Public repo and README

The first screen

A README a stranger can act on answers four questions before any prose:

1. What is being built: a tensor-lowering path from linalg to RISC-V Vector 1.0, compiled and measured under QEMU.
2. What is measured: median GFLOP/s at 512 cubed, on rv64,v=true,vlen=256,elen=64, with the spread beside it.
3. How to reproduce it: make bench.
4. What it ran on: the commit hash and the pinned versions in VERSIONS.md.

Write those four lines on day one, before there are numbers. “No results yet; make bench prints them” is a true statement about the repository. A placeholder figure is not.

Numbers come from files

The results table is generated from the output of make bench, and the README links to the raw file it came from, named with the commit that produced it. Retyping a table breaks the link between claim and artifact, and the reader cannot tell a stale digit from a fresh one.

State the rig in the same paragraph as the numbers. Every figure here is an emulated figure: it supports comparisons between kernels in this repository, not claims about any RISC-V core. A reader who assumes otherwise will over-read your first speedup, and that is a failure of the README, not of the reader.

Licence and layout

Put a licence in LICENSE, keep build outputs out of the tree with .gitignore, and keep dated notes in notes/. The record of what you tried is the part that makes the repository worth reading in month nine.

Things that break

The failure is a README whose numbers drift from the code. Notice it the way a stranger would: clone into a fresh directory, run the single command, and compare against the table. If that takes a step you did not write down, the command is incomplete, and the fix belongs in the Makefile rather than in a longer README.