Autonomy needs evidence.

Leaving the default 3-run sample intact with the threshold set to 60 seconds, click 'Audit three runs'. The page parses three runs: r-001 ($5.00 revenue − $2.00 cost = $3.00 margin), r-002 ($0.00 revenue − $1.50 cost = -$1.50 margin, flagged for wallet action), and r-003 ($1.00 revenue − $0.50 cost = $0.50 margin, running for 95 seconds). The metrics display total tokens (3,000), total cost ($4.00), total revenue ($6.00), net margin ($2.00), return on cost (50.0%), and average duration (41.67 s). Run r-003 is flagged as stuck because 95 seconds exceeds the 60-second limit.

The script uses Decimal.js to avoid floating-point drift when performing financial reconciliations. For each record, margin is evaluated as token_revenue_usd minus token_cost_usd. Across the dataset, return on cost is computed as the total aggregate margin divided by total token cost, yielding 50.0% on the default sample ($2.00 net margin on $4.00 total cost).

An execution job is designated as stuck if and only if its status string equals 'running' and its duration_seconds value is greater than the numerical threshold input (defaulting to 60 s). Sensitive operations are isolated through the wallet_action boolean flag, allowing operators to filter high-privilege runs independently of overall token profitability.

The three illustrative runs contain fifteen hundred, one thousand and five hundred tokens when input and output are added. Total is three thousand. The tool sums supplied row values, not a provider usage ledger. It accepts successful, failed and running statuses with unique run IDs and requires wallet action to be a real boolean. Numeric fields must be finite and nonnegative. No platform, wallet or chain is connected. The sample costs sum to four dollars and revenues to six, yielding two dollars net margin. Return on cost is two divided by four, or fifty percent; zero cost produces null instead of division by zero. The failed wallet-marked run has margin minus one point five, even though aggregate margin is positive. Wallet filtering changes visible rows, not audited totals. These economics use declared amounts rather than actual token pricing or financial transfers. A row is stuck only when status is running and duration is strictly greater than the positive threshold. The ninety five second running sample exceeds sixty and is flagged; exactly sixty would pass. A completed run over sixty is not flagged by that rule. Average sample duration is one hundred twenty five divided by three seconds. Exports preserve audited rows and flags, but their fixed three-row messages can misdescribe another input size. Local library readiness is required and no audit is claimed if unavailable.