Rare failures can dominate an agent fleet loss

A failure can be uncommon by count and dominant by economic damage. In the historical perpetuals fleet analysis, 205 liquidations were 3.2% of 6,400 closed positions and carried 74% of gross loss. We read those as separate frequency and severity measures. The first tells us how often a labeled outcome occurred; the second tells us where recorded losses accumulated.

The continuous record companion publishes these aggregates for a pre-alpha research fleet frozen at an August 15, 2026 data cutoff. Most fleet fills were paper, using live prices with zero slippage and zero funding, with a maintenance-margin placeholder whose liquidation timing differed from the real venue. The economic concentration inherits those accounting and execution assumptions.

Count share and damage share measure different outcomes

Dividing 205 by 6,400 gives approximately 3.20%, consistent with the rounded headline. The remaining 6,195 positions are about 96.8% of the closed-position count. The 74% number concerns gross loss; net portfolio P&L is a separate metric. Profitable trades and the definition of loss aggregation affect the relationship between those two totals.

A useful illustrative ledger has 100 positions, three liquidations and $10,000 of gross losses. If the three liquidations account for $7,400, their average loss is roughly $2,467. The other 97 positions collectively account for $2,600 of gross loss, but some may be profitable or flat. Dividing $2,600 by all 97 yields a contribution per position. An average among losing positions needs that narrower denominator.

This small example prevents a common mistake: comparing a liquidation average with an average over every non-liquidation position while naming both averages loss severity. A valid severity comparison needs the number of losing positions in each group and the same currency and accounting convention. Reconstructing that calculation requires fields beyond the public headline.

A risk review should preserve two columns

We would make a failure table with event count, eligible position count, gross loss contribution and net contribution. A separate column would state paper or live mode and valuation assumptions. This proposed artifact lets an operator prioritize high-damage failures without mistaking rarity for safety or treating every failure as equally expensive.

The first deployment, described in our controls paper companion, ran for 21 days in a twelve-token real-capital Base market under a frozen model family. Its policy-valid settlement success of 99.9% measures a different kind of failure on a different denominator. Successful settlement can still produce a losing investment outcome.

The fleet result suggests looking at the left tail when evaluating a candidate repair, with measured severity carried alongside incidence. Its liquidation rate and loss distribution remain bounded to the historical pre-alpha corpus. A present product claim would require its own exposure window and accounting record.