Re-run the numbers

Benchmarks from Ahmed's projects, ported line for line so they run in your browser. Seed 0 reproduces the README; any other seed tests it.

Open Proof.app ▸

Don't trust the README. Re-run it.

Proof.app ports each project's benchmark from Python to JavaScript, line for line, including Python's own Mersenne Twister random number generator, so seed 0 in your browser reproduces the table in the repository exactly. Change the seed and watch whether the claim holds. Each port was checked against the real Python repository on inputs it was never tuned on.

The 8 claims you can re-run

• structstream: Raw parsing recovers 7% of realistic malformed output. One repair pass recovers 100%. Source on GitHub
• speculabench: Speculative decoding: 1.41x at 50% agreement, 2.86x at 90%. Source on GitHub
• kvsqueeze: At 40% of full cache, heavy-hitter holds 57% recall while plain recency manages 15%. Source on GitHub
• crisprradar: The 149-base demo finds one exact target and two ranked off-targets. Source on GitHub
• agentbudget: Step limits catch 0/3 stalled agent traces. Loop detection catches all 3. Source on GitHub
• toolrouter: On genuinely ambiguous queries, always routing fires the wrong tool 5/5 times. Confidence gating abstains on all 5. Source on GitHub
• phenopacketlint: Nine packets with planted defects: nine caught, and nothing fired on the clean one. Source on GitHub
• vrsbridge: Normalize the allele and four VCF records collapse into two variants. Source on GitHub