---
license: cc-by-4.0
pretty_name: Galahad — SHA-256 Proof (measured benchmark results)
tags:
  - benchmark
  - kv-cache
  - llm-inference
  - reproducibility
configs:
  - config_name: default
    data_files: index.jsonl
---

# Galahad — SHA-256 Proof

Timestamped, independently verifiable **measured** benchmark results behind the Galahad paper.

📄 **Paper:** https://huggingface.co/papers/2607.23806
📚 **arXiv:** https://arxiv.org/abs/2607.23806
🧪 **Live testbench (Space):** https://huggingface.co/spaces/Corbenic/galahad-bench
💾 **Proof mirror (GitHub):** https://github.com/corbenicai/galahad-bench/tree/main/proof

## What this is

This dataset is the **integrity proof** for the paper's claims: the raw measured outputs (pure numbers —
token counts, latency in ms, VRAM in MB, accuracy, pass/fail) plus a full SHA-256 manifest. It contains
**no source code, no engine internals, and no solver contents** — only measurements.

- **`results/`** — the raw measured benchmark JSON files.
- **`index.jsonl`** — one row per result file (filename, its SHA-256, key numeric measurements) so the
  dataset viewer renders a quick overview.
- **`results.sha256`** — the hashes in `sha256sum -c` format.
- **`SHA256_PROOF.txt`** — the full manifest: verifiable hashes **plus sealed commitments** to files kept
  private (source, verified-solver memory, internal notes, engine binaries). A sealed hash proves a file
  existed and was fixed on the stated date **without revealing its contents**.

## Verify it yourself

```bash
git clone https://huggingface.co/datasets/Corbenic/galahad-bench-proof
cd galahad-bench-proof
sha256sum -c results.sha256          # expect: every line "OK"
```

## Highlights (measured, single GPU)

- **Movable window** — `results/window6m.json`: a **6,000,000-token** context on one GPU at **flat VRAM**
  (base ≈ peak MB), correct retrieval at every depth, passing negative control.
- **Head-to-head** — `results/bench_vllm.json`, `results/bench_sglang.json`: same GPU / model class; vLLM
  hard-stops at ~30k tokens, SGLang truncates past ~32k, Galahad runs to millions.
- **Verified reuse** — `results/reuse_*.json`, `results/frommem_*.json`: 0-decode-token reuse across four
  model families; `frommem_EMPTY.json` is the negative control (empty memory solves nothing).

## Scope & honesty

These are **measurements**, not the method. The engine, the verified-solver memory, and the internal
method notes are intentionally **not** published — their SHA-256 commitments are in `SHA256_PROOF.txt` so
the record is complete and tamper-evident. No trade-secret content is exposed by a hash.

`load_dataset("Corbenic/galahad-bench-proof")` loads the `index.jsonl` overview; the full result files live
under `results/`.

## Cite

```
arxiv.org/abs/2607.23806
```