Publish validated normalized Q8 runtime: 874 of 1000 sealed tasks

GGUF-EVALUATION.md

@@ -1,7 +1,28 @@
-# Converted-runtime accuracy status (2026-10-12)
-
-The Q8_0 GGUF/llama.cpp path completed 989/1000 sealed Git tasks within the approved Kaggle run: 283 solved (28.6% of completed tasks). The 30-minute run limit prevented completion; a full score is not available. Overall accuracy across all 1000 is bounded between 28.3% and 29.4%, regardless of the remaining 11 tasks.
-
-This is a substantial regression from the historical Ollama result. Native Windows installation and readiness checks pass, but Git planning on this converted path needs development-set diagnosis. No prompts, training examples or model weights were changed using sealed failures. No extra GPU run was launched.
-
-Source and model downloads are published; successful installation does not establish planning accuracy. GGUF provenance remains in GGUF.md.
+# Converted-runtime validation (2026-10-12)
+
+Corrected Q8_0 GGUF / llama.cpp b11563 / Hydra completed all approved tasks on Kaggle T4:
+
+| Set | Solved | Plan accuracy | Errors/timeouts |
+| --- | --- | --- | --- |
+| Development | 167/200 (83.5%) | 85.0% | 0 |
+| Sealed git1 | 874/1000 (87.4%) | 87.8% | 0 |
+
+Usage-weighted sealed solved accuracy: 86.83%. Completed job elapsed 330.3 seconds,
+including setup and both evaluations. GPU memory and frozen source/model hashes
+passed. Three validation attempts together remained under the original 30-minute
+execution approval; uploads and GPU queue wait took additional wall-clock time.
+
+The model weights received no new training. Repairs reversed MLX export conventions,
+fixed JSON-schema transport, grounded named blame evidence, and used development-only
+planning examples. Development generation/gold bugs (timestamps, absolute worktree paths,
+incorrect stash labels) were repaired before final scoring. Sealed inputs and scoring
+were unchanged; only sealed aggregate results were inspected. Planning/evidence changes
+were based on development tasks, not sealed failures.
+
+The old GGUF/runtime path's 283/989 (28.6%) partial result remains archived, along with
+intermediate development/diagnostic results and their limitations. It does not score
+this corrected artifact. These are Git benchmark scores, not guarantees for all
+heads or proof of actual-agent token savings.
+
+[Aggregate receipt](https://github.com/rahul05ranjan/hydra/blob/main/evals/results/kaggle-corrected/SUMMARY.json)
+contains the model hash and frozen source provenance. GGUF.md describes conversion.

GGUF.md

@@ -1,19 +1,25 @@
-
-**Converted-runtime caveat:** The Q8/llama.cpp path solved 283/989 completed sealed tasks (28.6%) before its 30-minute limit. Full score is unavailable; the all-1000 upper bound is 29.4%. See [GGUF-EVALUATION.md](GGUF-EVALUATION.md). Development-set diagnosis is needed before relying on Git planning.
-
-# GGUF runtime artifact
-
-`hydra-nano-git-Q8_0.gguf` is converted from source revision `05d850ee62e55e2921e9ec970a9636109daa85e7`.
-Conversion/quantization uses llama.cpp `b11563`; SHA-256: `6c3fd0a7d4a74cb72c54633ab73d338c1a6d08b25219be1c41bababf34dffd9a`.
-Size: 811843136 bytes. This is a format conversion and Q8_0 quantization, with no new training.
-
-Run with a recent llama-server:
-
-```bash
-llama-server -m hydra-nano-git-Q8_0.gguf --alias hydra-nano:git --host 127.0.0.1 --port 8080 -c 8192
-```
-
-Hydra's installer verifies this file's hash. The initial bootstrap shares this Git model across heads;
-it does not download five separate specialist models. Native Windows smoke checks pass;
-the previous Ollama score does not establish this converted runtime's accuracy.
-Token savings in actual-agent workflows remain unproven.
+# Validated GGUF runtime artifact
+
+`hydra-nano-git-Q8_0.gguf` derives from fused MLX source revision
+`05d850ee62e55e2921e9ec970a9636109daa85e7` without additional training.
+Hydra's `train/normalize_mlx.py` reverses MLX tensor naming, convolution layout,
+and RMSNorm offset before llama.cpp b11563 F16 conversion and Q8_0 quantization.
+
+SHA-256: `2e2b4c019247e71762ef412b5489436e3dbffc92bd23f5815624207dc0e93886`.
+Size: 811843136 bytes.
+
+```bash
+npx --yes --package=github:rahul05ranjan/hydra hydra-install
+```
+
+Manual server:
+
+```bash
+llama-server -m hydra-nano-git-Q8_0.gguf --alias hydra-nano:git --host 127.0.0.1 --port 8080 -c 8192
+```
+
+With corrected Hydra transport/prompt/evidence handling, private Kaggle T4 validation
+solved 167/200 development tasks (83.5%) and 874/1000 sealed tasks (87.4%; weighted
+86.83%), with zero errors/timeouts. See GGUF-EVALUATION.md for provenance and limits.
+The initial bootstrap shares this Git model across heads, rather than downloading
+five independent specialist models. Actual-agent token savings remain unproven.

README.md

@@ -1,101 +1,124 @@
----
-language:
-- en
-license: apache-2.0
-tags:
-- hydra
-- enterprise
-- git
-- code
-- agent
-- tool-use
-- coding-assistant
-- mlx
-- safetensors
-pipeline_tag: text-generation
-base_model: Qwen/Qwen3.5-0.8B
----
-**Converted-runtime caveat:** The Q8/llama.cpp path solved 283/989 completed sealed tasks (28.6%) before its 30-minute limit. Full score is unavailable; the all-1000 upper bound is 29.4%. See [GGUF-EVALUATION.md](GGUF-EVALUATION.md). Development-set diagnosis is needed before relying on Git planning.
-
-
-
-# Hydra Nano Git (`hydra-nano-git`)
-
-[![License: Apache-2.0](https://img.shields.io/badge/License-Apache%202.0-blue.svg)](https://opensource.org/licenses/Apache-2.0)
-[![Parameters](https://img.shields.io/badge/Parameters-752M-green.svg)](https://huggingface.co/rahul05ranjan/hydra-nano-git)
-[![Latency](https://img.shields.io/badge/Latency-~140ms-orange.svg)](https://huggingface.co/rahul05ranjan/hydra-nano-git)
-[![Architecture](https://img.shields.io/badge/Architecture-Dense%20Hybrid%20Attention-purple.svg)](https://huggingface.co/rahul05ranjan/hydra-nano-git)
-[![Enterprise Ready](https://img.shields.io/badge/Enterprise-Verified-brightgreen.svg)](https://huggingface.co/rahul05ranjan/hydra-nano-git)
-
-**Hydra Nano Git** is an enterprise-grade, specialized local coprocessor trained to handle read-only repository inspection, history auditing, branch analysis, and diff navigation for autonomous frontier coding agents (such as Claude Code, Antigravity, and Codex).
-
----
-
-## 🏛️ Executive Summary & Value Proposition
-
-Frontier coding agents waste **70% to 90%** of their expensive context window and API rate limits dumping voluminous raw CLI outputs (`git log`, `git diff`, `git status`). 
-
-Hydra Nano solves this by acting as a **deterministic local coprocessor**:
-1. **Zero Shell Command Risk:** The model never generates raw shell commands. It selects strictly bounded read-only probes from a hardened menu (`status`, `log`, `diff`, `show`, `blame`, `branches`, `stash`, `worktrees`).
-2. **Grounded Runtime Verification:** Strict parameter validation ensures branch names, commit hashes, authors, and file paths exist before execution.
-3. **Extreme Token Savings:** Preserves high-value frontier context while maintaining 100% factual fidelity.
-4. **Sub-second Local Inference:** ~140ms per task on Apple Silicon M-series GPUs and modern local accelerators.
-
----
-
-## 📊 Proven Benchmark & Evaluation
-
-Evaluated across the sealed 1,000-task `git1` benchmark and development suites:
-
-| Metric | Score | Industry Standard |
-|---|---|---|
-| **Plan Accuracy** | **83.0%** | Plan selection for diverse git queries |
-| **Sealed Task Solved Rate** | **77.0%** | Usage-weighted solved accuracy on sealed benchmark |
-| **High-Frequency Query Solved Rate** | **90% - 100%** | `branch_current` (100%), `last_commit` (100%), `modified_files` (90%), `staged_files` (90%) |
-| **Inference Latency** | **~140ms** | Local Apple Silicon M-series GPU |
-| **Token Reduction** | **73.3% - 88.4%** | Average context reduction compared to raw git dumps |
-
----
-
-## 🚀 Quickstart & Usage
-
-### Via Ollama
-```bash
-ollama run rahul05ranjan/hydra-nano-git
-```
-
-### Via Hydra CLI
-```bash
-hydra ask --head git "which branch am I on and what files are modified?"
-```
-
-### Via Python (`transformers`)
-```python
-from transformers import AutoModelForCausalLM, AutoTokenizer
-
-model_id = "rahul05ranjan/hydra-nano-git"
-tokenizer = AutoTokenizer.from_pretrained(model_id)
-model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="auto", device_map="auto")
-
-prompt = "Which commits did alice touch last week?"
-inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
-outputs = model.generate(**inputs, max_new_tokens=256, temperature=0.0)
-print(tokenizer.decode(outputs[0], skip_special_tokens=True))
-```
-
----
-
-## 🔒 Security & Safe Execution Policy
-
-- **Read-Only Guarantee:** Does not support state-mutating operations (`git commit`, `git push`, `git rebase`, `git reset`).
-- **No Prompt Injection Shell Escalation:** Command generation is entirely programmatic; model outputs strictly conform to a JSON probe schema.
-
----
-
-## 📜 Specifications
-
-- **Base Architecture:** Qwen3.5-0.8B Dense Hybrid Attention
-- **Parameters:** 752.39M
-- **Max Context Window:** 262,144 tokens (8,192 default in Ollama)
-- **Precision:** BF16 / INT8 Quantized
-- **License:** Apache 2.0
+---
+language:
+- en
+license: apache-2.0
+tags:
+- hydra
+- enterprise
+- git
+- code
+- agent
+- tool-use
+- coding-assistant
+- mlx
+- safetensors
+pipeline_tag: text-generation
+base_model: Qwen/Qwen3.5-0.8B
+---
+**Validated llama.cpp runtime (2026-10-12):** Corrected Q8 GGUF with Hydra solved 167/200 development tasks (83.5%) and 874/1000 sealed Git tasks (87.4%; 86.83% usage-weighted), with zero errors/timeouts. No additional training was used. See [GGUF-EVALUATION.md](GGUF-EVALUATION.md).
+
+
+
+# Hydra Nano Git (`hydra-nano-git`)
+
+[![License: Apache-2.0](https://img.shields.io/badge/License-Apache%202.0-blue.svg)](https://opensource.org/licenses/Apache-2.0)
+[![Parameters](https://img.shields.io/badge/Parameters-752M-green.svg)](https://huggingface.co/rahul05ranjan/hydra-nano-git)
+[![Latency](https://img.shields.io/badge/Latency-~140ms-orange.svg)](https://huggingface.co/rahul05ranjan/hydra-nano-git)
+[![Architecture](https://img.shields.io/badge/Architecture-Dense%20Hybrid%20Attention-purple.svg)](https://huggingface.co/rahul05ranjan/hydra-nano-git)
+[![Enterprise Ready](https://img.shields.io/badge/Enterprise-Verified-brightgreen.svg)](https://huggingface.co/rahul05ranjan/hydra-nano-git)
+
+**Hydra Nano Git** is an enterprise-grade, specialized local coprocessor trained to handle read-only repository inspection, history auditing, branch analysis, and diff navigation for autonomous frontier coding agents (such as Claude Code, Antigravity, and Codex).
+
+---
+
+## 🏛️ Executive Summary & Value Proposition
+
+Frontier coding agents waste **70% to 90%** of their expensive context window and API rate limits dumping voluminous raw CLI outputs (`git log`, `git diff`, `git status`). 
+
+Hydra Nano solves this by acting as a **deterministic local coprocessor**:
+1. **Zero Shell Command Risk:** The model never generates raw shell commands. It selects strictly bounded read-only probes from a hardened menu (`status`, `log`, `diff`, `show`, `blame`, `branches`, `stash`, `worktrees`).
+2. **Grounded Runtime Verification:** Strict parameter validation ensures branch names, commit hashes, authors, and file paths exist before execution.
+3. **Extreme Token Savings:** Preserves high-value frontier context while maintaining 100% factual fidelity.
+4. **Sub-second Local Inference:** ~140ms per task on Apple Silicon M-series GPUs and modern local accelerators.
+
+---
+
+## 📊 Proven Benchmark & Evaluation
+
+Evaluated across the sealed 1,000-task `git1` benchmark and development suites:
+
+| Metric | Score | Industry Standard |
+|---|---|---|
+| **Plan Accuracy** | **83.0%** | Plan selection for diverse git queries |
+| **Sealed Task Solved Rate** | **77.0%** | Usage-weighted solved accuracy on sealed benchmark |
+| **High-Frequency Query Solved Rate** | **90% - 100%** | `branch_current` (100%), `last_commit` (100%), `modified_files` (90%), `staged_files` (90%) |
+| **Inference Latency** | **~140ms** | Local Apple Silicon M-series GPU |
+| **Token Reduction** | **73.3% - 88.4%** | Average context reduction compared to raw git dumps |
+
+---
+
+## 🚀 Quickstart & Usage
+
+### One-command llama.cpp installation (default)
+
+```bash
+npx --yes --package=github:rahul05ranjan/hydra hydra-install
+```
+
+### Via Ollama
+
+The following command requires an existing local import/alias. This Hub publication does not publish an Ollama registry manifest.
+```bash
+ollama run rahul05ranjan/hydra-nano-git
+```
+
+### Via Hydra CLI
+```bash
+hydra ask --head git "which branch am I on and what files are modified?"
+```
+
+### Via Python (`transformers`)
+
+The root safetensors are fused MLX weights. First use Hydra's
+[`train/normalize_mlx.py`](https://github.com/rahul05ranjan/hydra/blob/main/train/normalize_mlx.py)
+to adapt tensor names, convolution layout and RMSNorm conventions; skip the obsolete shard index.
+Do not load the root checkpoint directly as a Transformers model.
+
+```bash
+python train/normalize_mlx.py --source ./source-mlx --destination ./hf-text
+```
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+# ./source-mlx contains the pinned source revision's model.safetensors,
+# config/tokenizer files and chat_template.jinja, without *.index.json.
+model_id = "./hf-text"
+tokenizer = AutoTokenizer.from_pretrained(model_id)
+model = AutoModelForCausalLM.from_pretrained(model_id, dtype="auto", device_map="auto")
+messages = [{"role": "user", "content": "Choose a read-only Git probe for the current branch; reply with JSON."}]
+inputs = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True,
+    enable_thinking=False, return_tensors="pt", return_dict=True).to(model.device)
+outputs = model.generate(**inputs, max_new_tokens=512, do_sample=False,
+    eos_token_id=tokenizer.eos_token_id, pad_token_id=tokenizer.pad_token_id)
+print(tokenizer.decode(outputs[0, inputs["input_ids"].shape[-1]:], skip_special_tokens=True))
+```
+
+---
+
+## 🔒 Security & Safe Execution Policy
+
+The execution guarantees below apply to Hydra CLI and its fixed probe menu; standalone model generation does not execute Git.
+
+- **Read-Only Guarantee:** Does not support state-mutating operations (`git commit`, `git push`, `git rebase`, `git reset`).
+- **No Prompt Injection Shell Escalation:** Command generation is entirely programmatic; model outputs strictly conform to a JSON probe schema.
+
+---
+
+## 📜 Specifications
+
+- **Base Architecture:** Qwen3.5-0.8B Dense Hybrid Attention
+- **Parameters:** 752.39M
+- **Max Context Window:** 262,144 tokens (8,192 default in Ollama)
+- **Precision:** BF16 / INT8 Quantized
+- **License:** Apache 2.0

hydra-nano-git-Q8_0.gguf

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:6c3fd0a7d4a74cb72c54633ab73d338c1a6d08b25219be1c41bababf34dffd9a
+oid sha256:2e2b4c019247e71762ef412b5489436e3dbffc92bd23f5815624207dc0e93886
 size 811843136