# WSP Hub (wsphub.dev) — Sovereign Agentic & LLM Protocol Directory

> Cryptographically signed (Ed25519) P0 Config Fixes and P1 Sparse Weight-Space Patches for open-weight language models (Qwen, Llama, Gemma, DeepSeek), plus WSP Verify regression & safety attestation via Supernova MCP.
> Full specification and mathematical formulas: https://wsphub.dev/llms-full.txt

## Core Value Proposition
- **Instant Weight Surgery (<5s)**: Eliminates 16 GB full-model re-downloads for small defect fixes (e.g. 64-column MLP tweaks or generation config errors).
- **Copy-on-Write Safetensors**: Applies surgical modifications directly onto existing local checkpoints without duplicating base model storage.
- **Goodhart-Resistant Verification**: Evaluated on secret perturbed holdouts, 10,000 paired bootstrap resamples, permutation tests, and StrongREJECT/XSTest safety gates.
- **Zero-Trust Attestation**: Every patch and evaluation event is signed with Ed25519 and committed to a public SHA-256 transparency ledger.

## Quickstart CLI (Apache-2.0 Open Source)
```bash
# 1. Install CLI
pip install wsp-cli

# 2. Apply verified patch to local model or Hugging Face cache
wsp apply Qwen/Qwen3.5-0.8B --patch wsp-qwen3.5-0.8b-math-loop-exp8 --out ./qwen-patched

# 3. Bit-exactly revert back to pristine original weights
wsp revert ./qwen-patched

# 4. Empirically verify locked benchmark scores locally
wsp reproduce ./qwen-patched --patch wsp-qwen3.5-0.8b-math-loop-exp8
```

## Streamable HTTP Model Context Protocol (MCP) Server
AI coding agents (Cursor, Windsurf, Claude Code, Cline, Supernova) can connect natively:
- **Free Discovery Endpoint**: `POST https://api.wsphub.dev/api/v1/mcp`
  - Tool: `wsp.search_verified_patches` (query by model repo or defect type)
- **Metered Escrow Settlement Gateway**: `https://supernova.cool/api/v1/mcp/` (`X-Supernova-Key`)
  - `wsp.diagnose_model_defect`: $0.25 (synchronous zero-GPU triage) | $4.90 (live GPU sampling sweep)
  - `wsp.trigger_full_eval`: $9.00 (<=4B) | $29.00 (7B-14B) multi-seed locked benchmark + holdout
  - `wsp.request_custom_surgery`: $149.00 (50-pass stateful `gemini-4-argon` causal probing report + candidate `.wsp`)
  - `wsp.request_publisher_attestation`: $49.00 (official release attestation badge + hash-chained certificate)

## Verified Launch Catalog
1. `wsp-qwen3.5-0.8b-math-loop-exp8` (P1_WEIGHTS)
   - Target: `Qwen/Qwen3.5-0.8B`
   - Defect: Catastrophic 93.6% 32k-token repetition loops on complex reasoning tasks (MATH-500, GSM8K).
   - Solution: 64 delta columns across MLP layers 15 & 19 (131,072 modified parameters, 0.016% of total).
   - Result: Cures thinking loops with +48.1% MATH-500 accuracy uplift in under 4 seconds.
2. `wsp-qwen3.5-0.8b-missing-genconfig-eos-fix` (P0_CONFIG)
   - Target: `Qwen/Qwen3.5-0.8B`
   - Defect: Missing stop-token declarations causing runaway generation and infinite loops in non-vLLM runtimes.
   - Solution: Injects canonical EOS/EOT tokens and anti-loop presence penalty in `generation_config.json`.
3. `wsp-qwen2.5-coder-1.5b-rep-penalty-syntax-fix` (P0_CONFIG)
   - Target: `Qwen/Qwen2.5-Coder-1.5B-Instruct`
   - Defect: Severe JSON/brace syntax truncation triggered by aggressive default repetition penalties.
   - Solution: Neutralizes `repetition_penalty` to 1.0 with safe temperature (0.2) and top-p (0.95).

## Citations & Verification URLs
- Web Portal: https://wsphub.dev/
- Benchmark Replay Studio: https://wsphub.dev/studio
- Verification CI/CD Pricing: https://wsphub.dev/verify
- Append-Only Transparency Ledger: https://wsphub.dev/transparency
- Unabridged Technical Specification: https://wsphub.dev/llms-full.txt