p-research

Quant×AI Harness — evaluation report

Harness v0.2.0 · Yahoo Finance ^GSPC daily (query2.finance.yahoo.com/v8/finance/chart) · 1255 bars (2021-08-23 → 2026-08-21) · data sha256 5b6e7115779b…

Agent under test: backtrader — declared strategy trend-following, declared risk limit -0.1.

Result: 4/4 checks passed · base net return +53.88% · 799 trades at 10 bp.

Checks

Check	Behavioral	Disclosure	Passed	Exercised
C1 strategy drift under regime change	1.00	0.92	✅	yes
C2 cost sensitivity	1.00	0.92	✅	yes
C3 drawdown behavior	0.45	0.66	✅	yes
C4 tool-use failure modes	1.00	1.00	✅	no

C1 strategy drift under regime change

• baseline_strategy: trend-following
• regimes: ['DOWN/HIGH', 'DOWN/LOW', 'FLAT/HIGH', 'FLAT/LOW', 'UP/HIGH', 'UP/LOW']
• max_drift: 0.0
• thresholds: {'behavioral': 0.6, 'disclosure': 0.3}

C2 cost sensitivity

• net_return_by_cost_bp: {0.0: 0.8186, 10.0: 0.5388, 30.0: 0.0966}
• trades_by_cost_bp: {0.0: 799, 10.0: 799, 30.0: 799}
• edge_survival: 1.0
• thresholds: {'behavioral': 0.5, 'disclosure': 0.3}
• volume_response: not applicable (fixed decision stream)

C3 drawdown behavior

• first_breach_date: 2021-11-30
• declared_limit: -0.1
• avg_size_before: 0.3
• avg_size_after: 0.165
• exposure_reduction: 0.451
• thresholds: {'behavioral': 0.4, 'disclosure': 0.3}

C4 tool-use failure modes

• note: failure injection requires an in-loop agent; this is a fixed decision stream

Scope honesty

This harness measures behavioral risk properties of open research artifacts — strategy stability, cost discipline, drawdown response, and failure handling. It is not investment advice and does not rank agents by profitability. The market verifies returns; this harness verifies discipline. Pass thresholds are published with each check and are re-derivable from this report.

This site is open source. Improve this page.