▣
BLACKBOX
flight recorder for AI agents

Production Insights

FROM RECORDED TRAFFIC
Pattern analysis mined from your actual recorded runs — fault classes, cost distribution, and trend signals. Every number comes from immutable production data, not benchmarks.
▣
⚠ 13/37 runs (35%) faulted. Primary fault class: C2. Fault rate is trending up.
↑ RISING
TOTAL RUNS
37
RUNS WITH FAULTS
1335.1% fault rate
COST P50 / P95
$0.0035p95: $0.0376
AVG EVENTS / RUN
8.4
FAULT CLASS BREAKDOWN
C23 occurrences
100%
run_meridian_003
COST DISTRIBUTION
Median run cost
$0.0035
95th percentile
$0.0376
Cost spread (p95/p50)
10.6×
Avg events per run
8.4
FAULT TREND (7-DAY WINDOW)
↑ RISINGFault rate is increasing. Review recent model changes or prompt changes.
NEXT ACTIONS
⚡
Run a Challenge
Test a challenger model against your best run
▲
Upgrade Test
HOLD or PROCEED verdict across all runs
◎
Leaderboard
Quality score and cost by model
MODEL DISTRIBUTION — USAGE SHARE
MODELRUNSSHAREAVG COSTFAULT RATE
claude-haiku-4-5-2025100115
40%
$0.0018833.3%
claude-sonnet-4-614
37%
$0.0129235.7%
claude-sonnet-4-6@sn-4.6-05195
13%
$0.0333260.0%
anthropic/claude-sonnet-4-63
8%
$0.010800.0%
claude-opus-4-61
3%
$0.00000100.0%
COMMON PROMPT PHRASES — TOP N-GRAMS ACROSS GENESIS EVENTS
“covenant monitoring”
13×34%
“monitoring agent”
13×34%
“agent meridian”
13×34%
“meridian capital”
13×34%
“capital test”
13×34%
“test covenants”
13×34%
“covenants §6.1”
13×34%
“§6.1 against”
13×34%
“against latest”
13×34%
“latest audited”
13×34%
Phrases appearing in ≥2 system prompts. Aggregate only — no individual run attribution.