Refresh org card: two-tier framing, 23 verified (22 PPL + 1 ViT), v0.6.27, clean encoding, drop stale/banned terms

README.md

@@ -1,4 +1,4 @@
----
+---
 title: README
 emoji: 🧊
 colorFrom: indigo
@@ -8,48 +8,50 @@ pinned: false
 ---
 # Sipsa Labs
 
-> **Sipsa Labs** is an experimental and deep tech-and-software company. We invent and ship across the full breadth of tech and software.
+> **Sipsa Labs** is a deep-tech company that invents and ships across the breadth of AI infrastructure.
 >
 > **Two products live today:**
 >
-> 1. **UltraCompress** — lossless 5-bit transformer compression. Flagship. `pip install ultracompress`.
-> 2. **Sipsa Inference** — OpenAI-compatible serving API for our compressed weights. `https://api.sipsalabs.com/v1`. Pro $99/mo + Team $499/mo.
+> 1. **UltraCompress** — near-lossless 5-bit transformer compression (plus a genuinely lossless archival tier), with SHA-256-verifiable reconstruction. Flagship. `pip install ultracompress`.
+> 2. **Sipsa Inference** — an OpenAI-compatible serving API for our compressed weights at `https://api.sipsalabs.com/v1`.
 >
-> **Compression-as-a-Service** engagements are open ($5K Phase 0 POC). More products in flight — we don't pre-announce.
+> Phase 0 POC engagements are open (scoped, customer-picked model). More in flight — we don't pre-announce.
 
 ---
 
-## v0.6.22 current — minimal clean release
+## Current release — v0.6.27 (minimal, clean)
 
-Only `pip install ultracompress` (v0.6.22) is on PyPI. The package is the minimal SHA-256 verifier; codec internals are patent-protected and not shipped in the public package. Prior dev releases were retired during the 2026-05-18 cleanup. Mistral-7B-v0.3 hits 1.00548× — a strong dense 7B-class lossless 5-bit result.
+Only `pip install ultracompress` (v0.6.27) is on PyPI: a minimal SHA-256 verifier. Codec internals are patent-protected and are not shipped in the public package.
 
 ---
 
 ## Product 1: UltraCompress
 
-Production-grade lossless 5-bit transformer compression across 22 shipped architectures (20 PPL-verified: 19 transformer + 1 SSM with comparator-note caveat), dense + Mixture-of-Experts + state-space, 0.6B to 405B parameters. Mathematically lossless customer-side reconstruction: SHA-256 over reconstructed tensor bytes matches the reference manifest, verified by `uc verify`.
+**Two tiers, both with reproducible, SHA-256-verifiable reconstruction:**
+
+- A **genuinely lossless archival tier** — reconstructs the original weights bit-for-bit (~1.3–1.6× smaller) for exact storage and hand-off.
+- A **near-lossless 5-bit tier** — ~1% perplexity vs bf16 (lossy), ~3× smaller, for running frontier-scale models on small GPUs.
+
+**23 architectures verified**: 22 PPL-verified end-to-end (21 transformer — 17 dense + 4 MoE — plus 1 SSM with a comparator-note caveat) and 1 Vision Transformer cosine-verified (DINOv2-Large). Dense + MoE + SSM + ViT, 0.6B to 405B. `uc verify` checks SHA-256 over the reconstructed tensor bytes against the reference manifest.
 
-### Tightest verified PPL ratios at 5 bpw
-*(perplexity ratio = compressed PPL / bf16 baseline PPL; FineWeb-edu held-out tail; seq_len = 1024; n = 30-50; seed = 42)*
+### Verified 5-bit PPL ratios (sample)
 
-| Model | Params | Type | **PPL ratio** | Notes |
+*(perplexity ratio = compressed PPL / bf16 baseline PPL; FineWeb-edu held-out tail; seq_len = 1024; n = 30–50; seed = 42)*
+
+| Model | Params | Type | PPL ratio | Notes |
 |---|---:|:---|---:|:---|
-| **Mixtral-8x7B** | 47B | MoE | **1.00368×** | tightest MoE result |
-| Qwen3-1.7B-Base | 1.7B | dense | **1.00401×** | small-decoder record |
-| Qwen3-14B | 14B | dense | **1.00403×** | 14B-class record |
-| Yi-1.5-9B | 8.8B | dense | **1.00414×** | >8B record |
-| Qwen3-8B | 8B | dense | **1.00440×** | 8B-class record |
-| **Mistral-7B-v0.3** | 7B | dense | **1.00548×** | NEW — a strong dense 7B-class result |
-| Phi-3-mini-4k | 3.8B | dense | **1.00262×** | tightest dense ratio (seq_len=128 caveat) |
-| **Hermes-3-Llama-3.1-405B** | 405B | dense | **1.0066×** | 405B-class lossless on a single 32 GB consumer GPU |
-| Qwen3-0.6B | 0.6B | dense | **1.0069×** | |
-| OLMo-2-0425-1B | 1B | dense | **1.0073×** | |
-| SmolLM2-1.7B-Instruct | 1.7B | dense | **1.0075×** | |
-| SmolLM2-1.7B | 1.7B | dense | **1.0085×** | |
-| Mamba-2.8B | 2.8B | SSM | **1.0059×** | state-space model (uses SSM-compatible comparator; canonical transformer streaming reconstruction is architecture-incompatible with SSMs) |
-| Llama-3.1-8B | 8B | dense | **1.0125×** | |
-
-22 architectures shipped, 20 PPL-verified end-to-end (19 transformer + 1 SSM with comparator-note caveat). Full matrix below + at github.com/sipsalabs/ultracompress.
+| **Mixtral-8x7B** | 47B | MoE | **1.00368×** | tightest MoE result |
+| Qwen3-1.7B-Base | 1.7B | dense | **1.00401×** | small-decoder record |
+| Qwen3-14B | 14B | dense | **1.00403×** | 14B-class record |
+| Yi-1.5-9B | 8.8B | dense | **1.00414×** | >8B record |
+| Qwen3-8B | 8B | dense | **1.00440×** | 8B-class record |
+| Mistral-7B-v0.3 | 7B | dense | **1.00548×** | strong dense 7B-class result |
+| Phi-3-mini-4k | 3.8B | dense | **1.00262×** | tightest dense ratio (seq_len=128 caveat) |
+| **Hermes-3-Llama-3.1-405B** | 405B | dense | **1.0066×** | 405B-class near-lossless on a single 32 GB GPU (streamed from disk) |
+| Mamba-2.8B | 2.8B | SSM | **1.0059×** | state-space model — SSM-compatible comparator (the canonical transformer pipeline is architecture-incompatible with SSMs) |
+| Llama-3.1-8B | 8B | dense | **1.0125×** | honest worst case (pipeline-specific floor, not architectural) |
+
+Full matrix at [github.com/sipsalabs/ultracompress](https://github.com/sipsalabs/ultracompress).
 
 ### Try UltraCompress (3 commands)
 
@@ -59,53 +61,49 @@ hf download SipsaLabs/qwen3-1.7b-base-uc-v3-bpw5 --local-dir ./pack
 uc verify ./pack
 ```
 
-Expected: `VERIFY: PASS — bit-identical reconstruction guaranteed.`
+Expected: `VERIFY: PASS — bit-identical reconstruction.`
 
 ---
 
 ## Product 2: Sipsa Inference
 
-OpenAI-compatible inference API serving our compressed weights. Drop-in replacement for OpenAI's `base_url` — same `openai` Python SDK works unchanged.
+OpenAI-compatible inference API serving our compressed weights — a drop-in `base_url` swap; the `openai` Python SDK works unchanged.
 
 ```bash
 export OPENAI_BASE_URL=https://api.sipsalabs.com/v1
 curl $OPENAI_BASE_URL/models
 ```
 
-22 models live in the catalog with `sipsa-*` prefix (e.g. `model="sipsa-hermes-3-llama-3.1-405b"`). Backed by dual RTX 5090 over Cloudflare Tunnel.
-
-**Pricing**: Free $5 credit on signup (no card). **Pro $99/mo** (600 RPM, $100 included credit). **Team $499/mo** (2400 RPM, $500 included credit). Full pricing + bill estimator at sipsalabs.com/pricing.
+23 models live in the catalog with the `sipsa-*` prefix (e.g. `model="sipsa-hermes-3-llama-3.1-405b"`). **Free $5 credit on signup (no card).** Full pricing and a bill estimator at [sipsalabs.com/pricing](https://sipsalabs.com/pricing).
 
 ---
 
-## Service: Compression-as-a-Service (CaaS)
-
-Bring a model, we deliver a verified-lossless 5-bit pack you can run on your hardware. **Phase 0 POC is $5K / 5 business days / customer-picked model**. Day 7 deliverable is a pack you self-verify with `uc verify` + benchmark with `uc bench`. Acceptance gate: `uc verify PASS` + PPL ratio within 1.5% on your eval set.
+## Service: Compression-as-a-Service
 
-Email founder@sipsalabs.com.
+Bring a model; we deliver a verified 5-bit pack you run on your own hardware. Phase 0 POC is a scoped, customer-picked-model engagement with a self-verifiable (`uc verify`) deliverable and a PPL-ratio acceptance gate. Email founder@sipsalabs.com or see [sipsalabs.com/poc](https://sipsalabs.com/poc).
 
 ---
 
 ## Why this matters
 
-Defense / FDA-regulated healthcare / SR 11-7 model validation / frontier lab red-team eval — places where "good enough on MMLU" isn't enough. Bit-identical reconstruction is the actual reason to talk to us in those settings. As a side-effect of the streaming compression path, it lets us put a 405B-parameter model through a single 32 GB consumer GPU without renting an H100 cluster.
+Defense, FDA-regulated healthcare, SR 11-7 model validation, frontier-lab red-team eval — settings where "good enough on a benchmark" isn't enough. **SHA-256-verifiable reconstruction** — proving that what runs in production is exactly the validated artifact — is the reason to talk to us there. As a side effect, the streaming reconstruction path runs a 405B-parameter model through a single 32 GB consumer GPU.
 
 ---
 
 ## License + IP
 
-- **PyPI v0.6+** under BUSL-1.1 with **Additional Use Grant**: free for sub-$1M ARR companies, research, and individuals. Auto-converts to Apache 2.0 four years after each release.
-- **v0.5.x** stays under Apache-2.0 forever on the `legacy/0.5.x` branch.
-- Codec internals patent-protected. Continuations through 2027.
+- **PyPI v0.6+** under BUSL-1.1 with an **Additional Use Grant**: free for sub-$1M-ARR companies, research, and individuals. Auto-converts to Apache 2.0 four years after each release.
+- **v0.5.x** stays Apache-2.0 on the `legacy/0.5.x` branch.
+- Codec internals are patent-protected (continuations pending).
 
 ---
 
 ## Contact
 
-- **Commercial / Phase 0 POC** → founder@sipsalabs.com
-- **Patents / licensing** → legal@sipsalabs.com
-- **Press / media** → press@sipsalabs.com
-- **Security disclosure** → security@sipsalabs.com
-- **General** → hello@sipsalabs.com
+- **Commercial / Phase 0 POC** → founder@sipsalabs.com
+- **Patents / licensing** → legal@sipsalabs.com
+- **Press / media** → press@sipsalabs.com
+- **Security disclosure** → security@sipsalabs.com
+- **General** → hello@sipsalabs.com
 
-[sipsalabs.com](https://sipsalabs.com) · [GitHub](https://github.com/sipsalabs) · [PyPI](https://pypi.org/project/ultracompress/) · [Pricing](https://sipsalabs.com/pricing)
+[sipsalabs.com](https://sipsalabs.com) · [GitHub](https://github.com/sipsalabs) · [PyPI](https://pypi.org/project/ultracompress/) · [Pricing](https://sipsalabs.com/pricing)