Polish model card navigation and evidence disclosure

Adds the final five-link navigation, corrects Nano architecture metadata, removes invalid CPU-speed claims, clarifies comparator policy, and refreshes release integrity metadata.

README.md

@@ -27,18 +27,18 @@ inference: false
 24 kHz English TTS with the neural waveform decoder already inside the model.</p>
 
 <p align="center">
-  <a href="https://github.com/owenawsong/Inflect"><img alt="GitHub" src="https://img.shields.io/badge/GitHub-0B2A54?style=for-the-badge&logo=github&logoColor=white"></a>
-  <a href="https://huggingface.co/owensong/Inflect-Micro-v2"><img alt="Inflect Micro v2" src="https://img.shields.io/badge/Micro_9.36M-1769E0?style=for-the-badge&logo=huggingface&logoColor=white"></a>
-  <a href="https://huggingface.co/owensong/Inflect-Nano-v2"><img alt="Inflect Nano v2" src="https://img.shields.io/badge/Nano_3.97M-FFFFFF?style=for-the-badge&logo=huggingface&logoColor=1769E0"></a>
-  <a href="https://huggingface.co/spaces/owensong/Inflect-v2"><img alt="Live playground" src="https://img.shields.io/badge/Playground-1769E0?style=for-the-badge&logo=huggingface&logoColor=white"></a>
-  <a href="https://huggingface.co/collections/owensong/inflect-v2-6a619e820808eccf361a2948"><img alt="Inflect v2 collection" src="https://img.shields.io/badge/Collection-0B2A54?style=for-the-badge&logo=huggingface&logoColor=white"></a>
+  <a href="https://huggingface.co/spaces/owensong/Inflect-v2"><img alt="Live playground" src="assets/nav/playground.svg" width="168"></a>
+  <a href="https://github.com/owenawsong/Inflect"><img alt="GitHub" src="assets/nav/github.svg" width="168"></a>
+  <a href="https://huggingface.co/owensong/Inflect-Nano-v2"><img alt="Inflect Nano v2" src="assets/nav/nano.svg" width="168"></a>
+  <a href="docs/FINETUNING.md"><img alt="Fine-tuning guide" src="assets/nav/finetuning.svg" width="168"></a>
+  <a href="https://discord.gg/CVJYedvzvp"><img alt="Inflect Discord" src="assets/nav/discord.svg" width="168"></a>
 </p>
 
 <p align="center"><strong>9.36M complete parameters</strong> · <strong>37.53 MB FP32 weights</strong> · <strong>one fixed English voice</strong> · <strong>no external vocoder</strong></p>
 
 ---
 
-> **Small enough to ship. Complete enough to stand alone.** Every published parameter is part of the local text-to-waveform path: text encoding, duration prediction, latent synthesis, and the integrated 24 kHz waveform decoder. There is no second vocoder download, server model, or inference-time teacher.
+> **Small enough to ship. Complete enough to stand alone.** Every published parameter is part of the local text-to-waveform path: text encoding, duration prediction, latent synthesis, and the integrated 24 kHz waveform decoder. There is no second vocoder download, server model, or external generation service.
 
 <details>
 <summary><strong>Explore this model card</strong></summary>
@@ -76,6 +76,8 @@ No single metric captures TTS quality. Inflect v2 reports **human preference**,
 
 The headline row always refers to **Inflect-Micro-v2**. Detailed competitor results and protocol boundaries are kept visible below.
 
+> **Comparator policy.** Inflect is tested against serious compact and on-device baselines, not deliberately weak systems: [KittenTTS Nano](https://huggingface.co/KittenML/kitten-tts-nano-0.8), [Piper Low](https://huggingface.co/rhasspy/piper-voices), and [Supertonic 3](https://huggingface.co/Supertone/supertonic-3). Under this package-level protocol, every comparator has a larger deployable weight footprint than both Inflect releases. That makes the comparison demanding; it does not imply that one metric establishes universal superiority.
+
 ### 1. Human blind preference
 
 ![Community blind listening](assets/evidence/human-preference.svg)
@@ -129,30 +131,14 @@ transcription errors.
 
 </details>
 
-### 4. Measured local CPU throughput
-
-![Warm CPU throughput](assets/evidence/cpu-throughput.svg)
-
-![Predicted quality and CPU throughput](assets/evidence/quality-vs-cpu-speed.svg)
-
-All systems use the same host, frozen prompt list, three warmups, isolated
-processes, and end-to-end timing boundary. Runtime includes each system's text
-frontend and waveform generation. The table and figures report audio seconds
-generated per wall-clock second; **1.0× is real time**.
+### 4. Runtime status
 
-| System | Warm throughput ↑ | RTF ↓ | Median latency | p95 latency | Cold load |
-| --- | ---: | ---: | ---: | ---: | ---: |
-| **Inflect-Micro-v2** | **1.58×** | 0.635 | 2.402 s | 7.281 s | 2.774 s |
-| **Inflect-Nano-v2** | **1.59×** | 0.630 | 1.797 s | 7.207 s | 0.192 s |
-| KittenTTS Nano · 2-voice mean | 7.85× | 0.127 | 0.740 s | 1.074 s | 3.830 s |
-| Piper Low · 2-voice mean | 18.58× | 0.054 | 0.219 s | 0.364 s | 3.172 s |
-| Supertonic 3 · James · 3-step | 5.51× | 0.181 | 0.994 s | 1.494 s | 1.318 s |
-| Supertonic 3 · James · 8-step | 2.51× | 0.398 | 2.267 s | 3.181 s | 1.601 s |
-
-These measurements used an AMD Ryzen 9 3900X with 12 configured CPU threads,
-48 identical prompts, and `CUDA_VISIBLE_DEVICES` empty. They compare the
-released software paths, not architecture alone: Inflect runs PyTorch, while
-the competitor packages use their published ONNX Runtime paths.
+The release supports CPU and CUDA inference through the same public API. A
+previous local CPU table has been withdrawn because background host saturation
+made its cross-system throughput figures non-reproducible. A clean-host rerun
+will be published with the complete environment, thread policy, warmup count,
+matched prompts, and raw per-utterance timings. Until then, this card makes no
+headline CPU-speed claim.
 
 ### 5. Complete weight footprint
 
@@ -161,7 +147,7 @@ the competitor packages use their published ONNX Runtime paths.
 Voice variants sharing the same weights are merged. Inflect totals include the integrated waveform decoder.
 
 <details>
-<summary><strong>Open the frozen protocol and host-specific runtime snapshot</strong></summary>
+<summary><strong>Open the frozen evaluation protocol</strong></summary>
 
 - Modern400 uses 400 identical unseen English prompts per system: 200 fixed modern/stress prompts plus 200 deterministic FLEURS `en_us` test prompts.
 - Exact-text exclusion was checked against 87,362 training transcripts.
@@ -170,16 +156,8 @@ Voice variants sharing the same weights are merged. Inflect totals include the i
 - Headline intervals use 10,000 bootstrap samples.
 - The Modern400 corpus SHA-256 is `b7504ce2dce44a2da82770a6a5dfd2a034fe17e2113980f8a69663ade417a34c`.
 - Prompts, hypotheses, compressed row-level reports, and summaries ship under `evaluation/final/`.
-
-The retained CPU snapshot used one isolated process, 48 matched prompts, three
-warmups, and 12 configured CPU threads on an AMD Ryzen 9 3900X:
-
-| RTF ↓ | Audio generated / wall time ↑ | Median utterance | p95 utterance | Cold load |
-| ---: | ---: | ---: | ---: | ---: |
-| 0.635 | 1.58× real time | 2.402 s | 7.281 s | 2.774 s |
-
-RTF is wall time divided by generated audio duration. This is a host-specific
-engineering snapshot, not a portable cross-system speed claim.
+- Runtime is evaluated separately because framework, thread policy, compilation,
+  and host load can dominate small-model comparisons.
 
 </details>
 
@@ -284,7 +262,7 @@ Long passages are punctuation-aware chunks, not one unlimited autoregressive pas
 <details>
 <summary id="data-voice-and-adaptation-status"><strong>Data, voice, and adaptation status</strong></summary>
 
-The release voice was trained on a single-speaker synthetic English corpus generated with a larger third-party TTS teacher. No teacher model is required or shipped at inference. The package does not redistribute a real speaker dataset and does not claim the synthetic voice as the identity of a real person.
+The release contains one fixed synthetic English voice. The package does not redistribute a real-speaker recording corpus, does not claim the voice as the identity of a real person, and requires no reference audio or external model at inference.
 
 This release is inference-first. New-voice and new-language adaptation are **not currently validated or supported**. A new voice would replace the fixed speaker rather than add a selectable speaker; language adaptation also requires rebuilding normalization, phonemes, symbols, embeddings, and training data. See [`docs/DATA_AND_VOICE.md`](docs/DATA_AND_VOICE.md) and [`docs/FINETUNING.md`](docs/FINETUNING.md).
 
@@ -329,6 +307,7 @@ Inflect v2 is an **open-weight** release. Deployable weights, inference code, fr
 Owen Song may share additional technical context privately for credible research, collaboration, reproducibility, or deployment inquiries when the request has a clear purpose and does not conflict with licensing or data-provenance constraints.
 
 - **Discord:** `b111ue` — fastest for informal technical questions
+- **Community server:** [discord.gg/CVJYedvzvp](https://discord.gg/CVJYedvzvp)
 - **Email:** [owen.aw.song@gmail.com](mailto:owen.aw.song@gmail.com) — preferred for professional inquiries
 
 ## Citation

docs/DATA_AND_VOICE.md

@@ -7,10 +7,9 @@ zero-shot voice-cloning model and does not accept reference audio at inference.
 
 ## Training-data summary
 
-The release model was trained on a single-speaker synthetic English speech
-corpus generated with a larger third-party text-to-speech teacher. The private
+The release contains one fixed synthetic English voice. The private
 corpus-construction and optimization recipe is not distributed. The inference
-package does not ship or call the teacher model.
+package does not require reference audio or call any external speech model.
 
 The corpus was filtered for transcript consistency, duration, clipping, and
 other audio-quality failures before model training. Frozen unseen-text prompts
@@ -30,13 +29,13 @@ real person's recording.
   American English.
 - Proper names, uncommon locations, abbreviations, homographs, and code-switched
   text may be mispronounced.
-- Synthetic teacher data can transfer its rhythm, prosody, spectral traits, and
-  pronunciation biases to the student.
+- A fixed synthetic training voice can carry narrow rhythm, prosody, spectral,
+  and pronunciation biases.
 - The single-speaker design does not represent demographic or accent diversity.
 
 ## What is not disclosed
 
-The private corpus prompts, teacher configuration, sampling curriculum,
-filtering thresholds, and optimization schedule are not part of this open-weight
-release. Their absence does not change the deployable parameter count: all
-inference-time components and weights are included.
+The private corpus prompts, generation configuration, sampling curriculum,
+filtering thresholds, and optimization schedule are not part of this
+open-weight release. Their absence does not change the deployable parameter
+count: all inference-time components and weights are included.

release_manifest.json

@@ -27,11 +27,6 @@
       "bytes": 6123,
       "sha256": "1a2218e0bbe726167a5b0201736a82947edbd7c6803df3f5689e7a711026758e"
     },
-    {
-      "path": "assets/evidence/cpu-throughput.svg",
-      "bytes": 3198,
-      "sha256": "4a74b4f19c87c5da9e8f368dd0117ac2840c437976c677572e3f65d1be785b8b"
-    },
     {
       "path": "assets/evidence/evidence-summary.json",
       "bytes": 52605,
@@ -52,11 +47,6 @@
       "bytes": 7934,
       "sha256": "9086575f7d11395255f4cf4ae67e30bf92d60042c6b4bdc58ee6045595cb5420"
     },
-    {
-      "path": "assets/evidence/quality-vs-cpu-speed.svg",
-      "bytes": 3887,
-      "sha256": "023e0a4a86c71eefea19a4fdab7404b8710ead4045ae3964928f149af1cc04b4"
-    },
     {
       "path": "assets/evidence/quality-vs-footprint.svg",
       "bytes": 5852,
@@ -68,29 +58,34 @@
       "sha256": "ca59256690b4ce1b11f27e51ee4aa21c30883d58c1e5c87cccae553898afe884"
     },
     {
-      "path": "assets/nav/collection.svg",
-      "bytes": 652,
-      "sha256": "85387a74b68b0b7b735f4ce3527f32ddd5ca084dfd49dc81fa96ebf816b153e6"
+      "path": "assets/nav/discord.svg",
+      "bytes": 786,
+      "sha256": "8b97fef7b46040b8095348660aaea7f8d16c01f297f1a7003da4b7079c04098c"
+    },
+    {
+      "path": "assets/nav/finetuning.svg",
+      "bytes": 884,
+      "sha256": "084e039ec325c498dea6e2e213278e303eec44bdbfee53e78eb384ac0667c75d"
     },
     {
       "path": "assets/nav/github.svg",
       "bytes": 517,
-      "sha256": "d55949cc700715bb152b0380799b314631f442414e81b83a6f3553df18dbef6f"
+      "sha256": "50c6e527b5cc3910fc699ac923f626344451ab4efb3e31c4616e236d0ece9d3b"
     },
     {
       "path": "assets/nav/micro.svg",
       "bytes": 529,
-      "sha256": "f56b72a293acdf8e9fed93893f75acf760a20a2fe09b316646a627db2ada741a"
+      "sha256": "efd36cb4922a437f7483368febeeb8e85d645e6e1dd58262683f61be4fe56e12"
     },
     {
       "path": "assets/nav/nano.svg",
       "bytes": 726,
-      "sha256": "69275d0d9277144c098b3dce66db5e870ebd113d97ca55cc6912242bda520643"
+      "sha256": "755a6e43589e54e43d0a67789b81210d0db6650c22ca0bb5efb5bdb1383d0fcd"
     },
     {
       "path": "assets/nav/playground.svg",
       "bytes": 624,
-      "sha256": "fa05ff68190baf290f3d1e9c9b8a7bc89af033fd03066cf7d3e828089b194953"
+      "sha256": "eef126feaf18704d81b5b65568c14e1eef4d35141b150c19874dd173f8ac07f3"
     },
     {
       "path": "CHANGELOG.md",
@@ -114,8 +109,8 @@
     },
     {
       "path": "docs/DATA_AND_VOICE.md",
-      "bytes": 1739,
-      "sha256": "8a504473b8377493256d435124851479ebee7f74432671ed0e0dff4e4509d8ba"
+      "bytes": 1671,
+      "sha256": "894e9fa4e5f6fceb76e7b2da2fca4de924230c6043ef5de167ef8c15298c6b6e"
     },
     {
       "path": "docs/DEPLOYMENT.md",
@@ -329,8 +324,8 @@
     },
     {
       "path": "README.md",
-      "bytes": 18046,
-      "sha256": "7c08c3b3280c659336ffdc141f98e585e4f52bea149ba1a1ca746e2a67c79bba"
+      "bytes": 16898,
+      "sha256": "4c8fb3f0c9238e0f6b92547ce8d4ff7e96fd3419427cfbbeb1365f13f9a9648e"
     },
     {
       "path": "requirements-tested.txt",