[loopx-agent]

Goal And Delivered Outcome

This PR delivers the complete empirical study, causal attribution analysis, and publication-grade vector visualizations for the EdgeBench Sol 6.1-xhigh five-arm comparative research across Portfolio, Grid, and Ad Placement tasks (todo_994c1e0336be03576fb5a15b).

Key Causal Findings

1. The "Micro-tuning Plateau" ("原地微调"): Under fixed-frequency completed-todo replanning (Explore-Off), the agent easily passes loose hypothesis novelty checks by making trivial parameter tweaks (e.g., timeout limits 28s -> 27s), leading to a 9.5-hour plateau on Grid Turing Robot with only a 1.7% cost improvement.
2. Breakthroughs Require Evidence-Driven Paradigm Shifts: True performance leaps consistently stem from evidence-conditioned hypothesis shifts (e.g., switching from Python simulation to C++ DP accelerator in Grid; aligning daily VaR timing in Portfolio).
3. Incumbent Retention Guards Against Degradation: Exploring in late hours frequently introduces degraded candidates; strict candidate acceptance gates in Explore-On and Stride prevented workspace contamination and preserved the winning solution in Ad Placement.
4. Hierarchical Stride Optimization: Replacing rigid todo clocks with effective-turn evidence conditioning (evidence_conditioned_v1) improves the valid jump ratio from 11.4% to 62.5% and cuts protocol tax from 23.9% down to 12.4%.

Deliverables

• benchmark/edgebench/studies/five-arm-sol61-causal-attribution/:
  • README.md: Bilingual technical report with rigorous empirical methodology, data tables, and actionable engineering prescriptions for next-gen agent harnesses.
  • data/summary_metrics.json: Structured empirical results across 5 arms x 3 tasks.
  • data/replan_causal_breakdown.json: Causal taxonomy and per-arm breakdown of 86 replans.
  • generate_charts.py: Standalone, zero-external-dependency script rendering 3 vector SVGs:
    • assets/five_arm_trajectory_comparison.svg
    • assets/causal_action_attribution.svg
    • assets/replan_efficiency_frontier.svg
• benchmark/tests/test_five_arm_causal_attribution.py: Automated pytest suite verifying schema integrity, mathematical invariants, and chart generation.

Verification

• uv run --extra test python -m pytest benchmark/tests/test_five_arm_causal_attribution.py (2 passed)
• uv run ruff check passed cleanly
• git diff --check passed cleanly
• Public/private boundary check: no credentials, local paths, or unredacted raw traces exposed