "name": "profile:llamacpp_q4_k_m_coding_context:artifact_published",
+      "status": "pass"
+    },
+    {
+      "detail": "evidence exists: benchmarks/llamacpp-quant-benchmarks.md",
+      "evidence": [
+        "hf_upload/benchmarks/llamacpp-quant-benchmarks.md"
+      ],
+      "name": "profile:llamacpp_q4_k_m_coding_context:evidence:benchmark",
+      "status": "pass"
+    },
+    {
+      "detail": "evidence exists: probes/behavior-llamacpp-q4-script-default.json",
+      "evidence": [
+        "hf_upload/probes/behavior-llamacpp-q4-script-default.json"
+      ],
+      "name": "profile:llamacpp_q4_k_m_coding_context:evidence:behavior",
+      "status": "pass"
+    },
+    {
+      "detail": "profile uses tensor_parallel_size=1",
+      "metrics": {
+        "tensor_parallel_size": 1
+      },
+      "name": "profile:llamacpp_q3_k_m_short_context:tp1",
+      "status": "pass"
+    },
+    {
+      "detail": "serving script exists: scripts/serve_llamacpp_gpu0.sh",
+      "evidence": [
+        "hf_upload/scripts/serve_llamacpp_gpu0.sh"
+      ],
+      "name": "profile:llamacpp_q3_k_m_short_context:script",
+      "status": "pass"
+    },
+    {
+      "detail": "profile sets REASONING=off",
+      "metrics": {
+        "REASONING": "off"
+      },
+      "name": "profile:llamacpp_q3_k_m_short_context:reasoning_off",
+      "status": "pass"
+    },
+    {
+      "detail": "artifact is published: ornith-1.0-35b-Q3_K_M.gguf",
+      "metrics": {
+        "artifact": "ornith-1.0-35b-Q3_K_M.gguf"
+      },
+      "name": "profile:llamacpp_q3_k_m_short_context:artifact_published",
+      "status": "pass"
+    },
+    {
+      "detail": "evidence exists: benchmarks/llamacpp-quant-benchmarks.md",
+      "evidence": [
+        "hf_upload/benchmarks/llamacpp-quant-benchmarks.md"
+      ],
+      "name": "profile:llamacpp_q3_k_m_short_context:evidence:benchmark",
+      "status": "pass"
+    },
+    {
+      "detail": "evidence exists: benchmarks/kld-quant-vs-bf16-top64.md",
+      "evidence": [
+        "hf_upload/benchmarks/kld-quant-vs-bf16-top64.md"
+      ],
+      "name": "profile:llamacpp_q3_k_m_short_context:evidence:kld",
+      "status": "pass"
+    },
+    {
+      "detail": "evidence exists: probes/behavior-llamacpp-q3-k-m-parallel16.json",
+      "evidence": [
+        "hf_upload/probes/behavior-llamacpp-q3-k-m-parallel16.json"
+      ],
+      "name": "profile:llamacpp_q3_k_m_short_context:evidence:behavior",
+      "status": "pass"
+    },
+    {
+      "detail": "profile uses tensor_parallel_size=1",
+      "metrics": {
+        "tensor_parallel_size": 1
+      },
+      "name": "profile:llamacpp_iq4_xs_short_context:tp1",
+      "status": "pass"
+    },
+    {
+      "detail": "serving script exists: scripts/serve_llamacpp_gpu0.sh",
+      "evidence": [
+        "hf_upload/scripts/serve_llamacpp_gpu0.sh"
+      ],
+      "name": "profile:llamacpp_iq4_xs_short_context:script",
+      "status": "pass"
+    },
+    {
+      "detail": "profile sets REASONING=off",
+      "metrics": {
+        "REASONING": "off"
+      },
+      "name": "profile:llamacpp_iq4_xs_short_context:reasoning_off",
+      "status": "pass"
+    },
+    {
+      "detail": "artifact is published: ornith-1.0-35b-IQ4_XS.gguf",
+      "metrics": {
+        "artifact": "ornith-1.0-35b-IQ4_XS.gguf"
+      },
+      "name": "profile:llamacpp_iq4_xs_short_context:artifact_published",
+      "status": "pass"
+    },
+    {
+      "detail": "evidence exists: benchmarks/llamacpp-quant-benchmarks.md",
+      "evidence": [
+        "hf_upload/benchmarks/llamacpp-quant-benchmarks.md"
+      ],
+      "name": "profile:llamacpp_iq4_xs_short_context:evidence:benchmark",
+      "status": "pass"
+    },
+    {
+      "detail": "evidence exists: benchmarks/kld-quant-vs-bf16-top64.md",
+      "evidence": [
+        "hf_upload/benchmarks/kld-quant-vs-bf16-top64.md"
+      ],
+      "name": "profile:llamacpp_iq4_xs_short_context:evidence:kld",
+      "status": "pass"
+    },
+    {
+      "detail": "evidence exists: probes/behavior-llamacpp-iq4-xs-parallel16.json",
+      "evidence": [
+        "hf_upload/probes/behavior-llamacpp-iq4-xs-parallel16.json"
+      ],
+      "name": "profile:llamacpp_iq4_xs_short_context:evidence:behavior",
+      "status": "pass"
+    },
+    {
+      "detail": "profile uses tensor_parallel_size=1",
+      "metrics": {
+        "tensor_parallel_size": 1
+      },
+      "name": "profile:llamacpp_iq4_xs_mtp_graft_adaptive:tp1",
+      "status": "pass"
+    },
+    {
+      "detail": "serving script exists: scripts/serve_llamacpp_gpu0.sh",
+      "evidence": [
+        "hf_upload/scripts/serve_llamacpp_gpu0.sh"
+      ],
+      "name": "profile:llamacpp_iq4_xs_mtp_graft_adaptive:script",
+      "status": "pass"
+    },
+    {
+      "detail": "profile sets REASONING=off",
+      "metrics": {
+        "REASONING": "off"
+      },
+      "name": "profile:llamacpp_iq4_xs_mtp_graft_adaptive:reasoning_off",
+      "status": "pass"
+    },
+    {
+      "detail": "artifact is published: ornith-1.0-35b-IQ4_XS-MTP-graft-headQ6.gguf",
+      "metrics": {
+        "artifact": "ornith-1.0-35b-IQ4_XS-MTP-graft-headQ6.gguf"
+      },
+      "name": "profile:llamacpp_iq4_xs_mtp_graft_adaptive:artifact_published",
+      "status": "pass"
+    },
+    {
+      "detail": "evidence exists: benchmarks/llamacpp-iq4-xs-mtp-graft-adaptive-profile.md",
+      "evidence": [
+        "hf_upload/benchmarks/llamacpp-iq4-xs-mtp-graft-adaptive-profile.md"
+      ],
+      "name": "profile:llamacpp_iq4_xs_mtp_graft_adaptive:evidence:benchmark",
+      "status": "pass"
+    },
+    {
+      "detail": "evidence exists: patches/llamacpp-mtp-adaptive-serving-and-recurrent-rollback.patch",
+      "evidence": [
+        "hf_upload/patches/llamacpp-mtp-adaptive-serving-and-recurrent-rollback.patch"
+      ],
+      "name": "profile:llamacpp_iq4_xs_mtp_graft_adaptive:evidence:patch",
+      "status": "pass"
+    },
+    {
+      "detail": "adaptive MTP profile uses draft-mtp with max drafting slots 1 and n_max 2",
+      "metrics": {
+        "LLAMA_SPEC_MAX_DRAFTING_SLOTS": "1",
+        "SPEC_DRAFT_N_MAX": "2",
+        "SPEC_TYPE": "draft-mtp"
+      },
+      "name": "profile:llamacpp_iq4_xs_mtp_graft_adaptive:adaptive_mtp_throttle",
+      "status": "pass"
+    },
+    {
+      "detail": "MTP saturated-c16 caveat is documented",
+      "metrics": {
+        "caveats": [
+          "Always-on MTP is not recommended at saturated c16 because it was slower than target-only decoding.",
+          "Adaptive throttling uses MTP for low-concurrency requests and disables it under saturated batches."
+        ]
+      },
+      "name": "profile:llamacpp_iq4_xs_mtp_graft_adaptive:mtp_caveat",
+      "status": "pass"
+    }
+  ],
+  "generated_at": "2026-06-28T00:56:49.528851+00:00",
+  "inputs": {
+    "catalog": "configs/serving_profiles.yaml",
+    "hf_root": "hf_upload",
+    "manifest": "hf_upload/ARTIFACT_MANIFEST.json"
+  },
+  "profiles": [
+    {
+      "artifact": "ornith-1.0-35b-Q4_K_M.gguf",
+      "backend": "llama.cpp",
+      "command": {
+        "env": {
+          "BATCH_SIZE": "4096",
+          "CTX_SIZE": "8192",
+          "PARALLEL": "16",
+          "PORT": "8002",
+          "QUANT": "Q4_K_M",
+          "REASONING": "off",
+          "UBATCH_SIZE": "512"
+        },
+        "script": "scripts/serve_llamacpp_gpu0.sh"
+      },
+      "evidence": {
+        "behavior": "probes/behavior-llamacpp-q4-script-default.json",
+        "benchmark": "benchmarks/llamacpp-quant-benchmarks.md"
+      },
+      "id": "llamacpp_q4_k_m_short_context",
+      "measured": {
+        "behavior_gate": "pass",
+        "c16_output_tps": 655.6,
+        "c16_p95_ttft_ms": 650.0,
+        "c1_output_tps": 243.3,
+        "loaded_vram_gib": 21.31
+      },
+      "notes": [
+        "Short-context benchmark profile exposes about 512 tokens per slot because CTX_SIZE is split across PARALLEL slots."
+      ],
+      "quant": "Q4_K_M",
+      "status": "recommended_speed_benchmark",
+      "tensor_parallel_size": 1
+    },
+    {
+      "artifact": "ornith-1.0-35b-Q4_K_M.gguf",
+      "backend": "llama.cpp",
+      "command": {
+        "env": {
+          "BATCH_SIZE": "4096",
+          "CTX_SIZE": "131072",
+          "PARALLEL": "16",
+          "PORT": "8002",
+          "QUANT": "Q4_K_M",
+          "REASONING": "off",
+          "UBATCH_SIZE": "512"
+        },
+        "script": "scripts/serve_llamacpp_gpu0.sh"
+      },
+      "evidence": {
+        "behavior": "probes/behavior-llamacpp-q4-script-default.json",
+        "benchmark": "benchmarks/llamacpp-quant-benchmarks.md"
+      },
+      "id": "llamacpp_q4_k_m_coding_context",
+      "measured": {
+        "behavior_gate": "pass",
+        "c16_output_tps_short_context": 655.6,
+        "loaded_vram_gib_short_context": 21.31
+      },
+      "notes": [
+        "Practical coding-serving profile with 16 slots and about 8k context per slot.",
+        "Throughput numbers come from the matched short-context sweep."
+      ],
+      "quant": "Q4_K_M",
+      "status": "recommended_default",
+      "tensor_parallel_size": 1
+    },
+    {
+      "artifact": "ornith-1.0-35b-Q3_K_M.gguf",
+      "backend": "llama.cpp",
+      "command": {
+        "env": {
+          "BATCH_SIZE": "4096",
+          "CTX_SIZE": "8192",
+          "PARALLEL": "16",
+          "PORT": "8002",
+          "QUANT": "Q3_K_M",
+          "REASONING": "off",
+          "UBATCH_SIZE": "512"
+        },
+        "script": "scripts/serve_llamacpp_gpu0.sh"
+      },
+      "evidence": {
+        "behavior": "probes/behavior-llamacpp-q3-k-m-parallel16.json",
+        "benchmark": "benchmarks/llamacpp-quant-benchmarks.md",
+        "kld": "benchmarks/kld-quant-vs-bf16-top64.md"
+      },
+      "id": "llamacpp_q3_k_m_short_context",
+      "measured": {
+        "behavior_gate": "pass",
+        "c16_output_tps": 493.0,
+        "c1_output_tps": 240.5,
+        "gguf_gib": 15.61,
+        "loaded_vram_gib": 17.27,
+        "mean_kld_nats_vs_bf16_top64": 0.362
+      },
+      "quant": "Q3_K_M",
+      "status": "lowest_memory_validated",
+      "tensor_parallel_size": 1
+    },
+    {
+      "artifact": "ornith-1.0-35b-IQ4_XS.gguf",
+      "backend": "llama.cpp",
+      "command": {
+        "env": {
+          "BATCH_SIZE": "4096",
+          "CTX_SIZE": "8192",
+          "PARALLEL": "16",
+          "PORT": "8002",
+          "QUANT": "IQ4_XS",
+          "REASONING": "off",
+          "UBATCH_SIZE": "512"
+        },
+        "script": "scripts/serve_llamacpp_gpu0.sh"
+      },
+      "evidence": {
+        "behavior": "probes/behavior-llamacpp-iq4-xs-parallel16.json",
+        "benchmark": "benchmarks/llamacpp-quant-benchmarks.md",
+        "kld": "benchmarks/kld-quant-vs-bf16-top64.md"
+      },
+      "id": "llamacpp_iq4_xs_short_context",
+      "measured": {
+        "behavior_gate": "pass",
+        "c16_output_tps": 476.0,
+        "c1_output_tps": 234.1,
+        "gguf_gib": 17.64,
+        "loaded_vram_gib": 19.34,
+        "mean_kld_nats_vs_bf16_top64": 0.1426
+      },
+      "quant": "IQ4_XS",
+      "status": "middle_footprint_validated",
+      "tensor_parallel_size": 1
+    },
+    {
+      "artifact": "ornith-1.0-35b-IQ4_XS-MTP-graft-headQ6.gguf",
+      "backend": "llama.cpp",
+      "caveats": [
+        "Always-on MTP is not recommended at saturated c16 because it was slower than target-only decoding.",
+        "Adaptive throttling uses MTP for low-concurrency requests and disables it under saturated batches."
+      ],
+      "command": {
+        "env": {
+          "ALIAS": "Ornith-1.0-35B-GGUF-IQ4_XS-MTP",
+          "BATCH_SIZE": "4096",
+          "CTX_SIZE": "8192",
+          "LLAMA_MTP_DRAFT_TEMP": "1",
+          "LLAMA_MTP_DRAFT_TOP_K": "1",
+          "LLAMA_MTP_DRAFT_TOP_P": "1",
+          "LLAMA_MTP_FAST_BACKEND_SAMPLE": "1",
+          "LLAMA_SPEC_MAX_DRAFTING_SLOTS": "1",
+          "MODEL_PATH": "ornith-1.0-35b-IQ4_XS-MTP-graft-headQ6.gguf",
+          "PARALLEL": "16",
+          "PORT": "8002",
+          "REASONING": "off",
+          "SPEC_DRAFT_BACKEND_SAMPLING": "1",
+          "SPEC_DRAFT_N_MAX": "2",
+          "SPEC_DRAFT_N_MIN": "0",
+          "SPEC_TYPE": "draft-mtp",
+          "UBATCH_SIZE": "512"
+        },
+        "script": "scripts/serve_llamacpp_gpu0.sh"
+      },
+      "evidence": {
+        "benchmark": "benchmarks/llamacpp-iq4-xs-mtp-graft-adaptive-profile.md",
+        "patch": "patches/llamacpp-mtp-adaptive-serving-and-recurrent-rollback.patch"
+      },
+      "id": "llamacpp_iq4_xs_mtp_graft_adaptive",
+      "measured": {
+        "c16_adaptive_tps": 564.48,
+        "c16_ar_tps": 568.57,
+        "c16_threshold1_tps": 592.45,
+        "c1_acceptance_by_position": [
+          0.953,
+          0.865
+        ],
+        "c1_ar_tps": 221.8,
+        "c1_mtp_adaptive_tps": 319.53
+      },
+      "quant": "IQ4_XS_with_Q6_MTP_head",
+      "status": "experimental_adaptive_mtp",
+      "tensor_parallel_size": 1
+    }
+  ],
+  "summary": {
+    "checks": 43,
+    "failed_checks": 0,
+    "status": "pass",
+    "status_counts": {
+      "pass": 43
+    },
+    "warnings": 0
+  }
+}

probes/serving-profile-validation.md

@@ -0,0 +1,68 @@
+# Serving Profile Validation
+
+Generated: 2026-06-28T00:56:49.528851+00:00
+
+## Summary
+
+- Status: `pass`
+- Checks: 43
+- Failures: 0
+- Warnings: 0
+
+## Profiles
+
+| Profile | Status | Backend | Artifact |
+|---|---|---|---|
+| `llamacpp_q4_k_m_short_context` | `recommended_speed_benchmark` | llama.cpp | `ornith-1.0-35b-Q4_K_M.gguf` |
+| `llamacpp_q4_k_m_coding_context` | `recommended_default` | llama.cpp | `ornith-1.0-35b-Q4_K_M.gguf` |
+| `llamacpp_q3_k_m_short_context` | `lowest_memory_validated` | llama.cpp | `ornith-1.0-35b-Q3_K_M.gguf` |
+| `llamacpp_iq4_xs_short_context` | `middle_footprint_validated` | llama.cpp | `ornith-1.0-35b-IQ4_XS.gguf` |
+| `llamacpp_iq4_xs_mtp_graft_adaptive` | `experimental_adaptive_mtp` | llama.cpp | `ornith-1.0-35b-IQ4_XS-MTP-graft-headQ6.gguf` |
+
+## Checks
+
+| Check | Status | Detail |
+|---|---:|---|
+| `catalog:exists` | `pass` | catalog exists: configs/serving_profiles.yaml |
+| `policy:tp1_only` | `pass` | serving policy is single-GPU tp=1 |
+| `policy:multi_gpu_forbidden` | `pass` | multi-GPU serving and tensor_parallel_size>1 are explicitly forbidden |
+| `policy:reasoning_off` | `pass` | default serving policy disables reasoning final-answer leakage |
+| `profiles:count` | `pass` | 5 serving profiles are listed |
+| `profiles:default_profile` | `pass` | default_profile points to a listed profile |
+| `profiles:lowest_memory_profile` | `pass` | lowest_memory_profile points to a listed profile |
+| `profiles:middle_footprint_profile` | `pass` | middle_footprint_profile points to a listed profile |
+| `profiles:adaptive_mtp_profile` | `pass` | adaptive_mtp_profile points to a listed profile |
+| `profile:llamacpp_q4_k_m_short_context:tp1` | `pass` | profile uses tensor_parallel_size=1 |
+| `profile:llamacpp_q4_k_m_short_context:script` | `pass` | serving script exists: scripts/serve_llamacpp_gpu0.sh |
+| `profile:llamacpp_q4_k_m_short_context:reasoning_off` | `pass` | profile sets REASONING=off |
+| `profile:llamacpp_q4_k_m_short_context:artifact_published` | `pass` | artifact is published: ornith-1.0-35b-Q4_K_M.gguf |
+| `profile:llamacpp_q4_k_m_short_context:evidence:benchmark` | `pass` | evidence exists: benchmarks/llamacpp-quant-benchmarks.md |
+| `profile:llamacpp_q4_k_m_short_context:evidence:behavior` | `pass` | evidence exists: probes/behavior-llamacpp-q4-script-default.json |
+| `profile:llamacpp_q4_k_m_coding_context:tp1` | `pass` | profile uses tensor_parallel_size=1 |
+| `profile:llamacpp_q4_k_m_coding_context:script` | `pass` | serving script exists: scripts/serve_llamacpp_gpu0.sh |
+| `profile:llamacpp_q4_k_m_coding_context:reasoning_off` | `pass` | profile sets REASONING=off |
+| `profile:llamacpp_q4_k_m_coding_context:artifact_published` | `pass` | artifact is published: ornith-1.0-35b-Q4_K_M.gguf |
+| `profile:llamacpp_q4_k_m_coding_context:evidence:benchmark` | `pass` | evidence exists: benchmarks/llamacpp-quant-benchmarks.md |
+| `profile:llamacpp_q4_k_m_coding_context:evidence:behavior` | `pass` | evidence exists: probes/behavior-llamacpp-q4-script-default.json |
+| `profile:llamacpp_q3_k_m_short_context:tp1` | `pass` | profile uses tensor_parallel_size=1 |
+| `profile:llamacpp_q3_k_m_short_context:script` | `pass` | serving script exists: scripts/serve_llamacpp_gpu0.sh |
+| `profile:llamacpp_q3_k_m_short_context:reasoning_off` | `pass` | profile sets REASONING=off |
+| `profile:llamacpp_q3_k_m_short_context:artifact_published` | `pass` | artifact is published: ornith-1.0-35b-Q3_K_M.gguf |
+| `profile:llamacpp_q3_k_m_short_context:evidence:benchmark` | `pass` | evidence exists: benchmarks/llamacpp-quant-benchmarks.md |
+| `profile:llamacpp_q3_k_m_short_context:evidence:kld` | `pass` | evidence exists: benchmarks/kld-quant-vs-bf16-top64.md |
+| `profile:llamacpp_q3_k_m_short_context:evidence:behavior` | `pass` | evidence exists: probes/behavior-llamacpp-q3-k-m-parallel16.json |
+| `profile:llamacpp_iq4_xs_short_context:tp1` | `pass` | profile uses tensor_parallel_size=1 |
+| `profile:llamacpp_iq4_xs_short_context:script` | `pass` | serving script exists: scripts/serve_llamacpp_gpu0.sh |
+| `profile:llamacpp_iq4_xs_short_context:reasoning_off` | `pass` | profile sets REASONING=off |
+| `profile:llamacpp_iq4_xs_short_context:artifact_published` | `pass` | artifact is published: ornith-1.0-35b-IQ4_XS.gguf |
+| `profile:llamacpp_iq4_xs_short_context:evidence:benchmark` | `pass` | evidence exists: benchmarks/llamacpp-quant-benchmarks.md |
+| `profile:llamacpp_iq4_xs_short_context:evidence:kld` | `pass` | evidence exists: benchmarks/kld-quant-vs-bf16-top64.md |
+| `profile:llamacpp_iq4_xs_short_context:evidence:behavior` | `pass` | evidence exists: probes/behavior-llamacpp-iq4-xs-parallel16.json |
+| `profile:llamacpp_iq4_xs_mtp_graft_adaptive:tp1` | `pass` | profile uses tensor_parallel_size=1 |
+| `profile:llamacpp_iq4_xs_mtp_graft_adaptive:script` | `pass` | serving script exists: scripts/serve_llamacpp_gpu0.sh |
+| `profile:llamacpp_iq4_xs_mtp_graft_adaptive:reasoning_off` | `pass` | profile sets REASONING=off |
+| `profile:llamacpp_iq4_xs_mtp_graft_adaptive:artifact_published` | `pass` | artifact is published: ornith-1.0-35b-IQ4_XS-MTP-graft-headQ6.gguf |
+| `profile:llamacpp_iq4_xs_mtp_graft_adaptive:evidence:benchmark` | `pass` | evidence exists: benchmarks/llamacpp-iq4-xs-mtp-graft-adaptive-profile.md |
+| `profile:llamacpp_iq4_xs_mtp_graft_adaptive:evidence:patch` | `pass` | evidence exists: patches/llamacpp-mtp-adaptive-serving-and-recurrent-rollback.patch |
+| `profile:llamacpp_iq4_xs_mtp_graft_adaptive:adaptive_mtp_throttle` | `pass` | adaptive MTP profile uses draft-mtp with max drafting slots 1 and n_max 2 |
+| `profile:llamacpp_iq4_xs_mtp_graft_adaptive:mtp_caveat` | `pass` | MTP saturated-c16 caveat is documented |

scripts/build_release_readiness_report.py

@@ -191,6 +191,33 @@ def check_behavior_matrix(path: Path, min_passing_artifacts: int) -> list[Check]
     return checks
 
 
+def check_serving_profile_validation(path: Path) -> list[Check]:
+    checks = [exists_check("serving_profiles:validation_exists", path, "serving profile validation report")]
+    if not path.exists():
+        return checks
+    report = load_json(path)
+    summary = report.get("summary", {})
+    checks.append(
+        Check(
+            name="serving_profiles:validation_status",
+            status=PASS if summary.get("status") == "pass" and summary.get("failed_checks") == 0 else FAIL,
+            detail=f"serving profile validation status is {summary.get('status')!r}",
+            evidence=[str(path)],
+            metrics=summary,
+        )
+    )
+    checks.append(
+        Check(
+            name="serving_profiles:profile_count",
+            status=PASS if len(report.get("profiles") or []) >= 5 else FAIL,
+            detail=f"{len(report.get('profiles') or [])} serving profiles are validated",
+            evidence=[str(path)],
+            metrics={"profiles": [profile.get("id") for profile in report.get("profiles") or []]},
+        )
+    )
+    return checks
+
+
 def check_sft_state(manifest_path: Path, allow_unreleased_sft: bool) -> list[Check]:
     if not manifest_path.exists():
         return [
@@ -253,6 +280,7 @@ def build_report(args: argparse.Namespace) -> dict[str, Any]:
     checks.extend(check_hf_verifier(args.hf_verification))
     checks.extend(check_manifest(args.manifest, args.min_gguf))
     checks.extend(check_behavior_matrix(args.behavior_matrix, args.min_passing_artifacts))
+    checks.extend(check_serving_profile_validation(args.serving_profile_validation))
     checks.extend(check_sft_state(args.manifest, args.allow_unreleased_sft))
     checks.extend(check_required_docs(args.required_doc or []))
 
@@ -297,6 +325,7 @@ def build_report(args: argparse.Namespace) -> dict[str, Any]:
             "manifest": str(args.manifest),
             "hf_verification": str(args.hf_verification),
             "behavior_matrix": str(args.behavior_matrix),
+            "serving_profile_validation": str(args.serving_profile_validation),
             "quant_benchmarks": str(args.quant_benchmarks),
             "kld_report": str(args.kld_report),
             "mtp_profile": str(args.mtp_profile),
@@ -372,6 +401,7 @@ def parse_args() -> argparse.Namespace:
     parser.add_argument("--manifest", type=Path, default=Path("hf_upload/ARTIFACT_MANIFEST.json"))
     parser.add_argument("--hf-verification", type=Path, default=Path("runs/hf-release-verification-behavior-matrix.json"))
     parser.add_argument("--behavior-matrix", type=Path, default=Path("runs/behavior-regression-matrix.json"))
+    parser.add_argument("--serving-profile-validation", type=Path, default=Path("runs/serving-profile-validation.json"))
     parser.add_argument("--quant-benchmarks", type=Path, default=Path("hf_upload/benchmarks/llamacpp-quant-benchmarks.md"))
     parser.add_argument("--kld-report", type=Path, default=Path("hf_upload/benchmarks/kld-quant-vs-bf16-top64.md"))
     parser.add_argument("--mtp-profile", type=Path, default=Path("hf_upload/benchmarks/llamacpp-iq4-xs-mtp-graft-adaptive-profile.md"))

scripts/validate_serving_profiles.py

@@ -0,0 +1,278 @@
+#!/usr/bin/env python3
+"""Validate the machine-readable serving profile catalog."""
+
+from __future__ import annotations
+
+import argparse
+import json
+from collections import Counter
+from datetime import UTC, datetime
+from pathlib import Path
+from typing import Any
+
+import yaml
+
+
+PASS = "pass"
+WARN = "warn"
+FAIL = "fail"
+
+
+def load_yaml(path: Path) -> dict[str, Any]:
+    with path.open("r", encoding="utf-8") as handle:
+        return yaml.safe_load(handle) or {}
+
+
+def load_json(path: Path) -> dict[str, Any]:
+    with path.open("r", encoding="utf-8") as handle:
+        return json.load(handle)
+
+
+def add_check(
+    checks: list[dict[str, Any]],
+    name: str,
+    status: str,
+    detail: str,
+    **extra: Any,
+) -> None:
+    item = {"name": name, "status": status, "detail": detail}
+    item.update(extra)
+    checks.append(item)
+
+
+def rel_exists(root: Path, rel: str) -> bool:
+    return (root / rel).exists()
+
+
+def validate_policy(checks: list[dict[str, Any]], catalog: dict[str, Any]) -> None:
+    policy = catalog.get("policy") or {}
+    add_check(
+        checks,
+        "policy:tp1_only",
+        PASS
+        if policy.get("serving_parallelism") == "tp1_only"
+        and int(policy.get("tensor_parallel_size") or 0) == 1
+        else FAIL,
+        "serving policy is single-GPU tp=1",
+        metrics=policy,
+    )
+    forbidden = set(policy.get("forbidden") or [])
+    add_check(
+        checks,
+        "policy:multi_gpu_forbidden",
+        PASS if {"tensor_parallel_size_gt_1", "multi_gpu_serving"} <= forbidden else FAIL,
+        "multi-GPU serving and tensor_parallel_size>1 are explicitly forbidden",
+        metrics={"forbidden": sorted(forbidden)},
+    )
+    add_check(
+        checks,
+        "policy:reasoning_off",
+        PASS if str(policy.get("reasoning")) == "off" else FAIL,
+        "default serving policy disables reasoning final-answer leakage",
+        metrics={"reasoning": policy.get("reasoning")},
+    )
+
+
+def validate_profiles(
+    checks: list[dict[str, Any]],
+    *,
+    catalog: dict[str, Any],
+    manifest: dict[str, Any],
+    hf_root: Path,
+) -> None:
+    profiles = catalog.get("profiles") or []
+    profile_ids = [profile.get("id") for profile in profiles]
+    manifest_artifacts = {
+        item.get("path")
+        for item in manifest.get("published_gguf", [])
+        if item.get("published") is True
+    }
+
+    add_check(
+        checks,
+        "profiles:count",
+        PASS if len(profiles) >= 5 else FAIL,
+        f"{len(profiles)} serving profiles are listed",
+        metrics={"profiles": profile_ids},
+    )
+    for field in ("default_profile", "lowest_memory_profile", "middle_footprint_profile", "adaptive_mtp_profile"):
+        value = catalog.get(field)
+        add_check(
+            checks,
+            f"profiles:{field}",
+            PASS if value in profile_ids else FAIL,
+            f"{field} points to a listed profile" if value in profile_ids else f"{field}={value!r} is not listed",
+            metrics={"value": value},
+        )
+
+    for profile in profiles:
+        profile_id = str(profile.get("id") or "unknown")
+        tp = int(profile.get("tensor_parallel_size") or 0)
+        add_check(
+            checks,
+            f"profile:{profile_id}:tp1",
+            PASS if tp == 1 else FAIL,
+            "profile uses tensor_parallel_size=1",
+            metrics={"tensor_parallel_size": tp},
+        )
+
+        command = profile.get("command") or {}
+        script = command.get("script")
+        add_check(
+            checks,
+            f"profile:{profile_id}:script",
+            PASS if script and rel_exists(hf_root, script) else FAIL,
+            f"serving script exists: {script}",
+            evidence=[str(hf_root / str(script))] if script else [],
+        )
+
+        env = command.get("env") or {}
+        reasoning = env.get("REASONING")
+        add_check(
+            checks,
+            f"profile:{profile_id}:reasoning_off",
+            PASS if reasoning == "off" else FAIL,
+            "profile sets REASONING=off",
+            metrics={"REASONING": reasoning},
+        )
+
+        artifact = profile.get("artifact")
+        add_check(
+            checks,
+            f"profile:{profile_id}:artifact_published",
+            PASS if artifact in manifest_artifacts else FAIL,
+            f"artifact is published: {artifact}",
+            metrics={"artifact": artifact},
+        )
+
+        evidence = profile.get("evidence") or {}
+        for key, rel in evidence.items():
+            if not isinstance(rel, str):
+                continue
+            add_check(
+                checks,
+                f"profile:{profile_id}:evidence:{key}",
+                PASS if rel_exists(hf_root, rel) else FAIL,
+                f"evidence exists: {rel}",
+                evidence=[str(hf_root / rel)],
+            )
+
+        if "mtp" in profile_id:
+            add_check(
+                checks,
+                f"profile:{profile_id}:adaptive_mtp_throttle",
+                PASS
+                if env.get("LLAMA_SPEC_MAX_DRAFTING_SLOTS") == "1"
+                and env.get("SPEC_TYPE") == "draft-mtp"
+                and env.get("SPEC_DRAFT_N_MAX") == "2"
+                else FAIL,
+                "adaptive MTP profile uses draft-mtp with max drafting slots 1 and n_max 2",
+                metrics={
+                    "LLAMA_SPEC_MAX_DRAFTING_SLOTS": env.get("LLAMA_SPEC_MAX_DRAFTING_SLOTS"),
+                    "SPEC_TYPE": env.get("SPEC_TYPE"),
+                    "SPEC_DRAFT_N_MAX": env.get("SPEC_DRAFT_N_MAX"),
+                },
+            )
+            caveats = profile.get("caveats") or []
+            add_check(
+                checks,
+                f"profile:{profile_id}:mtp_caveat",
+                PASS if any("saturated c16" in str(item) for item in caveats) else WARN,
+                "MTP saturated-c16 caveat is documented",
+                metrics={"caveats": caveats},
+            )
+
+
+def render_markdown(report: dict[str, Any]) -> str:
+    summary = report["summary"]
+    lines = [
+        "# Serving Profile Validation",
+        "",
+        f"Generated: {report['generated_at']}",
+        "",
+        "## Summary",
+        "",
+        f"- Status: `{summary['status']}`",
+        f"- Checks: {summary['checks']}",
+        f"- Failures: {summary['failed_checks']}",
+        f"- Warnings: {summary['warnings']}",
+        "",
+        "## Profiles",
+        "",
+        "| Profile | Status | Backend | Artifact |",
+        "|---|---|---|---|",
+    ]
+    for profile in report["profiles"]:
+        lines.append(
+            f"| `{profile.get('id')}` | `{profile.get('status')}` | {profile.get('backend')} | `{profile.get('artifact')}` |"
+        )
+    lines.extend(["", "## Checks", "", "| Check | Status | Detail |", "|---|---:|---|"])
+    for check in report["checks"]:
+        lines.append(f"| `{check['name']}` | `{check['status']}` | {check['detail']} |")
+    lines.append("")
+    return "\n".join(lines)
+
+
+def build_report(args: argparse.Namespace) -> dict[str, Any]:
+    catalog = load_yaml(args.catalog)
+    manifest = load_json(args.manifest)
+    checks: list[dict[str, Any]] = []
+
+    add_check(
+        checks,
+        "catalog:exists",
+        PASS if args.catalog.exists() else FAIL,
+        f"catalog exists: {args.catalog}",
+    )
+    validate_policy(checks, catalog)
+    validate_profiles(checks, catalog=catalog, manifest=manifest, hf_root=args.hf_root)
+
+    failures = [check for check in checks if check["status"] == FAIL]
+    warnings = [check for check in checks if check["status"] == WARN]
+    report = {
+        "generated_at": datetime.now(UTC).isoformat(),
+        "inputs": {
+            "catalog": str(args.catalog),
+            "manifest": str(args.manifest),
+            "hf_root": str(args.hf_root),
+        },
+        "summary": {
+            "status": PASS if not failures else FAIL,
+            "checks": len(checks),
+            "failed_checks": len(failures),
+            "warnings": len(warnings),
+            "status_counts": dict(Counter(check["status"] for check in checks)),
+        },
+        "profiles": catalog.get("profiles") or [],
+        "checks": checks,
+    }
+    return report
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--catalog", type=Path, default=Path("configs/serving_profiles.yaml"))
+    parser.add_argument("--manifest", type=Path, default=Path("hf_upload/ARTIFACT_MANIFEST.json"))
+    parser.add_argument("--hf-root", type=Path, default=Path("hf_upload"))
+    parser.add_argument("--output", type=Path, required=True)
+    parser.add_argument("--markdown-output", type=Path)
+    parser.add_argument("--fail-on-error", action="store_true")
+    return parser.parse_args()
+
+
+def main() -> int:
+    args = parse_args()
+    report = build_report(args)
+    args.output.parent.mkdir(parents=True, exist_ok=True)
+    args.output.write_text(json.dumps(report, indent=2, sort_keys=True) + "\n")
+    if args.markdown_output:
+        args.markdown_output.parent.mkdir(parents=True, exist_ok=True)
+        args.markdown_output.write_text(render_markdown(report), encoding="utf-8")
+    print(json.dumps(report["summary"], indent=2, sort_keys=True))
+    if args.fail_on_error and report["summary"]["status"] != PASS:
+        return 1
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

scripts/verify_hf_release.py

@@ -63,6 +63,7 @@ def collect_required_support_files(manifest: dict[str, Any]) -> set[str]:
     release_verification = manifest.get("release_verification", {})
     behavior_matrix = manifest.get("behavior_regression_matrix", {})
     release_readiness = manifest.get("release_readiness", {})
+    serving_profiles = manifest.get("serving_profiles", {})
 
     for key in ("scripts", "reports", "published_synthetic_data"):
         required.update(str(path) for path in repair.get(key, []) if path)
@@ -77,8 +78,8 @@ def collect_required_support_files(manifest: dict[str, Any]) -> set[str]:
         if value:
             required.add(str(value))
 
-    for section in (behavior_matrix, release_readiness):
-        for key in ("script", "report", "markdown_report"):
+    for section in (behavior_matrix, release_readiness, serving_profiles):
+        for key in ("script", "config", "report", "markdown_report"):
             value = section.get(key)
             if value:
                 required.add(str(value))
@@ -350,6 +351,26 @@ def check_release_readiness_summary(checks: list[dict[str, Any]], manifest: dict
     )
 
 
+def check_serving_profiles_summary(checks: list[dict[str, Any]], manifest: dict[str, Any]) -> None:
+    serving = manifest.get("serving_profiles")
+    if not serving:
+        add_check(
+            checks,
+            "serving_profiles:manifest_section",
+            "warn",
+            "serving profiles manifest section is not present",
+        )
+        return
+    summary = serving.get("summary", {})
+    add_check(
+        checks,
+        "serving_profiles:validation_status",
+        "pass" if summary.get("status") == "pass" and summary.get("failed_checks") == 0 else "fail",
+        "serving profile validation passes with zero failed checks",
+        metrics=summary,
+    )
+
+
 def check_local_and_remote_hashes(
     checks: list[dict[str, Any]],
     *,
@@ -523,6 +544,7 @@ def build_report(args: argparse.Namespace) -> dict[str, Any]:
     check_sft_gate_summaries(checks, manifest)
     check_behavior_matrix_summary(checks, manifest)
     check_release_readiness_summary(checks, manifest)
+    check_serving_profiles_summary(checks, manifest)
     hash_stats = check_local_and_remote_hashes(
         checks,
         repo_id=repo_id,