Sync AgentWeave benchmark from GitHub 0bd1097

README.md

@@ -0,0 +1,42 @@
+---
+license: apache-2.0
+language:
+- en
+tags:
+- agents
+- function-calling
+- tool-use
+- tool-routing
+- benchmark
+- llm
+pretty_name: AgentWeave Tool-Routing Benchmark
+task_categories:
+- text-generation
+---
+
+# AgentWeave Tool-Routing Benchmark
+
+Benchmark cases and evaluation artifacts for **AgentWeave**, a deterministic pre-inference routing layer for tool-rich language-model systems.
+
+## Contents
+
+- `data/benchmark_cases.jsonl` — normalized benchmark prompts for Hugging Face Dataset Viewer.
+- `raw/evaluation/` — frozen AgentWeave evaluation artifacts from the GitHub repository.
+- `docs/` — BFCL result summaries and paper notes.
+
+## Paper
+
+**AgentWeave: Routing Before Reasoning for Efficient Function Calling in Tool-Rich Language Models**  
+arXiv: https://arxiv.org/abs/2608.23078
+
+## Source code
+
+GitHub: https://github.com/sauravsingla/agentweave
+
+## Notes
+
+This Hub repository publishes benchmark/evaluation data rather than model weights. The AgentWeave runtime and implementation remain in the canonical GitHub repository.
+
+## License
+
+Apache-2.0, matching the source repository.

data/benchmark_cases.jsonl

@@ -0,0 +1,40 @@
+{"id": 1, "query": "research and summarize evidence about a technical topic", "dataset": "AgentWeave Routing Benchmark v1", "description": "Domain-agnostic multi-capability requirements for routing and team-formation evaluation."}
+{"id": 2, "query": "analyze and verify a technical claim using multiple sources", "dataset": "AgentWeave Routing Benchmark v1", "description": "Domain-agnostic multi-capability requirements for routing and team-formation evaluation."}
+{"id": 3, "query": "plan and optimize an engineering workflow", "dataset": "AgentWeave Routing Benchmark v1", "description": "Domain-agnostic multi-capability requirements for routing and team-formation evaluation."}
+{"id": 4, "query": "retrieve research then analyze and summarize the findings", "dataset": "AgentWeave Routing Benchmark v1", "description": "Domain-agnostic multi-capability requirements for routing and team-formation evaluation."}
+{"id": 5, "query": "code analyze and verify a solution", "dataset": "AgentWeave Routing Benchmark v1", "description": "Domain-agnostic multi-capability requirements for routing and team-formation evaluation."}
+{"id": 6, "query": "research plan analyze and summarize a complex requirement", "dataset": "AgentWeave Routing Benchmark v1", "description": "Domain-agnostic multi-capability requirements for routing and team-formation evaluation."}
+{"id": 7, "query": "retrieve evidence and perform independent verification", "dataset": "AgentWeave Routing Benchmark v1", "description": "Domain-agnostic multi-capability requirements for routing and team-formation evaluation."}
+{"id": 8, "query": "design code and validate an implementation plan", "dataset": "AgentWeave Routing Benchmark v1", "description": "Domain-agnostic multi-capability requirements for routing and team-formation evaluation."}
+{"id": 9, "query": "analyze a dataset and summarize the main findings", "dataset": "AgentWeave Routing Benchmark v1", "description": "Domain-agnostic multi-capability requirements for routing and team-formation evaluation."}
+{"id": 10, "query": "plan research and verify the resulting recommendations", "dataset": "AgentWeave Routing Benchmark v1", "description": "Domain-agnostic multi-capability requirements for routing and team-formation evaluation."}
+{"id": 11, "query": "inspect code and identify correctness risks", "dataset": "AgentWeave Routing Benchmark v1", "description": "Domain-agnostic multi-capability requirements for routing and team-formation evaluation."}
+{"id": 12, "query": "research alternatives and create a concise comparison", "dataset": "AgentWeave Routing Benchmark v1", "description": "Domain-agnostic multi-capability requirements for routing and team-formation evaluation."}
+{"id": 13, "query": "retrieve technical facts and verify consistency", "dataset": "AgentWeave Routing Benchmark v1", "description": "Domain-agnostic multi-capability requirements for routing and team-formation evaluation."}
+{"id": 14, "query": "analyze an architecture and propose an implementation plan", "dataset": "AgentWeave Routing Benchmark v1", "description": "Domain-agnostic multi-capability requirements for routing and team-formation evaluation."}
+{"id": 15, "query": "summarize multiple analyses and resolve contradictions", "dataset": "AgentWeave Routing Benchmark v1", "description": "Domain-agnostic multi-capability requirements for routing and team-formation evaluation."}
+{"id": 16, "query": "plan a system and verify its operational constraints", "dataset": "AgentWeave Routing Benchmark v1", "description": "Domain-agnostic multi-capability requirements for routing and team-formation evaluation."}
+{"id": 17, "query": "research a topic then code a small validated prototype", "dataset": "AgentWeave Routing Benchmark v1", "description": "Domain-agnostic multi-capability requirements for routing and team-formation evaluation."}
+{"id": 18, "query": "analyze evidence and produce a structured summary", "dataset": "AgentWeave Routing Benchmark v1", "description": "Domain-agnostic multi-capability requirements for routing and team-formation evaluation."}
+{"id": 19, "query": "retrieve information, plan next steps, and verify assumptions", "dataset": "AgentWeave Routing Benchmark v1", "description": "Domain-agnostic multi-capability requirements for routing and team-formation evaluation."}
+{"id": 20, "query": "code a solution then analyze and summarize its tradeoffs", "dataset": "AgentWeave Routing Benchmark v1", "description": "Domain-agnostic multi-capability requirements for routing and team-formation evaluation."}
+{"id": 21, "query": "research, analyze, verify and summarize a complex technical question", "dataset": "AgentWeave Routing Benchmark v1", "description": "Domain-agnostic multi-capability requirements for routing and team-formation evaluation."}
+{"id": 22, "query": "plan, code and verify a robust implementation", "dataset": "AgentWeave Routing Benchmark v1", "description": "Domain-agnostic multi-capability requirements for routing and team-formation evaluation."}
+{"id": 23, "query": "retrieve, analyze and validate evidence from independent sources", "dataset": "AgentWeave Routing Benchmark v1", "description": "Domain-agnostic multi-capability requirements for routing and team-formation evaluation."}
+{"id": 24, "query": "summarize findings while checking for contradictory claims", "dataset": "AgentWeave Routing Benchmark v1", "description": "Domain-agnostic multi-capability requirements for routing and team-formation evaluation."}
+{"id": 25, "query": "analyze a system for reliability and produce a remediation plan", "dataset": "AgentWeave Routing Benchmark v1", "description": "Domain-agnostic multi-capability requirements for routing and team-formation evaluation."}
+{"id": 26, "query": "research implementation options and verify scalability assumptions", "dataset": "AgentWeave Routing Benchmark v1", "description": "Domain-agnostic multi-capability requirements for routing and team-formation evaluation."}
+{"id": 27, "query": "retrieve evidence, analyze risks and summarize recommendations", "dataset": "AgentWeave Routing Benchmark v1", "description": "Domain-agnostic multi-capability requirements for routing and team-formation evaluation."}
+{"id": 28, "query": "plan an experiment, analyze expected outcomes and verify constraints", "dataset": "AgentWeave Routing Benchmark v1", "description": "Domain-agnostic multi-capability requirements for routing and team-formation evaluation."}
+{"id": 29, "query": "code and verify a data-processing workflow", "dataset": "AgentWeave Routing Benchmark v1", "description": "Domain-agnostic multi-capability requirements for routing and team-formation evaluation."}
+{"id": 30, "query": "research a protocol, analyze compatibility and summarize gaps", "dataset": "AgentWeave Routing Benchmark v1", "description": "Domain-agnostic multi-capability requirements for routing and team-formation evaluation."}
+{"id": 31, "query": "analyze a graph problem and plan an efficient solution", "dataset": "AgentWeave Routing Benchmark v1", "description": "Domain-agnostic multi-capability requirements for routing and team-formation evaluation."}
+{"id": 32, "query": "retrieve domain evidence and verify a proposed design", "dataset": "AgentWeave Routing Benchmark v1", "description": "Domain-agnostic multi-capability requirements for routing and team-formation evaluation."}
+{"id": 33, "query": "summarize a complex requirement and plan implementation steps", "dataset": "AgentWeave Routing Benchmark v1", "description": "Domain-agnostic multi-capability requirements for routing and team-formation evaluation."}
+{"id": 34, "query": "research multiple options, analyze costs and verify conclusions", "dataset": "AgentWeave Routing Benchmark v1", "description": "Domain-agnostic multi-capability requirements for routing and team-formation evaluation."}
+{"id": 35, "query": "code a prototype, analyze performance and summarize results", "dataset": "AgentWeave Routing Benchmark v1", "description": "Domain-agnostic multi-capability requirements for routing and team-formation evaluation."}
+{"id": 36, "query": "retrieve documentation, analyze behavior and verify compliance", "dataset": "AgentWeave Routing Benchmark v1", "description": "Domain-agnostic multi-capability requirements for routing and team-formation evaluation."}
+{"id": 37, "query": "plan a migration and analyze reliability risks", "dataset": "AgentWeave Routing Benchmark v1", "description": "Domain-agnostic multi-capability requirements for routing and team-formation evaluation."}
+{"id": 38, "query": "research, retrieve and verify supporting evidence", "dataset": "AgentWeave Routing Benchmark v1", "description": "Domain-agnostic multi-capability requirements for routing and team-formation evaluation."}
+{"id": 39, "query": "analyze competing technical claims and resolve disagreement", "dataset": "AgentWeave Routing Benchmark v1", "description": "Domain-agnostic multi-capability requirements for routing and team-formation evaluation."}
+{"id": 40, "query": "summarize evidence, verify citations and plan follow-up work", "dataset": "AgentWeave Routing Benchmark v1", "description": "Domain-agnostic multi-capability requirements for routing and team-formation evaluation."}

docs/BFCL_V5_RESULTS.md

@@ -0,0 +1,28 @@
+# 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`](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.

docs/BFCL_V6_RESULTS.md

@@ -0,0 +1,55 @@
+# Frozen BFCL Routing-Pressure V6 Replication
+
+This page records the first successful scored run of **BFCL routing-pressure v6**, a larger untouched replication of the frozen v5 pilot.
+
+## Headline result
+
+On **48 fresh BFCL V4 `multiple` tasks with zero overlap with v5**, AgentWeave was the only strategy to achieve any native BFCL successes:
+
+| Strategy | Native BFCL success | Mean tools shown | Input tokens | Mean latency |
+|---|---:|---:|---:|---:|
+| Single-agent / all tools | 0/48 = **0.0%** | 16.00 | 124,821 | 55.92 s |
+| Random top-8 | 0/48 = **0.0%** | 8.00 | 69,790 | 30.56 s |
+| Semantic top-8 | 0/48 = **0.0%** | 8.00 | 70,271 | 35.87 s |
+| **AgentWeave** | **6/48 = 12.5%** | **4.77** | **47,805** | **27.43 s** |
+
+AgentWeave's paired native-success difference versus each matched baseline was **+12.5 percentage points**, with a **10,000-resample paired bootstrap 95% CI of +4.17 to +22.92 pp** and **exact McNemar p = 0.03125**.
+
+Against the all-tools baseline, AgentWeave used approximately **70.18% fewer visible tools**, **61.70% fewer input tokens**, and **50.95% lower mean local-model latency**.
+
+## Replication relationship to v5
+
+The frozen v5 pilot used 12 tasks and produced AgentWeave **2/12 = 16.67%** versus **0/12** for all three matched baselines, but with McNemar **p = 0.5** because the sample was small.
+
+V6 kept the v5 model, routing budgets, baselines, 16-tool pressure, generation settings, and native BFCL evaluation unchanged. It used a new content-blind sample and explicitly excluded all 12 frozen v5 task IDs. The resulting **6/48 = 12.5% vs 0/48** outcome therefore serves as a larger untouched replication of the directional v5 finding.
+
+## Routing diagnostics
+
+| Strategy | Mean original-candidate recall | All original candidates retained |
+|---|---:|---:|
+| Single-agent | 100.00% | 100.00% |
+| Random top-8 | 48.44% | 10.42% |
+| Semantic top-8 | **86.81%** | 66.67% |
+| AgentWeave | 76.91% | 54.17% |
+
+Semantic top-8 retained more original BFCL candidates than AgentWeave, yet still scored 0/48. AgentWeave achieved 6/48 while exposing fewer tools, so the native-success result cannot be explained by candidate retention alone.
+
+## Reproducibility freeze
+
+- Study: `bfcl-routing-pressure-v6`
+- Scored head: `ca6ff084da4fe5c670421b99e7ad413650e60c33`
+- Merged commit: `e8c5fd92a568002dfb1213ab5c0b2dd08f0e339b`
+- BFCL/Gorilla commit: `6ea57973c7a6097fd7c5915698c54c17c5b1b6c8`
+- Workflow run: `31983991285`
+- Artifact: `9275082744`
+- Artifact digest: `sha256:db9bcca4e7f0d83b6b837b23a0ddf8a5bcd43b24a5d66554156ba794dca48c0e`
+- Model: `MadeAgents/Hammer2.1-1.5b`
+- External API spend: **$0**
+
+The machine-readable frozen record is [`evaluation/bfcl-routing-pressure-v6-frozen.json`](evaluation/bfcl-routing-pressure-v6-frozen.json). The earlier frozen pilot is documented in [`BFCL_V5_RESULTS.md`](BFCL_V5_RESULTS.md).
+
+## Evidence boundary
+
+This is a **BFCL-derived routing-pressure replication**, not an official full BFCL leaderboard score. The experiment preserves untouched BFCL questions and native BFCL evaluation while deterministically augmenting the model-visible tool context. It is keyless, local, single-turn, and uses $0 external API spend.
+
+The v6 result is frozen. Any tuning or further replication must use a new study ID and a fresh untouched sample.

docs/PAPER.md

@@ -0,0 +1,26 @@
+# AgentWeave Paper
+
+## AgentWeave: Routing Before Reasoning for Efficient Function Calling in Tool-Rich Language Models
+
+**arXiv:** 2608.23078  
+**Paper:** https://arxiv.org/abs/2608.23078
+
+AgentWeave studies deterministic pre-inference routing for tool-rich language-model agents. The paper evaluates whether reducing the model-visible candidate action space before inference can improve function-calling reliability while leaving the downstream model unchanged.
+
+### Citation
+
+**AgentWeave: Routing Before Reasoning for Efficient Function Calling in Tool-Rich Language Models.** arXiv:2608.23078 (2026).
+
+### Software artifact
+
+**Evolving AgentWeave project — Concept DOI:** https://doi.org/10.5281/zenodo.22913459  
+
+**AgentWeave v0.7.0 — Version DOI:** https://doi.org/10.5281/zenodo.22913460  
+**Zenodo:** https://zenodo.org/records/22913460  
+**GitHub release:** https://github.com/sauravsingla/agentweave/releases/tag/v0.7.0
+
+Use the Concept DOI when referring to AgentWeave as an evolving software project. Use the version DOI when citing or reproducing the archived v0.7.0 software release. The arXiv identifier remains the citation for the research paper.
+
+### Research boundary
+
+The paper reports a frozen BFCL-derived routing-pressure study. It is not an official full BFCL leaderboard submission. The repository keeps frozen evaluation evidence separate from later framework and experimental extensions.

raw/evaluation/bfcl-native-live-v1.json

@@ -0,0 +1,66 @@
+{
+  "study_id": "bfcl-routing-pressure-v6",
+  "status": "preregistered-before-first-score",
+  "supersedes_study": "bfcl-routing-pressure-v5",
+  "supersession_reason": "V5 is frozen as a positive pilot: AgentWeave achieved 2/12 native BFCL successes while all three matched baselines achieved 0/12, together with substantial tool, token, and latency reductions, but the success difference was not statistically significant. V6 is a larger untouched replication with the same model, routing budgets, baselines, tool pressure, and generation settings. Only the sample and distractor seeds change, and all 12 scored V5 task ids are explicitly excluded.",
+  "frozen_v5_evidence": "evaluation/bfcl-routing-pressure-v5-frozen.json",
+  "benchmark": {
+    "name": "Berkeley Function Calling Leaderboard V4 routing-pressure replication",
+    "repository": "ShishirPatil/gorilla",
+    "commit": "6ea57973c7a6097fd7c5915698c54c17c5b1b6c8",
+    "category": "multiple",
+    "native_metric": "BFCL single-turn AST/function-call validity against the restored original BFCL evaluator data",
+    "sample_size": 48,
+    "augmented_tool_count": 16,
+    "sample_rule": "Exclude the 12 frozen V5 task ids, then take the 48 remaining task ids with the smallest SHA256('agentweave-bfcl-routing-pressure-v6:' + id) digest from BFCL_v4_multiple.json. No task-content or outcome filtering is permitted.",
+    "augmentation_rule": "For each sampled task, retain every original BFCL candidate function and append unique distractor functions from the same pinned BFCL_v4_multiple.json, ordered by SHA256('agentweave-bfcl-distractors-v6:' + task_id + ':' + function_name), until 16 model-visible functions are present. Restore original BFCL data before native evaluation."
+  },
+  "inference": {
+    "mode": "self-hosted CPU",
+    "model": "MadeAgents/Hammer2.1-1.5b",
+    "model_source": "Hugging Face public repository; automatically downloaded by workflow",
+    "temperature": 0,
+    "max_new_tokens": 128,
+    "bfcl_transport_handler": "MiniCPM-SALA-FC registry handler used only as an OpenAI-compatible BFCL transport/decoder shim; no external model API is called",
+    "external_api_spend_usd": 0.0
+  },
+  "strategies": {
+    "single-agent": {"description": "Same local model receives all 16 model-visible candidate functions."},
+    "random-router": {"description": "Deterministic content-independent top-8 control using the frozen routing implementation from V5.", "top_k_tools": 8},
+    "semantic-router": {"description": "Zero-shot MiniLM cosine-similarity top-8 router.", "embedding_model": "sentence-transformers/all-MiniLM-L6-v2", "top_k_tools": 8},
+    "agentweave": {"description": "MiniLM scores tool-provider groups and AgentWeave GlobalTeamOptimizer selects up to four provider agents, capped at six tools.", "embedding_model": "sentence-transformers/all-MiniLM-L6-v2", "max_provider_agents": 4, "max_tools": 6}
+  },
+  "primary_hypotheses": [
+    "H1: The V5 directional native-success advantage replicates: AgentWeave native BFCL task success exceeds the all-tools single-agent baseline on the fresh V6 sample.",
+    "H2: AgentWeave native BFCL task success exceeds the deterministic random top-8 baseline on the fresh V6 sample.",
+    "H3: AgentWeave native BFCL task success exceeds the semantic top-8 baseline on the fresh V6 sample.",
+    "H4: AgentWeave materially reduces model-visible tools, input tokens, and local inference latency relative to the all-tools baseline while preserving useful original-candidate recall."
+  ],
+  "statistics": {
+    "success_interval": "Wilson 95% confidence interval",
+    "paired_success_test": "exact McNemar test",
+    "paired_effect_interval": "10,000-resample paired bootstrap 95% confidence interval",
+    "alpha": 0.05,
+    "power_note": "n=48 is a larger confirmatory replication than the n=12 V5 pilot, but still a resource-constrained study. Statistical significance and confidence intervals are reported without post-score tuning."
+  },
+  "measurements": [
+    "BFCL native per-task success",
+    "original BFCL candidate recall after routing",
+    "all-original-candidates-retained rate",
+    "model-visible tools",
+    "local wall-clock model latency",
+    "input/output tokens",
+    "external API spend ($0)",
+    "all BFCL failure/error records",
+    "explicit V5/V6 task overlap count"
+  ],
+  "anti_tuning_rules": [
+    "The frozen V5 result is not altered or rescored.",
+    "All 12 frozen V5 scored task ids are excluded before V6 sampling, guaranteeing zero V5/V6 task overlap.",
+    "V6 keeps the V5 model, 16-tool pressure, four-agent/six-tool AgentWeave budget, all three baselines, temperature, and max generation length unchanged.",
+    "V6 uses a fresh content-blind sample seed and fresh content-independent BFCL distractor seed.",
+    "Original BFCL data is restored before evaluation so augmentation changes model-visible context, not evaluator ground truth.",
+    "Any modification after the first V6 score requires a new study id and a new untouched sample."
+  ],
+  "evidence_boundary": "This is a BFCL-derived routing-pressure replication, not an official full BFCL leaderboard score. It preserves untouched BFCL questions and native BFCL evaluation but deterministically augments model-visible tool context. V6 is keyless, local, single-turn, has $0 external API spend, and excludes every task used in the frozen V5 pilot."
+}

raw/evaluation/bfcl-routing-pressure-v5-frozen.json

@@ -0,0 +1,32 @@
+{
+  "study_id": "bfcl-routing-pressure-v5",
+  "status": "frozen-after-first-successful-score",
+  "scored_head_sha": "e73374a2ca80f6d0ac4ed387ac50497915ebde9d",
+  "workflow_run_id": 31981730530,
+  "artifact_id": 9273004440,
+  "artifact_digest": "sha256:d10c78ccf1d8575d9b242fbb2dd4d0d358af2cda90c40c0e764cd08a0676b759",
+  "benchmark_commit": "6ea57973c7a6097fd7c5915698c54c17c5b1b6c8",
+  "model": "MadeAgents/Hammer2.1-1.5b",
+  "sample_size": 12,
+  "augmented_tool_count": 16,
+  "agentweave_budget": {"max_provider_agents": 4, "max_tools": 6},
+  "results": {
+    "single-agent": {"successes": 0, "n": 12, "native_task_success": 0.0, "mean_original_candidate_recall": 1.0, "mean_tools_after": 16.0, "input_tokens": 30482, "mean_call_latency_seconds": 56.04385813716666},
+    "random-router": {"successes": 0, "n": 12, "native_task_success": 0.0, "mean_original_candidate_recall": 0.46527777777777773, "mean_tools_after": 8.0, "input_tokens": 16623, "mean_call_latency_seconds": 30.22037881383335},
+    "semantic-router": {"successes": 0, "n": 12, "native_task_success": 0.0, "mean_original_candidate_recall": 0.8472222222222222, "mean_tools_after": 8.0, "input_tokens": 17194, "mean_call_latency_seconds": 36.071879301249986},
+    "agentweave": {"successes": 2, "n": 12, "native_task_success": 0.16666666666666666, "wilson_95_ci": [0.04696514218385381, 0.4480308622529734], "mean_original_candidate_recall": 0.7708333333333334, "all_original_candidates_retained_rate": 0.6666666666666666, "mean_tools_after": 4.583333333333333, "input_tokens": 11189, "mean_call_latency_seconds": 25.261402219500003}
+  },
+  "comparisons": {
+    "agentweave_vs_single_agent": {"difference_pp": 16.666666666666664, "paired_bootstrap_95_ci_pp": [0.0, 41.66666666666667], "exact_mcnemar_p": 0.5},
+    "agentweave_vs_random_router": {"difference_pp": 16.666666666666664, "paired_bootstrap_95_ci_pp": [0.0, 41.66666666666667], "exact_mcnemar_p": 0.5},
+    "agentweave_vs_semantic_router": {"difference_pp": 16.666666666666664, "paired_bootstrap_95_ci_pp": [0.0, 41.66666666666667], "exact_mcnemar_p": 0.5}
+  },
+  "hypothesis_outcomes": {
+    "H1_candidate_recall_at_least_85_percent": "not_supported: 77.08%",
+    "H2_efficiency_vs_all_tools": "supported: 71.35% fewer visible tools, 63.29% fewer input tokens, 54.92% lower mean model latency",
+    "H3_native_success_no_worse_and_may_improve": "supported directionally: 16.67% vs 0%; pilot is not statistically significant (McNemar p=0.5)",
+    "H4_vs_random_routing": "supported on native success and candidate recall, with different tool budgets; interpret as pilot evidence"
+  },
+  "freeze_rule": "Do not alter or overwrite this v5 result. Any tuning, larger replication, model change, routing-budget change, sample change, or distractor change must use a new study id and fresh untouched sample.",
+  "evidence_boundary": "BFCL-derived routing-pressure stress test, not an official full BFCL leaderboard score. Untouched BFCL questions and native BFCL evaluation are retained, while model-visible tool context is deterministically augmented. Keyless, local, single-turn, $0 external API spend."
+}

raw/evaluation/bfcl-routing-pressure-v6-frozen.json

@@ -0,0 +1,34 @@
+{
+  "study_id": "bfcl-routing-pressure-v6",
+  "status": "frozen-after-first-successful-score",
+  "scored_head_sha": "ca6ff084da4fe5c670421b99e7ad413650e60c33",
+  "merged_commit_sha": "e8c5fd92a568002dfb1213ab5c0b2dd08f0e339b",
+  "workflow_run_id": 31983991285,
+  "artifact_id": 9275082744,
+  "artifact_digest": "sha256:db9bcca4e7f0d83b6b837b23a0ddf8a5bcd43b24a5d66554156ba794dca48c0e",
+  "benchmark_commit": "6ea57973c7a6097fd7c5915698c54c17c5b1b6c8",
+  "model": "MadeAgents/Hammer2.1-1.5b",
+  "sample_size": 48,
+  "v5_overlap": 0,
+  "augmented_tool_count": 16,
+  "agentweave_budget": {"max_provider_agents": 4, "max_tools": 6},
+  "results": {
+    "single-agent": {"successes": 0, "n": 48, "native_task_success": 0.0, "mean_original_candidate_recall": 1.0, "mean_tools_after": 16.0, "input_tokens": 124821, "mean_call_latency_seconds": 55.92130121310414},
+    "random-router": {"successes": 0, "n": 48, "native_task_success": 0.0, "mean_original_candidate_recall": 0.484375, "mean_tools_after": 8.0, "input_tokens": 69790, "mean_call_latency_seconds": 30.55668982414585},
+    "semantic-router": {"successes": 0, "n": 48, "native_task_success": 0.0, "mean_original_candidate_recall": 0.8680555555555555, "mean_tools_after": 8.0, "input_tokens": 70271, "mean_call_latency_seconds": 35.86756529608332},
+    "agentweave": {"successes": 6, "n": 48, "native_task_success": 0.125, "wilson_95_ci": [0.058570514385719794, 0.24700458286386834], "mean_original_candidate_recall": 0.7690972222222222, "all_original_candidates_retained_rate": 0.5416666666666666, "mean_tools_after": 4.770833333333333, "input_tokens": 47805, "mean_call_latency_seconds": 27.43061615410416}
+  },
+  "comparisons": {
+    "agentweave_vs_single_agent": {"difference_pp": 12.5, "paired_bootstrap_95_ci_pp": [4.166666666666666, 22.916666666666664], "exact_mcnemar_p": 0.03125},
+    "agentweave_vs_random_router": {"difference_pp": 12.5, "paired_bootstrap_95_ci_pp": [4.166666666666666, 22.916666666666664], "exact_mcnemar_p": 0.03125},
+    "agentweave_vs_semantic_router": {"difference_pp": 12.5, "paired_bootstrap_95_ci_pp": [4.166666666666666, 22.916666666666664], "exact_mcnemar_p": 0.03125}
+  },
+  "efficiency_vs_all_tools": {
+    "visible_tools_reduction_percent": 70.18229166666667,
+    "input_tokens_reduction_percent": 61.70115605547144,
+    "mean_latency_reduction_percent": 50.94782210168547
+  },
+  "interpretation": "V6 replicated the V5 directional native-success advantage on 48 fresh BFCL multiple tasks with zero V5 overlap. AgentWeave was the only strategy with native successes (6/48) and the paired exact McNemar p-value versus each matched baseline was 0.03125. This is statistically significant at alpha=0.05 for this preregistered BFCL-derived routing-pressure study, but it is not an official full BFCL leaderboard score.",
+  "freeze_rule": "Do not alter or overwrite this v6 result. Any tuning, model change, routing-budget change, sample change, distractor change, or further replication must use a new study id and a fresh untouched sample.",
+  "evidence_boundary": "BFCL-derived routing-pressure replication, not an official full BFCL leaderboard score. Untouched BFCL questions and native BFCL evaluation are retained, while model-visible tool context is deterministically augmented. Keyless, local, single-turn, $0 external API spend, and all frozen V5 tasks are excluded."
+}

raw/evaluation/hybrid-selection-issue38-v1.json

@@ -0,0 +1,107 @@
+{
+  "aggregate": {
+    "all_tools_dynamic": {
+      "delegation_calls": 0,
+      "malicious_candidates_exposed": 1,
+      "mean_candidate_reduction": 0.09111111111111109,
+      "mean_final_visible_tools": 13.444444444444445,
+      "mean_initial_visible_tools": 13.444444444444445,
+      "model_calls": 11,
+      "provenance_complete_rate": 1.0,
+      "recovery_success_rate": 0.1111111111111111,
+      "routing_calls": 0,
+      "scenarios": 9,
+      "search_calls": 0,
+      "task_success_rate": 1.0,
+      "token_proxy": 16170,
+      "total_latency_ms_proxy": 330.0,
+      "unauthorized_executions": 0,
+      "unauthorized_selection_attempts": 1,
+      "unnecessary_delegations": 0,
+      "unnecessary_routings": 0,
+      "unnecessary_searches": 0
+    },
+    "hierarchical_delegation": {
+      "delegation_calls": 11,
+      "malicious_candidates_exposed": 1,
+      "mean_candidate_reduction": 0.5444444444444445,
+      "mean_final_visible_tools": 5.0,
+      "mean_initial_visible_tools": 5.0,
+      "model_calls": 22,
+      "provenance_complete_rate": 1.0,
+      "recovery_success_rate": 0.1111111111111111,
+      "routing_calls": 0,
+      "scenarios": 9,
+      "search_calls": 0,
+      "task_success_rate": 1.0,
+      "token_proxy": 13860,
+      "total_latency_ms_proxy": 644.0,
+      "unauthorized_executions": 0,
+      "unauthorized_selection_attempts": 1,
+      "unnecessary_delegations": 2,
+      "unnecessary_routings": 0,
+      "unnecessary_searches": 0
+    },
+    "hybrid_bounded_dynamic": {
+      "delegation_calls": 0,
+      "malicious_candidates_exposed": 0,
+      "mean_candidate_reduction": 0.8177777777777778,
+      "mean_final_visible_tools": 2.0,
+      "mean_initial_visible_tools": 1.5555555555555556,
+      "model_calls": 12,
+      "provenance_complete_rate": 1.0,
+      "recovery_success_rate": 0.3333333333333333,
+      "routing_calls": 13,
+      "scenarios": 9,
+      "search_calls": 4,
+      "task_success_rate": 1.0,
+      "token_proxy": 3780,
+      "total_latency_ms_proxy": 388.0,
+      "unauthorized_executions": 0,
+      "unauthorized_selection_attempts": 1,
+      "unnecessary_delegations": 0,
+      "unnecessary_routings": 0,
+      "unnecessary_searches": 1
+    },
+    "pre_inference_routing": {
+      "delegation_calls": 0,
+      "malicious_candidates_exposed": 0,
+      "mean_candidate_reduction": 0.8622222222222222,
+      "mean_final_visible_tools": 1.5555555555555556,
+      "mean_initial_visible_tools": 1.5555555555555556,
+      "model_calls": 10,
+      "provenance_complete_rate": 1.0,
+      "recovery_success_rate": 0.0,
+      "routing_calls": 9,
+      "scenarios": 9,
+      "search_calls": 0,
+      "task_success_rate": 0.6666666666666666,
+      "token_proxy": 2450,
+      "total_latency_ms_proxy": 308.0,
+      "unauthorized_executions": 0,
+      "unauthorized_selection_attempts": 1,
+      "unnecessary_delegations": 0,
+      "unnecessary_routings": 0,
+      "unnecessary_searches": 0
+    }
+  },
+  "evidence_boundary": {
+    "frozen_historical_results_modified": false,
+    "latency_values": "modeled proxy",
+    "production_claim": false,
+    "selection_policy": "deterministic controlled model proxy",
+    "token_values": "modeled proxy"
+  },
+  "protocol": "issue-38-controlled-hybrid-selection-v1",
+  "scenarios": [
+    {"catalog_size": 10, "emergent": false, "malicious_permitted": false, "name": "obvious_request", "route_budget": 2, "routing_miss": false, "tool_failure": false, "unauthorized_injection": false, "unnecessary_search": false},
+    {"catalog_size": 10, "emergent": false, "malicious_permitted": false, "name": "equivalent_specialists", "route_budget": 2, "routing_miss": false, "tool_failure": false, "unauthorized_injection": false, "unnecessary_search": false},
+    {"catalog_size": 10, "emergent": true, "malicious_permitted": false, "name": "emergent_capability", "route_budget": 1, "routing_miss": false, "tool_failure": false, "unauthorized_injection": false, "unnecessary_search": false},
+    {"catalog_size": 10, "emergent": false, "malicious_permitted": false, "name": "routing_miss", "route_budget": 1, "routing_miss": true, "tool_failure": false, "unauthorized_injection": false, "unnecessary_search": false},
+    {"catalog_size": 10, "emergent": false, "malicious_permitted": false, "name": "tool_failure", "route_budget": 1, "routing_miss": false, "tool_failure": true, "unauthorized_injection": false, "unnecessary_search": false},
+    {"catalog_size": 50, "emergent": false, "malicious_permitted": false, "name": "noisy_catalog", "route_budget": 2, "routing_miss": false, "tool_failure": false, "unauthorized_injection": false, "unnecessary_search": false},
+    {"catalog_size": 10, "emergent": false, "malicious_permitted": true, "name": "malicious_permitted", "route_budget": 1, "routing_miss": false, "tool_failure": false, "unauthorized_injection": false, "unnecessary_search": false},
+    {"catalog_size": 10, "emergent": false, "malicious_permitted": false, "name": "unauthorized_injection", "route_budget": 2, "routing_miss": false, "tool_failure": false, "unauthorized_injection": true, "unnecessary_search": false},
+    {"catalog_size": 10, "emergent": false, "malicious_permitted": false, "name": "unnecessary_search", "route_budget": 2, "routing_miss": false, "tool_failure": false, "unauthorized_injection": false, "unnecessary_search": true}
+  ]
+}

raw/evaluation/paper-quality-v1.json

@@ -0,0 +1,51 @@
+{
+  "protocol_version": 1,
+  "status": "preregistered-before-confirmatory-outcome-run",
+  "frozen_research_commit": "b502fbf803fc906ba4971324efd46bc6cfb91ab7",
+  "confirmatory_scope": "controlled executable team outcome benchmark",
+  "confirmatory_design": {
+    "seeds": 30,
+    "workloads_per_seed": 12,
+    "strategies": [
+      "agentweave-team",
+      "single-best-agent",
+      "random-team",
+      "capability-only-team"
+    ],
+    "primary_outcome": "task completion",
+    "secondary_outcomes": ["quality", "cost", "latency", "recovery success"],
+    "hypotheses": [
+      {
+        "id": "H1",
+        "statement": "AgentWeave optimized-team task completion exceeds every baseline by at least 20 percentage points across the preregistered repeated-seed workload runs."
+      },
+      {
+        "id": "H2",
+        "statement": "AgentWeave optimized-team mean delivered quality exceeds every baseline by at least 0.10 on the same repeated-seed workload runs."
+      },
+      {
+        "id": "H3",
+        "statement": "When an induced failure is actually exercised, AgentWeave recovery succeeds in at least 90% of opportunities."
+      }
+    ],
+    "statistics": {
+      "confidence_level": 0.95,
+      "completion_interval": "Wilson score interval",
+      "paired_effect_interval": "paired bootstrap, 10000 resamples, seed 20260816",
+      "paired_significance": "exact McNemar test for binary completion when paired observations are available"
+    }
+  },
+  "exploratory_external_routing_scope": {
+    "benchmark": "General-AgentBench frozen untouched evaluation",
+    "benchmark_commit": "35f5c027c31ddcb3366b28674c6cb2957460c0e2",
+    "purpose": "post-hoc stronger semantic-router comparison on the already-frozen untouched task set; this is explicitly exploratory and is not represented as a preregistered untouched comparison",
+    "semantic_baselines": [
+      "sentence-transformers/all-MiniLM-L6-v2",
+      "BAAI/bge-small-en-v1.5"
+    ],
+    "baseline_design": "zero-shot nearest family-description embedding; no General-AgentBench label or example is used to fit the embedding baseline",
+    "statistics": ["Wilson 95% confidence intervals", "paired bootstrap accuracy-difference interval", "exact McNemar test"],
+    "failure_reporting": "store all AgentWeave errors and disagreements with each semantic baseline"
+  },
+  "evidence_boundary": "The confirmatory outcome benchmark physically executes controlled handlers over a synthetic agent catalog; it is an actual task-completion outcome within that benchmark, not a production-user or external-provider result. The semantic-router comparison is routing accuracy only and is exploratory/post-hoc because General-AgentBench had already been observed before these embedding baselines were added."
+}

raw/evaluation/router-v2-holdout.json

@@ -0,0 +1,58 @@
+{
+  "protocol_version": 1,
+  "status": "preregistered-before-first-scored-run",
+  "router_name": "PrototypeFamilyRouterV2",
+  "router_source": "research/router_v2.py",
+  "development_source": {
+    "name": "General-AgentBench",
+    "repository": "cxcscmu/General-AgentBench",
+    "commit": "35f5c027c31ddcb3366b28674c6cb2957460c0e2",
+    "use": "development/training only",
+    "families": ["mathhay", "mcpbench", "search", "swebench", "tau2bench", "terminalbench"]
+  },
+  "external_holdout": {
+    "sampling": "first 24 published tasks from each source in repository order; 72 tasks total",
+    "label_visibility": "source/family labels are withheld until after prediction",
+    "sources": [
+      {
+        "family": "mathhay",
+        "name": "GSM8K",
+        "repository": "openai/grade-school-math",
+        "commit": "3101c7d5072418e28b9008a6636bde82a006892c",
+        "path": "grade_school_math/data/test.jsonl",
+        "text_field": "question",
+        "tasks": 24
+      },
+      {
+        "family": "swebench",
+        "name": "HumanEval",
+        "repository": "openai/human-eval",
+        "commit": "6d43fb980f9fee3c892a914eda09951f772ad10d",
+        "path": "data/HumanEval.jsonl.gz",
+        "text_field": "prompt",
+        "tasks": 24
+      },
+      {
+        "family": "terminalbench",
+        "name": "InterCode NL2Bash",
+        "repository": "princeton-nlp/intercode",
+        "commit": "c3e46d827cfc9d4c704ec078f7abf9f41e3191d8",
+        "path": "data/nl2bash/test_queries.json",
+        "text_field": "query",
+        "tasks": 24
+      }
+    ]
+  },
+  "primary_metric": "Router V2 family Hit@1 on the 72-task external holdout",
+  "secondary_metrics": [
+    "legacy frozen-router Hit@1 on the same holdout",
+    "Router V2 minus legacy absolute accuracy delta",
+    "macro family accuracy",
+    "per-family accuracy",
+    "mean Router V2 confidence",
+    "confusion matrix",
+    "5-fold deterministic development cross-validation on General-AgentBench"
+  ],
+  "anti_tuning_rule": "After the first successful scored holdout run, the Router V2 algorithm, family seeds, weighting, temperature, sampling rule, source pins, and prediction mapping must not be changed to improve this holdout score. Any later model must be versioned Router V3 and evaluated on a new holdout.",
+  "evidence_boundary": "This is a balanced three-source family-routing transfer test. It measures task-family routing only, not native GSM8K answer accuracy, HumanEval pass@k, or InterCode environment success."
+}

raw/evaluation/router-v3-holdout.json

@@ -0,0 +1,50 @@
+{
+  "protocol_version": 1,
+  "status": "frozen-before-first-run",
+  "router_v3_base_commit": "c940596cd0a9e4168cc301c32d47de7e40308d2e",
+  "development_data": {
+    "name": "General-AgentBench published baseline prompts",
+    "repository": "cxcscmu/General-AgentBench",
+    "commit": "35f5c027c31ddcb3366b28674c6cb2957460c0e2",
+    "usage": "Router V3 may fit on these development prompts only; Router V2's 72-task holdout is not reused for scoring. Its observed failure mode may motivate benchmark-independent semantic anchors."
+  },
+  "external_holdout": [
+    {
+      "name": "MBPP",
+      "repository": "google-research/google-research",
+      "commit": "589e977488f21a336a3d3da9b96da91ddbcf935e",
+      "path": "mbpp/mbpp.jsonl",
+      "family": "swebench",
+      "count": 24
+    },
+    {
+      "name": "TruthfulQA",
+      "repository": "sylinrl/TruthfulQA",
+      "commit": "d71c110897f5d31c5d7f309e7bc316c152f6f031",
+      "path": "TruthfulQA.csv",
+      "family": "search",
+      "count": 24
+    },
+    {
+      "name": "OSWorld",
+      "repository": "xlang-ai/OSWorld",
+      "commit": "091f5ef1d5544bc74953c77875d5feb5bed30108",
+      "path": "evaluation_examples/examples",
+      "family": "terminalbench",
+      "count": 24
+    }
+  ],
+  "sampling_rule": "For each source, collect valid task prompts, compute SHA256(source_id + '\\n' + prompt), sort ascending by digest, and select the first 24. The rule is fixed before the first score.",
+  "label_visibility": "Source family labels are attached only after Router V2 and Router V3 predictions are produced.",
+  "primary_metric": "Router V3 family Hit@1 over the balanced 72-task external holdout",
+  "secondary_metrics": [
+    "Router V2 Hit@1 on the identical new holdout",
+    "absolute V3-V2 Hit@1 delta",
+    "macro family accuracy",
+    "per-family accuracy",
+    "mean confidence",
+    "confusion matrix"
+  ],
+  "anti_tuning_rule": "After the first successful scored run, Router V3 algorithm, anchors, weights, source pins, sampling rule, and selected task count are frozen for this holdout. Any changed router must be Router V4 with a new holdout.",
+  "evidence_boundary": "This is a task-family routing transfer test only. It does not claim native MBPP code correctness, TruthfulQA truthfulness accuracy, or OSWorld environment success."
+}

raw/evaluation/router-v4-holdout.json

@@ -0,0 +1,53 @@
+{
+  "protocol_version": 1,
+  "status": "frozen-before-first-successful-score",
+  "router_v4_base_commit": "ca03b23c8485b87e7b5b9a45591a574909f3d432",
+  "development_data": {
+    "name": "General-AgentBench published baseline prompts",
+    "repository": "cxcscmu/General-AgentBench",
+    "commit": "35f5c027c31ddcb3366b28674c6cb2957460c0e2",
+    "usage": "Router V4 may fit on these development prompts only. Router V2 and Router V3 external holdouts are not reused for V4 scoring."
+  },
+  "external_holdout": [
+    {
+      "name": "CRUXEval",
+      "repository": "facebookresearch/cruxeval",
+      "commit": "190faf16d175b5847b0af05d937872b1fb395942",
+      "path": "data/cruxeval.jsonl",
+      "family": "swebench",
+      "count": 24,
+      "representation": "code plus input wrapped as a code-reasoning request"
+    },
+    {
+      "name": "WorkArena",
+      "repository": "ServiceNow/WorkArena",
+      "commit": "a772230a94cf1caf4166b8ead3983f3b3786455b",
+      "path": "src/browsergym/workarena/tasks",
+      "family": "terminalbench",
+      "count": 24,
+      "representation": "published user-facing goal templates extracted from goal-producing task code"
+    },
+    {
+      "name": "BrowserGym MiniWoB",
+      "repository": "ServiceNow/BrowserGym",
+      "commit": "9e779f087de9a65668b6974d11f9ce9816026e96",
+      "path": "browsergym/experiments/src/browsergym/experiments/benchmark/metadata/miniwob.csv",
+      "family": "terminalbench",
+      "count": 24,
+      "representation": "held-out MiniWoB benchmark task names converted to task requests"
+    }
+  ],
+  "sampling_rule": "For each source, collect valid prompts, compute SHA256(source_id + '\\n' + prompt), sort ascending by digest, and select the first requested count. The rule is fixed before the first successful score.",
+  "label_visibility": "Source family labels are attached only after Router V3 and Router V4 predictions are produced.",
+  "primary_metric": "Router V4 family Hit@1 over all 72 external holdout tasks",
+  "secondary_metrics": [
+    "Router V3 Hit@1 on the identical new holdout",
+    "absolute V4-V3 Hit@1 delta",
+    "macro family accuracy",
+    "per-source accuracy",
+    "mean confidence",
+    "confusion matrix"
+  ],
+  "anti_tuning_rule": "After the first successful scored run, Router V4 algorithm, hierarchy, weights, source pins, sampling rule, and selected task counts are frozen for this holdout. Any changed router must become Router V5 with a new holdout.",
+  "evidence_boundary": "This is task-family routing transfer only. It does not claim CRUXEval program-execution correctness, WorkArena task success, or MiniWoB browser-environment success."
+}

raw/evaluation/router-v5-holdout.json

@@ -0,0 +1,46 @@
+{
+  "protocol_version": 1,
+  "status": "frozen-before-first-successful-score",
+  "router_v5_base_commit": "075d1431709a4e4d6c26762a7d832c633050c670",
+  "development_data": {
+    "name": "General-AgentBench published baseline prompts",
+    "repository": "cxcscmu/General-AgentBench",
+    "commit": "35f5c027c31ddcb3366b28674c6cb2957460c0e2",
+    "usage": "Router V5 may fit on these development prompts only. No Router V2, V3, or V4 external holdout task is reused for V5 scoring."
+  },
+  "external_holdout": {
+    "name": "VisualWebArena",
+    "repository": "web-arena-x/visualwebarena",
+    "commit": "89f5af29305c3d1e9f97ce4421462060a70c9a03",
+    "family": "terminalbench",
+    "domains": [
+      {
+        "name": "classifieds",
+        "path": "config_files/vwa/test_classifieds.raw.json",
+        "count": 24
+      },
+      {
+        "name": "reddit",
+        "path": "config_files/vwa/test_reddit.raw.json",
+        "count": 24
+      },
+      {
+        "name": "shopping",
+        "path": "config_files/vwa/test_shopping.raw.json",
+        "count": 24
+      }
+    ]
+  },
+  "sampling_rule": "For each VisualWebArena domain, load published task intents, compute SHA256(domain + '\\n' + task_id + '\\n' + intent), sort ascending by digest, and select the first 24. The rule is fixed before the first successful score.",
+  "label_visibility": "The terminalbench family label is attached only after Router V4 and Router V5 predictions are produced.",
+  "primary_metric": "Interactive-family Hit@1 over the balanced 72-task VisualWebArena holdout",
+  "secondary_metrics": [
+    "Router V4 Hit@1 on the identical new holdout",
+    "absolute V5-V4 Hit@1 delta",
+    "per-domain accuracy",
+    "mean confidence",
+    "confusion matrix"
+  ],
+  "anti_tuning_rule": "After the first successful scored run, Router V5 algorithm, weights, source pin, sampling rule, and task counts are frozen for this holdout. Any changed router must become Router V6 with a new holdout.",
+  "evidence_boundary": "This is interactive task-family routing transfer only. It does not claim native VisualWebArena environment success, action execution correctness, or visual grounding performance."
+}

raw/evaluation/router-v6-holdout.json

@@ -0,0 +1,31 @@
+{
+  "protocol_version": 1,
+  "status": "frozen-before-first-successful-score",
+  "router_v6_base_commit": "b36e9fddc5f32aa89d31ebb4c43a9b0acff8ee1d",
+  "development_data": {
+    "name": "General-AgentBench published baseline prompts",
+    "repository": "cxcscmu/General-AgentBench",
+    "commit": "35f5c027c31ddcb3366b28674c6cb2957460c0e2",
+    "usage": "Router V6 may fit on these development prompts only. No Router V2-V5 external holdout task is reused for V6 scoring."
+  },
+  "external_holdout": {
+    "name": "WebArena",
+    "repository": "web-arena-x/visualwebarena",
+    "commit": "89f5af29305c3d1e9f97ce4421462060a70c9a03",
+    "path": "config_files/wa/test_webarena.raw.json",
+    "family": "terminalbench",
+    "count": 72,
+    "representation": "published WebArena natural-language intents"
+  },
+  "sampling_rule": "Load published WebArena task intents, compute SHA256(task_id + '\\n' + intent), sort ascending by digest, and select the first 72. The rule is fixed before the first successful score.",
+  "label_visibility": "The terminalbench family label is attached only after Router V5 and Router V6 predictions are produced.",
+  "primary_metric": "Interactive-family Hit@1 over the 72-task WebArena holdout",
+  "secondary_metrics": [
+    "Router V5 Hit@1 on the identical new holdout",
+    "absolute V6-V5 Hit@1 delta",
+    "mean confidence",
+    "confusion matrix"
+  ],
+  "anti_tuning_rule": "After the first successful scored run, Router V6 algorithm, weights, source pin, sampling rule, and task count are frozen for this holdout. Any changed router must become Router V7 with a new holdout.",
+  "evidence_boundary": "This is task-family routing transfer only. It does not claim native WebArena environment success, browser action correctness, or end-to-end task completion."
+}

raw/evaluation/router-v7-holdout.json

@@ -0,0 +1,35 @@
+{
+  "protocol_version": 1,
+  "status": "frozen-before-first-successful-score",
+  "router_v7_base_commit": "c8dc39f1d78a12e92a2e18a2ecd8dd059edbdb0d",
+  "development_data": {
+    "name": "General-AgentBench published baseline prompts",
+    "repository": "cxcscmu/General-AgentBench",
+    "commit": "35f5c027c31ddcb3366b28674c6cb2957460c0e2",
+    "usage": "Router V7 may fit on these development prompts only. No Router V2-V6 external holdout task is reused for V7 scoring."
+  },
+  "external_holdout": {
+    "name": "AssistantBench",
+    "repository": "AssistantBench/AssistantBench",
+    "host": "Hugging Face Datasets",
+    "commit": "482cbbc0400f6d048438c4021727f21a10cbff49",
+    "files": [
+      "assistant_bench_v1.0_dev.jsonl",
+      "assistant_bench_v1.0_test.jsonl"
+    ],
+    "family": "search",
+    "count": 72,
+    "representation": "published realistic information-seeking task text"
+  },
+  "sampling_rule": "Load both pinned AssistantBench JSONL files, deduplicate by id, compute SHA256(id + '\n' + task), sort ascending by digest, and select the first 72. The rule is fixed before the first successful score.",
+  "label_visibility": "The search family label is attached only after Router V6 and Router V7 predictions are produced.",
+  "primary_metric": "Search-family Hit@1 over the 72-task AssistantBench holdout",
+  "secondary_metrics": [
+    "Router V6 Hit@1 on the identical new holdout",
+    "absolute V7-V6 Hit@1 delta",
+    "mean confidence",
+    "confusion matrix"
+  ],
+  "anti_tuning_rule": "After the first successful scored run, Router V7 algorithm, weights, source pin, sampling rule, and task count are frozen for this holdout. Any changed router must become Router V8 with a new holdout.",
+  "evidence_boundary": "This is task-family routing transfer only. It does not claim AssistantBench answer accuracy, web browsing correctness, source quality, or end-to-end research completion."
+}

raw/evaluation/untouched-generalization.json

@@ -0,0 +1,33 @@
+{
+  "protocol_version": 1,
+  "status": "frozen-before-first-run",
+  "frozen_router_commit": "06b5381b1521c3db34944ac9ab2b8cb07be438f8",
+  "frozen_scope": "agentweave/",
+  "external_benchmark": {
+    "name": "General-AgentBench",
+    "repository": "cxcscmu/General-AgentBench",
+    "commit": "35f5c027c31ddcb3366b28674c6cb2957460c0e2"
+  },
+  "task_source": "published baseline trace user prompts only",
+  "canonical_trace_directories": {
+    "mathhay": "mathhay_Claude-Haiku-4.5_distraction_all",
+    "mcpbench": "mcpbench_Claude-Haiku-4.5_distraction_all",
+    "search": "search_Claude-Haiku-4.5_distraction_all",
+    "swebench": "swebench_Claude-Haiku-4.5_distraction_100",
+    "tau2bench": "tau2bench_Claude-Haiku-4.5_distraction_all",
+    "terminalbench": "terminalbench_Claude-Haiku-4.5_distraction_100"
+  },
+  "label_visibility": "benchmark labels are withheld until after requirement analysis and family prediction",
+  "prediction_rule_order": [
+    "mcpbench: mcp or tool-use capability",
+    "terminalbench: operating-system or shell capability",
+    "tau2bench: compliance capability/domain",
+    "swebench: coding, backend, frontend, database, or sql capability",
+    "search: research, retrieval, or knowledge-graph capability",
+    "mathhay: otherwise"
+  ],
+  "primary_metric": "family Hit@1",
+  "secondary_metrics": ["macro family accuracy", "per-family accuracy", "mean analyzer confidence", "confusion matrix"],
+  "anti_tuning_rule": "The benchmark fails if any file under agentweave/ differs from frozen_router_commit. Results must be published without changing the frozen router for this evaluation.",
+  "evidence_boundary": "This is an untouched external family-routing generalization test over published General-AgentBench prompts. It is not the native General-AgentBench end-to-end task score."
+}