Skip to content

Benchmarks

Caveat (read first)

All numbers below are measured on this project's own test fixtures. The harness and fixtures originate from context-management-for-antigravity (maybeanns) and were re-measured here on 2026-06-14 with 28 skills across five platforms. They are not guarantees for your production workloads.

Reproduce

python researcher/benchmarks/context-savings/benchmark_context_savings.py
python researcher/benchmarks/context-savings/benchmark_context_savings.py --markdown
python researcher/scripts/run_benchmarks.py   # CI gate wrapper

Runnable demo (subset)

python examples/demo/run_demo.py

Uses the same handoff fixture family as the benchmark; expect ~93% savings on token count for compact_session handoff mode.

Summary table (project fixtures)

Technique Before After Saved Savings %
Observation Masking 2,222 95 2,127 95.7%
Hierarchical Summarization 3,777 428 3,349 88.7%
Handoff Summary 3,777 255 3,522 93.2%
Combined Pipeline 2,936 234 2,702 92.0%
Progressive Disclosure 58,790 3,426 55,364 94.2%
Total (all techniques) 79,402 5,649 73,753 92.9%

Full table and methodology: root README benchmark section.

Validation gates

Benchmarks are one of four gates that must pass on every PR:

python researcher/scripts/validate_repo.py --strict
python researcher/scripts/skill_health.py --strict --no-history
python researcher/scripts/run_benchmarks.py
python researcher/scripts/check_activation_cases.py