docs/perf: define and measure the 10x evidence gates - #84
Conversation
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 9f8e876817
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| - candidate throughput: at least 6,349,502 reads/s, or | ||
| - candidate wall time: at most 38.8929 seconds, |
There was a problem hiding this comment.
When the required five-repeat baseline differs from the existing single-run measurement, these fixed thresholds are no longer a 10x comparison against that baseline: a candidate could pass while improving by less than 10x, or fail despite improving by more than 10x. This is especially problematic because the document requires capturing the repeated baseline before publishing the multiplier; define the gate as a candidate-to-baseline ratio using a specified repeat statistic, then derive any absolute thresholds from that frozen result.
Useful? React with 👍 / 👎.
dnncha
changed the title
1 participant
What this changes
make bench-10x-control, a deterministic indexed-Hamming control with a 256-read exhaustive-scan preflight and five-repeat CSV output;Evidence status
This PR does not claim that any 10x target has been met. The existing native/public correctness gates pass. The local synthetic control is diagnostic only and must not be substituted for the repeated full Sanson/Brunello workflow lane. Scientific and industry impact remain at zero verified external records until independent evidence exists.
Validation
make testmake cli-testmake docs-readymake build/bench_hamming_control