Publish validated Bonsai 2 Q2TP affine model and reproducible runtime patch

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+Ternary-Bonsai-2-27B-Q2TP-Affine.cmf filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,139 @@
+---
+license: apache-2.0
+base_model: prism-ml/Ternary-Bonsai-2-27B-mlx-2bit
+base_model_relation: quantized
+library_name: cortiq
+pipeline_tag: text-generation
+tags:
+  - cmf
+  - cortiq
+  - q2tp-affine
+  - ternary
+  - hadamard
+  - vulkan
+language:
+  - en
+  - ru
+---
+
+# Ternary Bonsai 2 27B — CMF Q2TP affine Prism
+
+This repository contains the CMF conversion of [Prism ML's Ternary Bonsai 2
+27B MLX pack](https://huggingface.co/prism-ml/Ternary-Bonsai-2-27B-mlx-2bit/tree/3f926b415992eaa2ae9dd7b573706494d6bbf787).
+Model file:
+
+`Ternary-Bonsai-2-27B-Q2TP-Affine.cmf`
+
+- Size: **8,247,698,536 bytes**
+- SHA-256: `fb0f1a9cdb0434bc9ce50eb6473bc2ca4a71d6994ad61314bbd7e2f11c4e29e8`
+- Repository: `infosave/Ternary-Bonsai-2-27B-cmf`
+
+The candidate is tested for **text generation only**. Vision tensors and
+metadata are preserved for provenance, but image/text fusion is not implemented
+or tested by this runtime. This is not a VLM release.
+
+## Required runtime
+
+The model uses standard CMF dtype16 Q2TP storage plus the mandatory
+`q2tp_affine` Prism operator and signed normalized Hadamard transforms. The
+operator descriptor is part of the model contract. A stock/public `cortiq`
+0.6.8 binary does not implement this profile; the tested development build
+retains the `cortiq 0.6.8` version label but is built from the pinned public
+source plus the patch in `source/patches/`.
+
+This is not a uniform ordinary two-bit dump: auxiliary, GDN, normalization,
+tokenizer, and vision records retain their declared precision and roles. The
+experimental dtype17/q2tp_t oracle is not part of this release. Older readers
+must reject the required affine/Hadamard features instead of silently decoding
+them as ordinary Q2TP.
+
+## Download and run
+
+The model is available at
+[infosave/Ternary-Bonsai-2-27B-cmf](https://huggingface.co/infosave/Ternary-Bonsai-2-27B-cmf):
+
+```sh
+wget -O Ternary-Bonsai-2-27B-Q2TP-Affine.cmf \
+  https://huggingface.co/infosave/Ternary-Bonsai-2-27B-cmf/resolve/main/Ternary-Bonsai-2-27B-Q2TP-Affine.cmf
+sha256sum Ternary-Bonsai-2-27B-Q2TP-Affine.cmf
+# expected:
+# fb0f1a9cdb0434bc9ce50eb6473bc2ca4a71d6994ad61314bbd7e2f11c4e29e8
+```
+
+Build the matching runtime from the pinned public base and the minimal patch;
+see [`source/SOURCE_BUILD.md`](source/SOURCE_BUILD.md). On Linux, install
+Rust/Cargo build prerequisites plus `libvulkan1` and `vulkan-tools`. The
+validated environment used stable Rust 1.98.1, Vulkan, and an NVIDIA RTX 4090.
+No HF token is needed for the public source or model repository.
+
+```sh
+export XDG_RUNTIME_DIR=/tmp WGPU_BACKEND=vulkan CMF_GPU=wgpu
+export CMF_Q2TP_GPU=1 CMF_GPU_PROBE=0 CMF_MTP=0
+cortiq verify Ternary-Bonsai-2-27B-Q2TP-Affine.cmf
+cortiq run Ternary-Bonsai-2-27B-Q2TP-Affine.cmf \
+  --prompt 'Return only a valid JSON object with exactly two keys: "city" with value "Казань" and "count" with integer value 3. Do not use Markdown.' \
+  --greedy --seed 42 --no-think --max-tokens 256
+```
+
+The validated response was exactly `{"city":"Казань","count":3}`. The
+explicit `CMF_Q2TP_GPU=1` enables the descriptor-aware Vulkan kernels; omit GPU
+variables only when intentionally running a CPU check.
+
+## Provenance and source
+
+- Upstream MLX source: `prism-ml/Ternary-Bonsai-2-27B-mlx-2bit`, revision
+  `3f926b415992eaa2ae9dd7b573706494d6bbf787`.
+- Upstream `model.safetensors`: 8,595,477,990 bytes,
+  SHA-256 `130de5925082c168b7866b2e91b52e44abbafc99017e3ca352b77b5b55a269ed`.
+- Runtime source: [`infosave2007/cmf`](https://github.com/infosave2007/cmf) at
+  `48a9ca5929ea42d28ced54279b4fc9d9c4c33d10`.
+- Minimal patch SHA-256:
+  `e6cd6a60149a8be764a323ab1f968668b524b36ec35bc1a193fcfb5218f669be`.
+- Changed-file hashes and the validated development binary identity are in
+  [`source/SOURCE_PROVENANCE.json`](source/SOURCE_PROVENANCE.json).
+- No runtime binary is redistributed here; the source recipe is the supported
+  reproduction path.
+
+## Quality checks
+
+The frozen mixed English/Russian/code corpus is included under `evaluation/`.
+It has SHA-256
+`181716ce3f0a71b94f7d5d4428232c241b88f3e26d68f4ed50a191506ddb78b5` and was
+tokenized once into 1,122 IDs. Scoring uses no synthetic BOS and two
+independent 512-input resets, scoring targets 1..1024. Raw reference rows and
+the exact recipe are included; this is an engineering correspondence check,
+not a universal benchmark.
+
+| System | Mean NLL | PPL | Result |
+|---|---:|---:|---|
+| Official Prism PQ2 CUDA reference | 1.81274952826 | 6.12727139849 | top-1 603/1024 |
+| CMF affine CPU | 1.812992284 | 6.128759009 | +0.000242756 nats/token |
+| CMF affine Vulkan | **1.813015751** | **6.128902833** | +0.000266223; primary gate passes |
+| Ordinary CMF Q2TP comparison | 1.886951899 | 6.599222896 | +0.074202371; outside +0.05 gate |
+
+The affine Vulkan result was reverified with the clean public-base port. The
+ordinary comparison remains experimental and is not presented as equivalent.
+Deterministic non-thinking text checks also passed arithmetic (`888`), exact
+JSON, fact-context (`5`), bilingual explanations, and a Python function
+shape check. No broad benchmark, long-context, training, or thinking-mode
+quality claim is made.
+
+## RTX 4090 measurement
+
+The accepted runtime4 source measured **8.5 tok/s prefill**, **6.88 s TTFT**,
+and **5.2 tok/s warm decode** on an RTX 4090 24 GB. This 5.2 figure is the
+runtime4 hardware measurement retained for context, **not a clean-port speed
+claim**: the clean public-base port was quality-reverified but not speed-
+rebenchmarked. It is not comparable to the upstream CUDA reference's 90.4
+tok/s figure. The runtime uses the conservative per-operation Prism path and
+CPU recurrence/state handoff rather than claiming full-graph throughput.
+
+## License and attribution
+
+The model and runtime have separate attribution. Preserve the complete bundle
+under [`licenses/`](licenses/): upstream model `LICENSE` and `NOTICE.txt`, the
+upstream MLX runtime MIT notice, and CMF `LICENSE`, `NOTICE`, and `PATENTS.md`.
+The upstream notice also identifies Qwen3.8-27B as Apache-2.0 software and
+links its license. Do not treat the CMF runtime license as replacing the model
+license, and retain all notices when redistributing the model or patched
+runtime.

STAGING_MANIFEST.json

@@ -0,0 +1,123 @@
+{
+  "schema": "ternary-bonsai-2-q2tp-affine-hf-staging-files-1",
+  "intended_repo": "infosave/Ternary-Bonsai-2-27B-cmf",
+  "intended_model_filename": "Ternary-Bonsai-2-27B-Q2TP-Affine.cmf",
+  "model_artifact_included": false,
+  "model_artifact_sha256": "fb0f1a9cdb0434bc9ce50eb6473bc2ca4a71d6994ad61314bbd7e2f11c4e29e8",
+  "model_artifact_bytes": 8247698536,
+  "self_hash": "see STAGING_MANIFEST.sha256 (self and sidecar are excluded from files)",
+  "file_count": 22,
+  "total_bytes": 485136,
+  "files": [
+    {
+      "path": "README.md",
+      "bytes": 6194,
+      "sha256": "ceb132413eca6ba7f70870f4e0ee5c1758635b368e2e77b9b01b46842d6d34eb"
+    },
+    {
+      "path": "evaluation/REFERENCE_SCORER_NOTICE.txt",
+      "bytes": 343,
+      "sha256": "3f74ecab64f3dd828e5a17bea61e36a0d2dc98dc25590c3e1b7f3cda4a44d276"
+    },
+    {
+      "path": "evaluation/SCORE_RECIPE.md",
+      "bytes": 2015,
+      "sha256": "c0ba28b5661f61eb831f7908a69bc48671e14e342b624f8d7f6a9d622e3a1eff"
+    },
+    {
+      "path": "evaluation/cmf-affine-score.txt",
+      "bytes": 395,
+      "sha256": "f56c6b7266dac9e3e534159b6d0eeb74bfbfda42b43a881fb02ecdd76973374f"
+    },
+    {
+      "path": "evaluation/cmf-affine-summary.json",
+      "bytes": 857,
+      "sha256": "c425e00f2d6f930c50c0600b85410904d90a2c0a4cbdd074c1f626db8f4de997"
+    },
+    {
+      "path": "evaluation/official-reference-summary.json",
+      "bytes": 637,
+      "sha256": "ecb493329c56d797ce58f05ad65672a805f6b51327875b30a1908c8d34054333"
+    },
+    {
+      "path": "evaluation/performance-summary.json",
+      "bytes": 357,
+      "sha256": "d508ea57641ca2100b6050a3d5e986180787c40e73ef4ad709217807c718cb41"
+    },
+    {
+      "path": "evaluation/prompts.jsonl",
+      "bytes": 2027,
+      "sha256": "68f6b0b2ffc8526d2f37df7bb2e494531d533e336b6d4df8c1feeebd2652c1f5"
+    },
+    {
+      "path": "evaluation/quality-corpus-token-ids.txt",
+      "bytes": 5774,
+      "sha256": "6dd2b70cb1d4a18a2955737b9dc3711bc20cd951c48a086ce144dcfbff8eb47b"
+    },
+    {
+      "path": "evaluation/quality-corpus.txt",
+      "bytes": 7029,
+      "sha256": "181716ce3f0a71b94f7d5d4428232c241b88f3e26d68f4ed50a191506ddb78b5"
+    },
+    {
+      "path": "evaluation/reference-teacher-score.cpp",
+      "bytes": 6984,
+      "sha256": "ca2a79e2ca72cd4fe5d1f8794d18a98eb2f08860717983e1070044f064a4eacd"
+    },
+    {
+      "path": "evaluation/reference-teacher-score.jsonl",
+      "bytes": 117173,
+      "sha256": "0fe3ce05070664d453186040a0013d1c66431f581252382bd2fe361f69577d7c"
+    },
+    {
+      "path": "licenses/README.md",
+      "bytes": 735,
+      "sha256": "5c42d4e3c3d1e43d18821d7c2669ae0f1a806139da66d594bb0f93bfe157f9e3"
+    },
+    {
+      "path": "licenses/cmf-LICENSE",
+      "bytes": 11345,
+      "sha256": "0bf59d136536dea5dcf80ce96c5988e46522ce69f3ae817666c7f5d10ded6618"
+    },
+    {
+      "path": "licenses/cmf-NOTICE",
+      "bytes": 680,
+      "sha256": "d8ee9168d48f5b0fa0012518848434a65cebb8288f3f2ed51f2738f2f4b347e4"
+    },
+    {
+      "path": "licenses/cmf-PATENTS.md",
+      "bytes": 2955,
+      "sha256": "607faa71dd5a600787f97dcc50e17221e04f57e21e19aa2d5e2a1399e21fb692"
+    },
+    {
+      "path": "licenses/upstream-mlx-runtime-LICENSE",
+      "bytes": 1066,
+      "sha256": "ccfab7ccb2ea306f71531c8ca77bb55507606cd90768b1e32b8b52ab5b48cf01"
+    },
+    {
+      "path": "licenses/upstream-model-LICENSE",
+      "bytes": 10174,
+      "sha256": "69849221bfb90053de2134ef5e6d540287b4b98062326492f1f96f5da685524b"
+    },
+    {
+      "path": "licenses/upstream-model-NOTICE.txt",
+      "bytes": 411,
+      "sha256": "de0e0c48fb6f691a31e74f338e3ccf93f9ecdfe2866ab769c4bf8b79a7636a30"
+    },
+    {
+      "path": "source/SOURCE_BUILD.md",
+      "bytes": 2854,
+      "sha256": "8522ca4b267d0f140a4d09c74ef67a7f130d6051ef25a63a5c96a2b2ea4a1671"
+    },
+    {
+      "path": "source/SOURCE_PROVENANCE.json",
+      "bytes": 8726,
+      "sha256": "648c789c1ba76f745fedea1fa894d5c41137022769e708da05f827795fd3656d"
+    },
+    {
+      "path": "source/patches/cmf-bonsai2-public-base-48a9.patch",
+      "bytes": 296405,
+      "sha256": "e6cd6a60149a8be764a323ab1f968668b524b36ec35bc1a193fcfb5218f669be"
+    }
+  ]
+}

STAGING_MANIFEST.sha256

@@ -0,0 +1 @@
+2915eb4ff8c2aaff6d55ceba0867a20708de502ea0007e6af58b0ebb7494acea  STAGING_MANIFEST.json

Ternary-Bonsai-2-27B-Q2TP-Affine.cmf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fb0f1a9cdb0434bc9ce50eb6473bc2ca4a71d6994ad61314bbd7e2f11c4e29e8
+size 8247698536

evaluation/REFERENCE_SCORER_NOTICE.txt

@@ -0,0 +1,5 @@
+The reference teacher-forcing helper in this directory was authored for this
+run and is released under Apache-2.0. It calls the public llama.cpp C API; it
+does not include or redistribute llama.cpp headers, libraries, CUDA binaries,
+or model weights. Obtain those separately from the pinned official release
+and retain that project's notices.

evaluation/SCORE_RECIPE.md

@@ -0,0 +1,48 @@
+# Frozen evaluation recipe
+
+The fixtures in this directory are run-authored engineering diagnostics,
+released under the CMF Apache-2.0 terms. They are not a general benchmark.
+`prompts.jsonl` contains the eight frozen text prompts. The likelihood corpus
+and its token IDs are separate fixtures; do not retokenize with a different
+model or add a BOS token.
+
+## Official reference
+
+The official PQ2 reference used llama.cpp/Prism release commit
+`9a9394a895b96003ca842a6041cb28ac49a108f7` with CUDA 12.8. The small
+`reference-teacher-score.cpp` helper links against the existing official
+`libllama`/ggml shared libraries and uses `llama_decode` plus
+`llama_get_logits_ith`; it does not build llama.cpp or redistribute its
+headers/libraries. Build the helper against that release's headers and
+libraries, then run:
+
+```sh
+./llama-teacher-score \
+  /path/to/Ternary-Bonsai-2-27B-PQ2_0.gguf \
+  evaluation/quality-corpus-token-ids.txt \
+  evaluation/reference-teacher-score.jsonl
+```
+
+The contract is two independent 512-input-token resets at starts 0 and 512,
+scoring the actual next IDs at targets 1..1024, with no synthetic BOS. The
+shipped raw JSONL records all 1,024 correct-token log-probabilities and top-1
+IDs without a full-vocabulary dump. Reference results: mean NLL
+`1.81274952826`, PPL `6.12727139849`, top-1 `603/1024`.
+
+## CMF affine comparison
+
+Build the patched runtime and its `score_fixed` example using
+`source/SOURCE_BUILD.md`, then use the same token-ID file and the model file:
+
+```sh
+export XDG_RUNTIME_DIR=/tmp WGPU_BACKEND=vulkan CMF_GPU=wgpu
+export CMF_Q2TP_GPU=1 CMF_GPU_PROBE=0 CMF_MTP=0
+./target/debug/examples/score_fixed \
+  Ternary-Bonsai-2-27B-Q2TP-Affine.cmf \
+  evaluation/quality-corpus-token-ids.txt
+```
+
+The validated affine Vulkan output was total NLL `1856.528128685`, mean NLL
+`1.813015751`, PPL `6.128902833`; CPU mean NLL was `1.812992284`. Ordinary
+Q2TP is retained only as a comparison: mean NLL `1.886951899`, delta
+`+0.074202371`, outside the adopted `+0.05` primary gate.

evaluation/cmf-affine-score.txt

@@ -0,0 +1,5 @@
+# Patched clean public-base runtime; frozen corpus; no BOS; reset inputs 0..513 and 512..1025
+# Env: XDG_RUNTIME_DIR=/tmp WGPU_BACKEND=vulkan CMF_GPU=wgpu CMF_Q2TP_GPU=1 CMF_GPU_PROBE=0 CMF_MTP=0
+chunk start=0 input_len=513 nll=1085.211889936 count=512
+chunk start=512 input_len=513 nll=771.316238749 count=512
+fixed-reset-1024 nll=1856.528128685 count=1024 mean_nll=1.813015751 ppl=6.128902833

evaluation/cmf-affine-summary.json

@@ -0,0 +1,17 @@
+{
+  "model_filename": "Ternary-Bonsai-2-27B-Q2TP-Affine.cmf",
+  "model_sha256": "fb0f1a9cdb0434bc9ce50eb6473bc2ca4a71d6994ad61314bbd7e2f11c4e29e8",
+  "corpus_sha256": "181716ce3f0a71b94f7d5d4428232c241b88f3e26d68f4ed50a191506ddb78b5",
+  "token_ids_sha256": "6dd2b70cb1d4a18a2955737b9dc3711bc20cd951c48a086ce144dcfbff8eb47b",
+  "policy": {"input_lengths": [513, 513], "starts": [0, 512], "targets": 1024, "bos": false},
+  "runtime": {"backend": "Vulkan", "gpu": "NVIDIA GeForce RTX 4090 24GB", "q2tp_gpu_opt_in": true, "gpu_probe": false},
+  "chunks": [
+    {"start": 0, "input_len": 513, "nll": 1085.211889936, "count": 512},
+    {"start": 512, "input_len": 513, "nll": 771.316238749, "count": 512}
+  ],
+  "nll": 1856.528128685,
+  "mean_nll": 1.813015751,
+  "ppl": 6.128902833,
+  "reference_mean_nll": 1.81274952826,
+  "delta_nats_per_token": 0.000266223
+}

evaluation/official-reference-summary.json

@@ -0,0 +1,16 @@
+{
+  "source": "official Prism PQ2 reference",
+  "source_model_sha256": "3907dc1658db1f78a9826bf8d5bcb8dc65db0d466388937af57f2294fae62ec1",
+  "source_runtime_commit": "9a9394a895b96003ca842a6041cb28ac49a108f7",
+  "corpus_sha256": "181716ce3f0a71b94f7d5d4428232c241b88f3e26d68f4ed50a191506ddb78b5",
+  "token_count": 1122,
+  "score_count": 1024,
+  "bos_injected": false,
+  "chunk_starts": [0, 512],
+  "mean_nll": 1.81274952826,
+  "ppl": 6.12727139849,
+  "top1_correct": 603,
+  "top1_accuracy": 0.5888671875,
+  "raw_file": "reference-teacher-score.jsonl",
+  "raw_sha256": "0fe3ce05070664d453186040a0013d1c66431f581252382bd2fe361f69577d7c"
+}

evaluation/performance-summary.json

@@ -0,0 +1,10 @@
+{
+  "measurement": "accepted runtime4 hardware run",
+  "hardware": "NVIDIA GeForce RTX 4090 24GB",
+  "prefill_tok_s": 8.5,
+  "ttft_s": 6.88,
+  "warm_decode_tok_s": 5.2,
+  "clean_port_quality_reverified": true,
+  "clean_port_speed_rebenchmarked": false,
+  "note": "5.2 tok/s is the runtime4 measurement retained for context, not a clean-port speed claim."
+}

evaluation/prompts.jsonl

@@ -0,0 +1,8 @@
+{"id": "en-explain", "thinking": false, "prompt": "In two concise sentences, explain why the sky looks blue.", "check": "coherent explanation of preferential scattering of short wavelengths"}
+{"id": "ru-explain", "thinking": false, "prompt": "Объясни по-русски в двух предложениях, зачем резервные копии нужны даже при наличии RAID.", "check": "Russian; RAID not replacement for independent backups"}
+{"id": "arithmetic", "thinking": false, "prompt": "Compute 37 * 24. Respond with the integer only.", "check": "888"}
+{"id": "python-function", "thinking": false, "prompt": "Write a Python function unique_in_order(items) that removes duplicates while preserving first occurrence order. Return only the function code. Items are hashable.", "check": "syntax parse and tests [],[1,1,2,1,3],[\"a\",\"b\",\"a\"]"}
+{"id": "strict-json", "thinking": false, "prompt": "Return only a valid JSON object with exactly two keys: \"city\" with value \"Казань\" and \"count\" with integer value 3. Do not use Markdown.", "check": "JSON exact keys/types/values"}
+{"id": "fact-context", "thinking": false, "prompt": "Use only these notes: The blue box contains 7 pencils. The green box contains 4 erasers. Nina moved 2 pencils from the blue box into an empty red box. How many pencils remain in the blue box? Answer with the number only.", "check": "5"}
+{"id": "thinking-math", "thinking": true, "prompt": "A shop discounts an item by 20%, then applies another 10% discount to the reduced price. The final price is 72 dollars. What was the original price?", "check": "100 dollars, with correct multiplicative reasoning"}
+{"id": "thinking-logic", "thinking": true, "prompt": "There are three boxes labeled apples, oranges, and mixed. Every label is wrong. You may take one fruit from one box without looking inside. Which labeled box should you choose to identify all three boxes, and why?", "check": "choose mixed-labeled box; infer other two using all labels wrong"}

evaluation/quality-corpus-token-ids.txt

@@ -0,0 +1,1122 @@
+2014
+14246
+35935
+1220
+30982
+264
+18100
+494
+449
+15673
+13
+357
+18100
+7189
+1092
+6716
+1172
+5024
+4530
+13
+1473
+15673
+52789
+3069
+424
+6716
+321
+1189
+1144
+2361
+9064
+13
+1368
+264
+1957
+8772
+10281
+1238
+3679
+4203
+11
+279
+17856
+7249
+1503
+524
+10229
+864
+2222
+8640
+13
+357
+5222
+11960
+4203
+799
+7911
+11
+13205
+279
+10804
+321
+11436
+4842
+8097
+11
+321
+42151
+279
+18100
+1238
+7948
+446
+13
+561
+1830
+1220
+1048
+2830
+26946
+11
+4598
+1056
+2476
+27482
+279
+6103
+467
+466
+3140
+16219
+8213
+13
+271
+32
+2526
+6092
+13205
+15687
+2919
+303
+1330
+25042
+13
+561
+1118
+12526
+20539
+2919
+494
+264
+1575
+11
+321
+279
+2018
+20539
+2919
+1238
+72386
+13
+357
+11661
+6473
+279
+3202
+303
+1754
+12526
+2957
+791
+9104
+13
+561
+3202
+628
+2222
+1881
+6473
+11
+748
+264
+6259
+1220
+524
+9370
+421
+279
+5362
+21254
+369
+1990
+4581
+13
+3095
+264
+11661
+17397
+28378
+11
+279
+1785
+1220
+3150
+421
+279
+869
+369
+9496
+13
+997
+62683
+279
+1483
+1324
+1973
+449
+25969
+5016
+628
+9959
+279
+6463
+1881
+264
+14733
+12526
+321
+264
+10548
+11661
+13
+271
+623
+264
+6522
+11
+264
+15922
+10426
+12361
+1332
+264
+2236
+1220
+381
+9476
+13
+1049
+369
+524
+264
+2880
+314
+279
+2236
+13
+357
+2716
+11782
+628
+1438
+264
+6396
+9973
+3315
+279
+2236
+369
+9468
+11
+694
+424
+4021
+54488
+22090
+449
+23410
+11316
+494
+279
+3889
+13
+33105
+11
+264
+16338
+12669
+1189
+20480
+279
+1928
+4221
+364
+1010
+8806
+1345
+12806
+3447
+4221
+364
+3672
+13
+561
+5222
+3296
+369
+524
+15756
+3315
+264
+12669
+369
+2526
+11
+694
+864
+4602
+628
+1990
+381
+18018
+28299
+494
+424
+13
+271
+35944
+264
+9403
+5832
+440
+2943
+17397
+13
+357
+5404
+8211
+506
+279
+29128
+11
+20406
+279
+62944
+11
+279
+2812
+11
+279
+26910
+11
+321
+279
+2981
+11
+321
+1179
+4523
+310
+279
+29128
+13
+561
+1906
+1189
+381
+51948
+539
+8306
+3979
+430
+1575
+430
+5821
+13
+561
+37847
+50717
+5832
+369
+524
+13902
+279
+24830
+40803
+5832
+13
+1368
+279
+26910
+15417
+2843
+1056
+279
+975
+10155
+11
+30957
+2166
+3988
+628
+1791
+264
+1248
+3655
+13
+357
+9412
+8646
+3749
+2107
+5640
+2942
+321
+2393
+10708
+13
+271
+81573
+207530
+173848
+177349
+172102
+150354
+17460
+180723
+84496
+7347
+63153
+11
+156412
+45480
+159052
+35587
+13
+68728
+183325
+159796
+177215
+18096
+211822
+11
+173347
+192481
+194711
+11
+45810
+177587
+172977
+6499
+148319
+181920
+13
+22220
+181124
+170429
+173598
+191835
+13785
+11
+149269
+194227
+148876
+69729
+154840
+45480
+158438
+16669
+13
+170113
+150354
+47552
+172073
+18096
+170000
+19244
+183325
+175655
+25
+168818
+148293
+177782
+227585
+11
+230487
+193784
+7422
+183145
+7347
+243190
+193346
+177488
+148803
+177371
+173697
+7347
+193107
+223895
+187849
+13
+271
+15729
+176544
+149193
+232456
+229142
+195738
+13
+68728
+173436
+191382
+173011
+18096
+42395
+28696
+7105
+11
+5624
+229142
+221469
+157976
+73781
+150604
+34083
+13
+68728
+13532
+154385
+68608
+148702
+152595
+6240
+84706
+181144
+153174
+170855
+179384
+11
+232254
+152984
+150984
+166449
+47745
+174039
+13
+167426
+35458
+45480
+191331
+5241
+220385
+169892
+226942
+18096
+176253
+177299
+42395
+28696
+1739
+13
+171985
+22346
+188665
+173347
+178962
+11
+7958
+194169
+152595
+182791
+223698
+179905
+11
+181618
+161573
+177007
+69829
+179224
+43572
+156359
+13
+182938
+179491
+176017
+180598
+7550
+149090
+150843
+148275
+7347
+160693
+148317
+206838
+16797
+54025
+148293
+42395
+28696
+1739
+13
+271
+19736
+152209
+153901
+242034
+149213
+154428
+208239
+11
+45810
+12696
+148427
+50800
+26149
+148445
+36183
+150165
+159226
+214880
+13
+22220
+176607
+171160
+150237
+153446
+178627
+11
+149269
+173983
+186794
+11
+19774
+5624
+245999
+183073
+159824
+173488
+13
+151459
+28696
+1739
+209970
+158353
+149195
+13
+193684
+152172
+26149
+34981
+148292
+13302
+27851
+214880
+11
+242034
+149213
+190234
+183407
+149925
+15630
+42317
+7347
+42395
+190793
+228043
+13
+159223
+19648
+151399
+2941
+56287
+11
+45810
+175564
+233208
+11
+19648
+148337
+220495
+175368
+7958
+170916
+157196
+150577
+13
+177566
+174771
+232254
+53921
+7422
+7721
+150182
+148369
+169929
+170267
+183407
+193331
+11
+170412
+44711
+161997
+29196
+231448
+54308
+148369
+172152
+219877
+36413
+85610
+36103
+148317
+13
+271
+173600
+155735
+155739
+171297
+20177
+153683
+2941
+53062
+42240
+153354
+172073
+183932
+177798
+11
+45810
+171272
+176952
+181593
+199418
+232680
+13
+170130
+154414
+60038
+38745
+11777
+170942
+174693
+78788
+45480
+149323
+148414
+173498
+13
+170130
+38745
+11777
+5241
+151387
+199019
+78788
+148554
+169969
+59551
+54308
+180808
+149076
+216494
+19648
+162793
+16669
+11
+150009
+152628
+168886
+162884
+170189
+196987
+148366
+17460
+13
+170130
+151184
+153305
+60038
+240406
+149323
+148414
+173498
+186065
+152930
+18096
+219593
+149323
+148414
+159787
+218842
+13
+171775
+38745
+164071
+182889
+11
+211525
+233431
+7347
+149925
+57868
+42308
+199849
+150149
+177007
+77955
+174921
+11
+19774
+18096
+230872
+66694
+13
+271
+727
+4752
+1201
+7621
+23573
+1590
+198
+262
+3753
+283
+716
+363
+198
+262
+1067
+283
+2958
+198
+262
+364
+1455
+303
+3470
+25
+198
+285
+413
+1455
+524
+303
+3753
+25
+198
+309
+3753
+1316
+5225
+8
+198
+309
+1067
+1989
+5225
+8
+198
+262
+460
+1067
+271
+727
+6992
+37314
+19490
+11
+3136
+1590
+198
+262
+413
+3136
+2564
+220
+15
+25
+198
+285
+4673
+14953
+437
+5334
+1902
+381
+6572
+871
+198
+262
+413
+2343
+19490
+8
+361
+3136
+25
+198
+285
+460
+2958
+198
+262
+2702
+283
+2542
+19490
+3337
+5334
+2387
+198
+262
+1067
+283
+498
+4873
+593
+3136
+60
+198
+262
+364
+1857
+303
+2020
+15438
+11
+2343
+19490
+35497
+198
+285
+2702
+1373
+2663
+5733
+60
+471
+2663
+5733
+471
+3136
+60
+198
+285
+1067
+1989
+22131
+593
+3136
+8
+198
+262
+460
+1067
+271
+727
+10562
+25146
+17145
+11
+1245
+1590
+198
+262
+25170
+283
+6249
+17145
+8
+198
+262
+364
+1328
+11
+1698
+303
+1245
+9328
+4406
+198
+285
+25170
+7900
+60
+283
+25170
+652
+4702
+11
+220
+15
+8
+478
+1698
+198
+262
+460
+25170
+271
+760
+1118
+709
+13205
+279
+1118
+30520
+314
+1754
+5010
+470
+1455
+13
+561
+2018
+709
+5533
+264
+32713
+2542
+11
+4598
+1056
+35970
+611
+286
+1396
+3136
+494
+18211
+13
+561
+4688
+709
+11121
+264
+491
+10675
+321
+1503
+524
+5427
+2892
+1879
+13
+4081
+5706
+628
+381
+12152
+24358
+25
+1906
+44026
+11
+18169
+11256
+11
+33625
+2961
+11
+321
+279
+18682
+314
+16015
+25532
+513
+2086
+5306
+314
+279
+4954
+13
+271
+32
+1228
+15705
+369
+1379
+36953
+948
+424
+5469
+2107
+4472
+321
+23595
+10804
+13
+1690
+264
+6992
+5289
+11
+449
+4147
+1103
+440
+264
+6572
+3136
+4523
+874
+47187
+13
+357
+3136
+314
+6942
+369
+8060
+321
+1220
+4673
+449
+1412
+13
+357
+3136
+5951
+310
+279
+2984
+314
+279
+1103
+18082
+6681
+799
+869
+13
+49146
+4947
+513
+18541
+795
+11
+524
+449
+1412
+2876
+13
+3744
+42092
+1439
+4886
+1518
+7873
+24469
+279
+59288
+310
+80403
+2316
+1238
+8847
+264
+1067
+13
+198

evaluation/quality-corpus.txt

@@ -0,0 +1,46 @@
+An engineering notebook should distinguish a measurement from an explanation. A measurement records what happened under specified conditions. An explanation proposes why it happened and may need another experiment. If a program becomes faster after several changes, the timing alone does not identify which change helped. A useful comparison changes one factor, keeps the inputs and hardware settings fixed, and repeats the measurement after warmup. The report should also include failures, rather than discarding the slowest or least convenient runs.
+
+A small town keeps drinking water in two tanks. The first tank receives water from a well, and the second receives water after filtration. A sensor reports the amount in each tank once per minute. The amount can change between reports, so a controller should not assume that the latest observation is still exact. When a sensor stops responding, the system should record that the value is unknown. Repeating the last number without an uncertainty flag can hide the difference between a stable tank and a broken sensor.
+
+In a library, a catalog tells readers where a book should be stored. It is not a copy of the book. A short summary can help a reader decide whether the book is relevant, but it cannot reliably reproduce an arbitrary sentence from the original. Similarly, a compact representation may preserve the information needed for some tasks while losing details needed for others. The useful question is not merely whether a representation is small, but which questions can still be answered accurately from it.
+
+Consider a delivery route with four stops. A driver starts at the warehouse, visits the bakery, the school, the clinic, and the market, and then returns to the warehouse. The order may be constrained by opening hours as well as distance. The shortest geometric route is not necessarily the fastest feasible route. If the clinic opens later than the other locations, arriving too early can create a long wait. A schedule therefore needs both travel times and service windows.
+
+Надёжная программа должна различать отсутствие значения и значение, равное нулю. Если датчик температуры не ответил, нельзя автоматически считать, что температура составляет ноль градусов. В первом случае информация неизвестна, во втором получено конкретное измерение. Это различие важно не только для датчиков: пустой список заказов, ошибка подключения к базе и запрет доступа означают разные ситуации и требуют разной реакции.
+
+В мастерской ведут журнал инструментов. Если новый инструмент ещё не записан, в журнал добавляют отдельную строку. Если у существующего инструмента изменилось место хранения, исправляют соответствующее поле. Повторное сообщение с теми же данными не требует новой записи. Когда из сообщения нельзя понять, о каком инструменте идёт речь, сначала уточняют его название или номер. Такой порядок помогает избежать дубликатов и случайной замены чужой записи.
+
+Однажды библиотекарь обнаружил, что на полке перепутаны три коробки. В первой должны лежать карты, во второй фотографии, а в третьей оба вида материалов. Подписи оказались неверными. Прежде чем переставлять коробки, библиотекарь решил проверить содержимое и записать наблюдения. Он отделил то, что действительно увидел, от предположений о других коробках. Благодаря этому исправление каталога можно было проверить независимо, даже если первоначальная догадка оказалась бы ошибочной.
+
+При сравнении двух вычислительных систем важно заранее определить, что именно считается одинаковым результатом. Для целых чисел часто требуется точное совпадение. Для чисел с плавающей точкой может быть допустимо небольшое отклонение, но величину допуска нужно обосновать. Для те��стовых ответов совпадение отдельных слов не гарантирует совпадения смысла. Поэтому численные проверки, проверка формата и содержательная оценка дополняют друг друга, а не заменяют.
+
+def unique_in_order(items):
+    seen = set()
+    result = []
+    for item in items:
+        if item not in seen:
+            seen.add(item)
+            result.append(item)
+    return result
+
+def moving_average(values, window):
+    if window <= 0:
+        raise ValueError("window must be positive")
+    if len(values) < window:
+        return []
+    total = sum(values[:window])
+    result = [total / window]
+    for index in range(window, len(values)):
+        total += values[index] - values[index - window]
+        result.append(total / window)
+    return result
+
+def merge_counts(left, right):
+    merged = dict(left)
+    for key, count in right.items():
+        merged[key] = merged.get(key, 0) + count
+    return merged
+
+The first function keeps the first occurrence of each hashable item. The second function uses a sliding sum, rather than recomputing every window from scratch. The third function creates a new dictionary and does not modify either input. These properties can be tested separately: order preservation, boundary handling, numerical results, and the absence of unexpected mutation are different parts of the contract.
+
+A test suite is most informative when it includes both normal and exceptional inputs. For a moving average, an empty list with a positive window returns no averages. A window of zero is invalid and should raise an error. A window equal to the length of the list produces exactly one value. Negative numbers are ordinary data, not an error condition. Describing these cases before implementation reduces the temptation to redefine success after seeing a result.

evaluation/reference-teacher-score.cpp

@@ -0,0 +1,208 @@
+#include "llama.h"
+
+#include <algorithm>
+#include <chrono>
+#include <cmath>
+#include <cstdint>
+#include <fstream>
+#include <iomanip>
+#include <iostream>
+#include <limits>
+#include <string>
+#include <vector>
+
+namespace {
+
+struct Score {
+    int index;
+    int input_index;
+    int target;
+    int top1;
+    double logprob;
+};
+
+bool read_tokens(const char * path, std::vector<llama_token> & tokens) {
+    std::ifstream in(path);
+    if (!in) {
+        std::cerr << "cannot open token file: " << path << "\n";
+        return false;
+    }
+    long long value = 0;
+    while (in >> value) {
+        if (value < std::numeric_limits<int32_t>::min() || value > std::numeric_limits<int32_t>::max()) {
+            std::cerr << "token out of int32 range\n";
+            return false;
+        }
+        tokens.push_back(static_cast<llama_token>(value));
+    }
+    if (!in.eof()) {
+        std::cerr << "malformed token file\n";
+        return false;
+    }
+    return true;
+}
+
+bool score_chunk(llama_context * ctx, const std::vector<llama_token> & tokens, int start,
+                 int n_vocab, std::vector<Score> & scores, std::string & error) {
+    constexpr int chunk = 512;
+    llama_batch batch = llama_batch_init(chunk, 0, 1);
+    if (!batch.token || !batch.pos || !batch.n_seq_id || !batch.seq_id || !batch.logits) {
+        llama_batch_free(batch);
+        error = "llama_batch_init returned incomplete batch";
+        return false;
+    }
+    for (int i = 0; i < chunk; ++i) {
+        batch.token[i] = tokens[start + i];
+        batch.pos[i] = i;
+        batch.n_seq_id[i] = 1;
+        batch.seq_id[i][0] = 0;
+        batch.logits[i] = 1;
+    }
+    batch.n_tokens = chunk;
+
+    const int rc = llama_decode(ctx, batch);
+    if (rc != 0) {
+        error = "llama_decode returned " + std::to_string(rc);
+        llama_batch_free(batch);
+        return false;
+    }
+
+    for (int i = 0; i < chunk; ++i) {
+        const float * logits = llama_get_logits_ith(ctx, i);
+        if (!logits) {
+            error = "llama_get_logits_ith returned null at row " + std::to_string(i);
+            llama_batch_free(batch);
+            return false;
+        }
+        float max_logit = -std::numeric_limits<float>::infinity();
+        int top1 = -1;
+        for (int v = 0; v < n_vocab; ++v) {
+            if (logits[v] > max_logit) {
+                max_logit = logits[v];
+                top1 = v;
+            }
+        }
+        double sum_exp = 0.0;
+        for (int v = 0; v < n_vocab; ++v) {
+            sum_exp += std::exp(static_cast<double>(logits[v]) - static_cast<double>(max_logit));
+        }
+        const int target = tokens[start + i + 1];
+        if (target < 0 || target >= n_vocab || !(sum_exp > 0.0)) {
+            error = "invalid target or logit normalization at row " + std::to_string(i);
+            llama_batch_free(batch);
+            return false;
+        }
+        const double logprob = static_cast<double>(logits[target]) - static_cast<double>(max_logit) - std::log(sum_exp);
+        scores.push_back({start + i + 1, start + i, target, top1, logprob});
+    }
+    llama_batch_free(batch);
+    return true;
+}
+
+} // namespace
+
+int main(int argc, char ** argv) {
+    if (argc != 4) {
+        std::cerr << "usage: llama-teacher-score MODEL TOKEN_IDS OUTPUT_JSONL\n";
+        return 2;
+    }
+
+    std::vector<llama_token> tokens;
+    if (!read_tokens(argv[2], tokens) || tokens.size() < 1025) {
+        std::cerr << "need at least 1025 token ids, got " << tokens.size() << "\n";
+        return 2;
+    }
+
+    llama_backend_init();
+    llama_model_params model_params = llama_model_default_params();
+    model_params.n_gpu_layers = -1;
+    llama_model * model = llama_model_load_from_file(argv[1], model_params);
+    if (!model) {
+        std::cerr << "llama_model_load_from_file failed\n";
+        llama_backend_free();
+        return 3;
+    }
+
+    llama_context_params context_params = llama_context_default_params();
+    context_params.n_ctx = 512;
+    context_params.n_batch = 512;
+    context_params.n_ubatch = 512;
+    context_params.n_seq_max = 1;
+    context_params.n_outputs_max = 512;
+    context_params.n_outputs_max_per_seq = 512;
+    context_params.n_threads = 16;
+    context_params.n_threads_batch = 48;
+    context_params.flash_attn_type = LLAMA_FLASH_ATTN_TYPE_AUTO;
+    context_params.op_offload = true;
+    context_params.kv_unified = false;
+    llama_context * ctx = llama_init_from_model(model, context_params);
+    if (!ctx) {
+        std::cerr << "llama_init_from_model failed\n";
+        llama_model_free(model);
+        llama_backend_free();
+        return 4;
+    }
+
+    const llama_vocab * vocab = llama_model_get_vocab(model);
+    const int n_vocab = llama_vocab_n_tokens(vocab);
+    std::vector<Score> scores;
+    scores.reserve(1024);
+    std::string error;
+    const auto started = std::chrono::steady_clock::now();
+    for (int start : {0, 512}) {
+        if (start != 0) {
+            llama_memory_clear(llama_get_memory(ctx), true);
+        }
+        if (!score_chunk(ctx, tokens, start, n_vocab, scores, error)) {
+            std::cerr << error << "\n";
+            llama_free(ctx);
+            llama_model_free(model);
+            llama_backend_free();
+            return 5;
+        }
+    }
+    llama_synchronize(ctx);
+    const auto elapsed = std::chrono::duration<double>(std::chrono::steady_clock::now() - started).count();
+
+    double nll = 0.0;
+    int top1_correct = 0;
+    for (const Score & s : scores) {
+        nll -= s.logprob;
+        top1_correct += s.top1 == s.target;
+    }
+    const double mean_nll = nll / static_cast<double>(scores.size());
+    const double ppl = std::exp(mean_nll);
+
+    std::ofstream out(argv[3]);
+    if (!out) {
+        std::cerr << "cannot open output: " << argv[3] << "\n";
+        llama_free(ctx);
+        llama_model_free(model);
+        llama_backend_free();
+        return 6;
+    }
+    out << std::setprecision(12);
+    out << "{\"type\":\"meta\",\"commit\":\"9a9394a895b96003ca842a6041cb28ac49a108f7\","
+           "\"vocab\":" << n_vocab << ",\"token_count\":" << tokens.size()
+        << ",\"chunk_size\":512,\"chunk_starts\":[0,512],\"bos_injected\":false,"
+           "\"score_count\":" << scores.size() << ",\"elapsed_seconds\":" << elapsed << "}\n";
+    for (const Score & s : scores) {
+        out << "{\"type\":\"score\",\"target_index\":" << s.index
+            << ",\"input_index\":" << s.input_index
+            << ",\"target\":" << s.target
+            << ",\"correct_logprob\":" << s.logprob
+            << ",\"top1\":" << s.top1 << "}\n";
+    }
+    out << "{\"type\":\"summary\",\"score_count\":" << scores.size()
+        << ",\"mean_nll\":" << mean_nll
+        << ",\"ppl\":" << ppl
+        << ",\"top1_correct\":" << top1_correct
+        << ",\"top1_accuracy\":" << (static_cast<double>(top1_correct) / scores.size())
+        << ",\"elapsed_seconds\":" << elapsed << "}\n";
+    out.close();
+
+    llama_free(ctx);
+    llama_model_free(model);
+    llama_backend_free();
+    return 0;
+}

licenses/README.md

@@ -0,0 +1,15 @@
+# License bundle
+
+This staging bundle contains both sides of the distribution:
+
+- `upstream-model-LICENSE` and `upstream-model-NOTICE.txt` are copied from
+  `prism-ml/Ternary-Bonsai-2-27B-mlx-2bit` at revision
+  `3f926b415992eaa2ae9dd7b573706494d6bbf787`.
+- `upstream-mlx-runtime-LICENSE` preserves the MIT notice shipped by the
+  upstream `runtime/` component.
+- `cmf-LICENSE`, `cmf-NOTICE`, and `cmf-PATENTS.md` cover the CMF runtime and
+  clean source patch. `cmf-NOTICE` references `cmf-PATENTS.md`.
+
+The upstream notice also identifies Qwen3.8-27B as Apache-2.0 software and
+links its license. This release does not remove or replace that attribution.
+Do not represent the CMF runtime license as the model license, or vice versa.

licenses/cmf-LICENSE

@@ -0,0 +1,201 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright 2026 Oleg Kirichenko
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

licenses/cmf-NOTICE

@@ -0,0 +1,17 @@
+CMF — Cortiq Model Format
+Copyright 2026 Oleg Kirichenko
+
+This product includes software developed by Oleg Kirichenko.
+
+Licensed under the Apache License, Version 2.0 (the "License").
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+------------------------------------------------------------------------
+
+This software implements methods that are the subject of patent
+applications filed by the author. The Apache License, Version 2.0 grants
+you an irrevocable, royalty-free patent license to those claims that are
+necessarily infringed by the software as distributed (see Section 3 of the
+License and the file PATENTS.md for details).

licenses/cmf-PATENTS.md

@@ -0,0 +1,50 @@
+# Patents
+
+CMF (Cortiq Model Format) implements methods that are the subject of the
+following United States patent applications filed by the author, Oleg
+Urevich Kirichenko:
+
+| # | Application Serial No. | Filed | Title |
+|---|------------------------|-------|-------|
+| 1 | 19/452,440 | January 19, 2026 | Resonance Routing (unsupervised task selection by reconstruction-error minimization) |
+| 2 | 19/452,464 | January 19, 2026 | Dynamic Task-Guided Mask Activation (DTG-MA) Compression |
+| 3 | 19/731,402 | July 6, 2026 | Unified Execution Architecture for Serving a Plurality of Specialized Language Models from a Single Shared Backbone via Dynamically Overlaid Compressed Delta Representations Without Materializing Separate Models |
+| 4 | 19/738,763 | July 13, 2026 | Training-Free Constant-Memory Streaming Attention Conversion |
+
+Where these methods appear in this repository:
+
+- **Resonance Routing (App. 19/452,440)** — the reconstruction-error skill
+  selector used by the runtime and the `selection descriptor` records in the
+  container.
+- **DTG-MA Compression (App. 19/452,464)** — the task-guided mask/skill
+  production path used by the converter to derive per-skill delta records.
+- **Unified Execution Architecture (App. 19/731,402)** — the CMF container
+  itself (shared backbone stored once + per-skill full-shape replacement
+  tensors + byte-offset delta index) and the overlay runtime
+  that reads replacement tensors *in place of* the backbone at forward time
+  without materializing a per-skill model.
+- **Streaming Attention Conversion (App. 19/738,763)** — the training-free
+  O(1) attention path: `cortiq convert --o1` (weights byte-identical), the
+  streaming sink/window/landmark-skeleton kernel with delayed insertion and
+  a single joint denominator (`nystrom.rs`, runtime seal that releases the
+  KV cache), and the generation-gated `cortiq fcd` restoration stage.
+
+## Patent grant
+
+This project is licensed under the **Apache License, Version 2.0**. Section 3
+of that license ("Grant of Patent License") gives every user a perpetual,
+worldwide, non-exclusive, no-charge, royalty-free, irrevocable patent license
+to make, have made, use, offer to sell, sell, import, and otherwise transfer
+the Work, for those patent claims of the above applications that are
+necessarily infringed by this software as distributed. That patent license
+terminates for any party that initiates patent litigation alleging that this
+software infringes a patent (Apache-2.0 §3).
+
+In plain terms: **you may use, modify, and redistribute this software —
+including for commercial purposes — under Apache-2.0, and you receive a
+patent license to the claims practiced by the code as shipped.** The
+underlying inventions remain patented; this grant is scoped to this software,
+consistent with Apache-2.0.
+
+This file is informational and does not modify the Apache License, Version 2.0,
+which is the sole governing license (see `LICENSE`).

licenses/upstream-mlx-runtime-LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright © 2023 Apple Inc.
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

licenses/upstream-model-LICENSE

@@ -0,0 +1,177 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+

licenses/upstream-model-NOTICE.txt

@@ -0,0 +1,4 @@
+This software is copyright 2026-present Prism ML, Inc. It is available under the Apache 2.0 license.
+If you publicly deploy or redistribute this software, we would appreciate attribution such as: "Created using Bonsai by Prism ML."
+
+This software is built from Qwen3.8-27B, Copyright 2026 Alibaba Cloud, which is available under the Apache 2.0 License: https://huggingface.co/Qwen/Qwen3.8-27B/blob/main/LICENSE

source/SOURCE_BUILD.md

@@ -0,0 +1,68 @@
+# Reproducible clean-source runtime
+
+This model requires the small source patch shipped at
+`source/patches/cmf-bonsai2-public-base-48a9.patch`. It is not compatible with
+the stock `cortiq` 0.6.8 release: the version label is retained by the pinned
+public base, while `q2tp_affine` and its signed-Hadamard descriptor are added
+by this patch. No runtime binary is bundled here.
+
+## Pin and patch
+
+```sh
+git clone https://github.com/infosave2007/cmf.git
+cd cmf
+git checkout --detach 48a9ca5929ea42d28ced54279b4fc9d9c4c33d10
+git apply --check /path/to/source/patches/cmf-bonsai2-public-base-48a9.patch
+git apply /path/to/source/patches/cmf-bonsai2-public-base-48a9.patch
+```
+
+The patch SHA-256 is
+`e6cd6a60149a8be764a323ab1f968668b524b36ec35bc1a193fcfb5218f669be`.
+The patch changes only the accepted 26-file Bonsai delta and preserves the
+public Qwen image modules and manifests. The source model provenance and all
+changed-file hashes are in `source/SOURCE_PROVENANCE.json`.
+
+## Linux build
+
+Install the normal Rust/Cargo build prerequisites plus Vulkan development and
+runtime support (`build-essential pkg-config cmake libssl-dev libvulkan1
+vulkan-tools`). The validated environment used stable Rust 1.98.1
+(`48a229cea`, 2026-09-01), Vulkan on an NVIDIA RTX 4090, and no HF token.
+
+```sh
+source "$HOME/.cargo/env"  # when rustup installed this file
+rustup toolchain install 1.98.1
+export RUSTUP_TOOLCHAIN=1.98.1
+export CARGO_TARGET_DIR="$PWD/target"
+export CARGO_BUILD_JOBS=8 CARGO_INCREMENTAL=0
+export RUSTFLAGS='-C opt-level=3'
+cargo build -p cortiq-cli --features gpu
+cargo build -p cortiq-cli --features gpu --example score_fixed
+export PATH="$CARGO_TARGET_DIR/debug:$PATH"
+```
+
+For Vulkan set `XDG_RUNTIME_DIR=/tmp WGPU_BACKEND=vulkan`. CPU-only checks can
+force `CMF_GPU=0`. The source validation passed core 6/6, q2tp CPU 3/3, Prism
+CPU 2/2, CLI 94 passed/1 ignored live-HF, and Vulkan q2tp 2/2.
+
+## Verify and run the uploaded model
+
+The model repository is
+`https://huggingface.co/infosave/Ternary-Bonsai-2-27B-cmf` and the
+file name is `Ternary-Bonsai-2-27B-Q2TP-Affine.cmf`. Download that file and verify its SHA-256 against
+`source/SOURCE_PROVENANCE.json`.
+
+```sh
+MODEL=Ternary-Bonsai-2-27B-Q2TP-Affine.cmf
+export XDG_RUNTIME_DIR=/tmp WGPU_BACKEND=vulkan CMF_GPU=wgpu
+export CMF_Q2TP_GPU=1 CMF_GPU_PROBE=0 CMF_MTP=0
+cortiq verify "$MODEL"
+cortiq run "$MODEL" \
+  --prompt 'Return only a valid JSON object with exactly two keys: "city" with value "Казань" and "count" with integer value 3. Do not use Markdown.' \
+  --greedy --seed 42 --no-think --max-tokens 256
+```
+
+The validated output was exactly `{"city":"Казань","count":3}`. This is a
+text-generation runtime check. Vision tensors and metadata are preserved in
+the CMF file, but image/text fusion is not implemented or tested by this
+runtime; do not use this file as a VLM.

source/SOURCE_PROVENANCE.json

@@ -0,0 +1,224 @@
+{
+  "schema": "ternary-bonsai-2-q2tp-affine-hf-staging-1",
+  "intended_repo": "infosave/Ternary-Bonsai-2-27B-cmf",
+  "intended_model_filename": "Ternary-Bonsai-2-27B-Q2TP-Affine.cmf",
+  "model": {
+    "bytes": 8247698536,
+    "sha256": "fb0f1a9cdb0434bc9ce50eb6473bc2ca4a71d6994ad61314bbd7e2f11c4e29e8",
+    "profile": "CMF dtype16 Q2TP storage + required q2tp_affine Prism operator",
+    "tested_modality": "text generation only"
+  },
+  "source": {
+    "runtime_repository": "https://github.com/infosave2007/cmf",
+    "public_base_commit": "48a9ca5929ea42d28ced54279b4fc9d9c4c33d10",
+    "patch_file": "source/patches/cmf-bonsai2-public-base-48a9.patch",
+    "patch_sha256": "e6cd6a60149a8be764a323ab1f968668b524b36ec35bc1a193fcfb5218f669be",
+    "changed_file_count": 26,
+    "changed_files": [
+      {
+        "path": "README.md",
+        "bytes": 8911,
+        "sha256": "bec1ab6c898dbb3be6afa304acc26f3b2b58c3f2b52bce0d164075af3e988aeb"
+      },
+      {
+        "path": "crates/cortiq-cli/examples/score_fixed.rs",
+        "bytes": 1230,
+        "sha256": "d1b909ed40995249fa75c445ddf40a367ca6adca001eba31d74ac370669b779c"
+      },
+      {
+        "path": "crates/cortiq-cli/src/convert.rs",
+        "bytes": 356447,
+        "sha256": "35aa91d49af5231f6ca990974a337823de0ccc10062091221a7bd833870beedc"
+      },
+      {
+        "path": "crates/cortiq-cli/src/gguf.rs",
+        "bytes": 98248,
+        "sha256": "e1f8afbf0b7e6d8fdf7d4921cc80747a5ca9bf9bf5bc599f72c589aba8ba8139"
+      },
+      {
+        "path": "crates/cortiq-cli/src/main.rs",
+        "bytes": 249728,
+        "sha256": "ea4cdaf09dda0e2fcb64a8159d35d8c11877137664e7616bd8c4dc4f88ebb3d0"
+      },
+      {
+        "path": "crates/cortiq-core/src/format.rs",
+        "bytes": 85313,
+        "sha256": "9a8fca48b8492fdfab4eb1ac122feb28604ece4cbdd7cab74f2c02a69cfec3a7"
+      },
+      {
+        "path": "crates/cortiq-core/src/hadamard.rs",
+        "bytes": 3315,
+        "sha256": "264eeba873021217b594965e904a6c6b604e384e9ceadd5e03163fd9f71dc428"
+      },
+      {
+        "path": "crates/cortiq-core/src/lib.rs",
+        "bytes": 775,
+        "sha256": "278d49a2b121b6851acabb812d2b5ace28f9ab0a5bdd3970045110c290969286"
+      },
+      {
+        "path": "crates/cortiq-core/src/types.rs",
+        "bytes": 39778,
+        "sha256": "a51f36c2112ed574b03aad086f3d96148787889becdb297904446454769eb8cf"
+      },
+      {
+        "path": "crates/cortiq-core/tests/loop_masks.rs",
+        "bytes": 6567,
+        "sha256": "6408c23b8de1cd6db3bb00875f315565270f82bd6028ba77253ccf69fc83529b"
+      },
+      {
+        "path": "crates/cortiq-core/tests/roundtrip.rs",
+        "bytes": 37607,
+        "sha256": "2ddb95cf55eee59a14f34884ae595132cf18a5c1e3bb2694e4f23c473c067091"
+      },
+      {
+        "path": "crates/cortiq-core/tests/skill_file.rs",
+        "bytes": 4886,
+        "sha256": "af99818e3671b992da2b4b343edf9f8055b388c8c8d4e11a883bab19d3ddcd4b"
+      },
+      {
+        "path": "crates/cortiq-engine/src/attention.rs",
+        "bytes": 70420,
+        "sha256": "246c9c720ed82a547a9af01f0c755b6f82601cb56a3344f231b985bbb4aadde5"
+      },
+      {
+        "path": "crates/cortiq-engine/src/gpu.rs",
+        "bytes": 121790,
+        "sha256": "bd6e651e5cafa585263e2c24049e70c2e8cc84a31ade294e3dd92c60b172d1ad"
+      },
+      {
+        "path": "crates/cortiq-engine/src/gpu_metal.rs",
+        "bytes": 597506,
+        "sha256": "78b4eecbee135aac938b30b774c57cb48ceb88a8e4f3a03330569f1b1898e431"
+      },
+      {
+        "path": "crates/cortiq-engine/src/gpu_wgpu.rs",
+        "bytes": 2089683,
+        "sha256": "2ade4ccc62940ba23453189f14ea0a1bf4cb315e8a86ee8be373f63d04eaf2b8"
+      },
+      {
+        "path": "crates/cortiq-engine/src/lib.rs",
+        "bytes": 3901,
+        "sha256": "517b35207c05307d4688349ff5da1505d877bb23b15b5f73a49a28c3aee4dd89"
+      },
+      {
+        "path": "crates/cortiq-engine/src/pipeline.rs",
+        "bytes": 584307,
+        "sha256": "a67dbc8768bd739c7b2f950385d6144dc5cee801245f95537ed7970aa3ea98f0"
+      },
+      {
+        "path": "crates/cortiq-engine/src/prism.rs",
+        "bytes": 12114,
+        "sha256": "16505deed5a0a00b1d38a7a842b746f00223d1d15d3a7dbb34c1bf3a79dfe45a"
+      },
+      {
+        "path": "crates/cortiq-engine/src/qtensor.rs",
+        "bytes": 512619,
+        "sha256": "57135f0a4ca811135f566c1c3288c3cad98122a2371483c4eb6c821607b1c3f7"
+      },
+      {
+        "path": "crates/cortiq-engine/tests/fcd_train.rs",
+        "bytes": 19591,
+        "sha256": "66c2998ed92a30001701f27928fa4ded0981b087dc4c13494f91e26daf31dc5a"
+      },
+      {
+        "path": "crates/cortiq-engine/tests/gpu_micro.rs",
+        "bytes": 25495,
+        "sha256": "b86eb91fe01a71bf21474f69a8789b8e4af458e4292952768fb147cf7a7214c9"
+      },
+      {
+        "path": "crates/cortiq-engine/tests/mask_quant.rs",
+        "bytes": 7199,
+        "sha256": "115453717aeed21e06c32741e617e074a9486d1b7fec1f413a37d6c622604976"
+      },
+      {
+        "path": "crates/cortiq-engine/tests/skills.rs",
+        "bytes": 15446,
+        "sha256": "235f27cb98f6e20c42a107a77cef2469d98df6b440baac2c4394c0ee831e5040"
+      },
+      {
+        "path": "crates/cortiq-engine/tests/x86_gemm.rs",
+        "bytes": 28107,
+        "sha256": "a3c1747f6179f1ee16e8f026ab4b687f013dbbe09a59b1407cfe5996142389d8"
+      },
+      {
+        "path": "python/cmf_reader.py",
+        "bytes": 14373,
+        "sha256": "54aca01be03f2c57f9d62ff867420394d7e61f579876a1bf795cec8ce497f67b"
+      }
+    ],
+    "runtime4_delta_sha256": "55add36a786bc9697817d5b2696b2e64f810926cf34663ae995fb3ad83a1f2dd",
+    "binary_not_distributed": true
+  },
+  "upstream_model": {
+    "repository": "prism-ml/Ternary-Bonsai-2-27B-mlx-2bit",
+    "revision": "3f926b415992eaa2ae9dd7b573706494d6bbf787",
+    "weights_file": "model.safetensors",
+    "weights_bytes": 8595477990,
+    "weights_sha256": "130de5925082c168b7866b2e91b52e44abbafc99017e3ca352b77b5b55a269ed",
+    "license": "Apache-2.0"
+  },
+  "validated_runtime_identity": {
+    "version_label": "cortiq 0.6.8",
+    "source_build_sha256": "0ccf849b132b2216aaff5f4e6762a1182a3a44fcf2329ad20503c2e896ab6eb3",
+    "build_id": "2b94566b1ce9d2cb7218df3735c2f24bc8838235",
+    "note": "development build from the pinned base plus this patch; not stock 0.6.8"
+  },
+  "evaluation": {
+    "prompts_file": "evaluation/prompts.jsonl",
+    "prompts_sha256": "68f6b0b2ffc8526d2f37df7bb2e494531d533e336b6d4df8c1feeebd2652c1f5",
+    "quality_corpus_file": "evaluation/quality-corpus.txt",
+    "quality_corpus_sha256": "181716ce3f0a71b94f7d5d4428232c241b88f3e26d68f4ed50a191506ddb78b5",
+    "token_ids_file": "evaluation/quality-corpus-token-ids.txt",
+    "token_ids_sha256": "6dd2b70cb1d4a18a2955737b9dc3711bc20cd951c48a086ce144dcfbff8eb47b",
+    "official_reference_raw": "evaluation/reference-teacher-score.jsonl",
+    "official_reference_raw_sha256": "0fe3ce05070664d453186040a0013d1c66431f581252382bd2fe361f69577d7c",
+    "official_reference_mean_nll": 1.81274952826,
+    "official_reference_ppl": 6.12727139849,
+    "official_reference_top1": "603/1024",
+    "cmf_affine_cpu_mean_nll": 1.812992284,
+    "cmf_affine_vulkan_mean_nll": 1.813015751,
+    "cmf_affine_vulkan_ppl": 6.128902833,
+    "ordinary_q2tp_vulkan_mean_nll": 1.886951899,
+    "ordinary_q2tp_vulkan_delta_nats": 0.074202371
+  },
+  "performance": {
+    "measurement_source": "accepted runtime4 hardware run on RTX 4090 24GB",
+    "prefill_tok_s": 8.5,
+    "ttft_s": 6.88,
+    "warm_decode_tok_s": 5.2,
+    "clean_port_quality_reverified": true,
+    "clean_port_speed_rebenchmarked": false,
+    "interpretation": "5.2 tok/s is not a clean-port speed claim; it is the runtime4 measurement retained for context."
+  },
+  "licenses": {
+    "files": [
+      {
+        "path": "licenses/upstream-model-LICENSE",
+        "sha256": "69849221bfb90053de2134ef5e6d540287b4b98062326492f1f96f5da685524b"
+      },
+      {
+        "path": "licenses/upstream-model-NOTICE.txt",
+        "sha256": "de0e0c48fb6f691a31e74f338e3ccf93f9ecdfe2866ab769c4bf8b79a7636a30"
+      },
+      {
+        "path": "licenses/upstream-mlx-runtime-LICENSE",
+        "sha256": "ccfab7ccb2ea306f71531c8ca77bb55507606cd90768b1e32b8b52ab5b48cf01"
+      },
+      {
+        "path": "licenses/cmf-LICENSE",
+        "sha256": "0bf59d136536dea5dcf80ce96c5988e46522ce69f3ae817666c7f5d10ded6618"
+      },
+      {
+        "path": "licenses/cmf-NOTICE",
+        "sha256": "d8ee9168d48f5b0fa0012518848434a65cebb8288f3f2ed51f2738f2f4b347e4"
+      },
+      {
+        "path": "licenses/cmf-PATENTS.md",
+        "sha256": "607faa71dd5a600787f97dcc50e17221e04f57e21e19aa2d5e2a1399e21fb692"
+      }
+    ],
+    "model_license": "Apache-2.0",
+    "cmf_runtime_license": "Apache-2.0",
+    "upstream_mlx_runtime_license": "MIT"
+  }
+}