# Hardware Inference Stack Ledger

> **Status**: Stage 3 (SWEEP) — 8 parallel research-agent sweeps in flight

## Purpose

Forensic survey of frontier inference-stack capability claims (open
frameworks + closed vendors + cloud silicon) on 2024–2026 corpus. Six
closure audits:

1. Benchmark-versus-real-workload audit (TTFT, tokens/sec at contended load)
2. Cost-per-token transparency
3. Frontier-model fidelity audit (quantized inference quality)
4. Batch-vs-streaming behavior
5. Commercial availability vs research-preview gap
6. Closed-vendor benchmark cherry-pick audit

## Empty-space hypothesis (predeclared before sweeps)

- **Bill 5 ★** — Cross-vendor benchmark stability (≤10% TTFT variance). Predicted empty.
- **Bill 8 ★** — Quantization fidelity (INT4/FP4 retains ≥95% FP16 capability). Predicted empty.
- **Bill 11 ★** — Universal inference-platform coverage (one framework runs Llama 4 + DeepSeek V3 + Qwen 3 + Mistral Large 2 at ≤20% variance). Predicted empty.

## Bridge tests

Hardware Inference is the **19th ledger** and provides the **purest
possible test of the rescoped B7 commercialization-vs-research axis**:
vLLM / SGLang / llama.cpp / MLX (open) vs Groq / Cerebras / SambaNova /
Etched (closed) creates the maximally separable cluster boundary in the
entire atlas.

## Files

- [purpose.md](./purpose.md) — threat model + bridge test specifics
- [bills_draft.md](./bills_draft.md) — 13 bills + 6 meta-costs + 3 escape gates
- [scripts/bill_classifier.py](./scripts/bill_classifier.py) — regex + arbitration rules
- `deep_loops/` — raw sweep JSONs (1301–1308)
- `wiki/` — Obsidian-compatible bill / paper / concept indices (post-LOCK)