pooria/benchmark-latency-budget

Latency Budget

Quality that only counts if it arrives before the deadline.

What it measures

Whether an agent gives the right answer before its deadline (voice 447-641 ms, UI 875-1494 ms, background 2025-3714 ms), each set at 1.5-2x the shortest correct path, with virtual per-tool latencies.

Category: agents. Status: alpha.

How to run

$ npx benchmark-latency-budget list
$ npx benchmark-latency-budget run --model <model-id> --provider openrouter
$ npx benchmark-latency-budget score results.json

--provider mock runs the whole benchmark offline with no keys. Real models run through OpenRouter (OPENROUTER_API_KEY) or any OpenAI-compatible base URL (OPENAI_BASE_URL, OPENAI_API_KEY).

Data fields

data/tasks.jsonl holds 36 tasks, one JSON object per line (kinds: 36 episode).

field	type
id	string
kind	string
prompt	string
expected	string
metadata	object

Provenance

Latency Budget provenance

This file is generated from the benchmark definition.

Seeded latency-budget task generator

• Kind: synthetic
• URL: https://github.com/pooriaarab/benchmark-latency-budget/blob/main/scripts/gen-tasks.mjs
• Licence: MIT
• Retrieved: 2026-10-15T00:00:00.000Z
• How to get it: Run node scripts/gen-tasks.mjs --seed 20261015 for the public tasks, or node scripts/gen-tasks.mjs --seed <n> --split hidden for a hidden split (written to gitignored artifacts/ with h- prefixed task ids).
• Contamination notes: Public tasks generated with seed 20261015, benchmark cutoff 2026-10-15. Public tasks may enter training data; use a separately seeded hidden split for final comparisons.
• Refresh: Pick a new seed, run the generator, review the JSONL diff. Regenerate quarterly or when saturation is shown.

Grader

qualityWithinBudget: a single named candidate value (bounded, case-insensitive) grounded in replayed tool output, scored 1 only when total time (measured or declared per-turn model time plus scheduled tool latencies) is within the task deadline. Reports p50/p95 time, lateButCorrect count, and a 0.5x-4x deadline curve.

Limitations

This dataset contains only the public task split (36 tasks). Hidden evaluation splits use a private seed and are deliberately not part of this dataset. Source kinds: synthetic. These are evaluation prompts for the benchmark-latency-budget benchmark, not training data; scoring requires the package's deterministic grader. See PROVENANCE.md for contamination notes.

Citation

@misc{benchmark_latency_budget,
  author = {Pooria Arab},
  title = {Latency Budget task data},
  year = {2026},
  url = {https://github.com/pooriaarab/benchmark-latency-budget}
}

Links

• npm: https://www.npmjs.com/package/benchmark-latency-budget
• GitHub: https://github.com/pooriaarab/benchmark-latency-budget
• Leaderboard: https://aibenchmarkindex.com/benchmarks/benchmark-latency-budget
• Dataset: https://huggingface.co/datasets/pooria/benchmark-latency-budget