Frozen BFCL Routing-Pressure V5 Result
AgentWeave's BFCL routing-pressure v5 study is frozen after its first successful scored run. The experiment uses untouched BFCL V4 multiple questions, deterministic BFCL-derived distractors to create 16-tool routing pressure, native BFCL evaluation, and the public MadeAgents/Hammer2.1-1.5b model running locally with no API key and $0 external API spend.
| Strategy | Native BFCL success | Original-candidate recall | Mean tools shown | Input tokens | Mean model latency |
|---|---|---|---|---|---|
| Single-agent / all tools | 0/12 (0%) | 100.0% | 16.00 | 30,482 | 56.04 s |
| Random top-8 | 0/12 (0%) | 46.5% | 8.00 | 16,623 | 30.22 s |
| Semantic top-8 | 0/12 (0%) | 84.7% | 8.00 | 17,194 | 36.07 s |
| AgentWeave | 2/12 (16.67%) | 77.1% | 4.58 | 11,189 | 25.26 s |
Against each matched baseline, AgentWeave's native-success difference is +16.67 percentage points. The paired bootstrap 95% interval is [0.0, +41.67] pp and exact McNemar p = 0.5. With only 12 tasks, this is a resource-constrained pilot and is not statistically significant.
Relative to the all-tools baseline, AgentWeave used 71.35% fewer model-visible tools, 63.29% fewer input tokens, and 54.92% lower mean model latency, while being the only strategy to solve any task in this frozen sample.
The preregistered >=85% original-candidate-recall target was not met: AgentWeave reached 77.08%. This negative hypothesis outcome is retained alongside the positive native-success and efficiency results.
Freeze and provenance
- Study:
bfcl-routing-pressure-v5 - Scored head:
e73374a2ca80f6d0ac4ed387ac50497915ebde9d - Workflow run:
31981730530 - Evidence artifact:
9273004440 - Artifact digest:
sha256:d10c78ccf1d8575d9b242fbb2dd4d0d358af2cda90c40c0e764cd08a0676b759 - Pinned Gorilla/BFCL commit:
6ea57973c7a6097fd7c5915698c54c17c5b1b6c8 - Frozen machine-readable record:
evaluation/bfcl-routing-pressure-v5-frozen.json
Evidence boundary: this is a BFCL-derived routing-pressure stress test, not an official full BFCL leaderboard score. It preserves untouched BFCL questions and native BFCL evaluation while deterministically augmenting model-visible tool context. Any future tuning or replication must use a new study id and a fresh untouched sample.