The capability curve and the reliability curve are not the same line
A model that clears a capability benchmark has only proven it can produce a correct answer under favorable conditions. It has not proven it can do so repeatedly, cheaply, and without accumulating loss. KellyBench makes the gap concrete: in its agentic coding evaluation, every frontier model posted an average deficit, and of 24 model-and-setting combinations only 3 avoided outright bankruptcy. Even the strongest performer, Opus, reached a sophistication score of just 32.6%. Passing a test says nothing about surviving a quarter of production traffic.
What an economic evaluation exposes that a pass rate hides
A single-shot pass rate rewards the best case. An economic evaluation scores cumulative profit and loss across many attempts, so every retry, wrong turn, and abandoned task is charged against the ledger. That framing surfaces the failure that a capability benchmark conceals: an agent that is right 70% of the time can still run a net loss if the remaining 30% burns tokens, triggers rollbacks, and consumes human review. Reliability is an economic property, not a peak-accuracy property.
The threat side grows on the same slope
Capability compounds on both sides of the fence. The UK AI Safety Institute estimates that frontier cyber-attack capability is doubling roughly every four months. An operator who treats a passed benchmark as a deployment approval is scaling exposure at the same rate as capability, without a matching gain in the reliability that would keep that exposure contained.
An A4-length practical guide: from planning to operations
Start planning by writing targets as numbers, not adjectives. Define the task success rate, the p95 latency, the cost ceiling per completed task, and an explicit violation budget such as zero PII leaks and zero unauthorized writes. Borrow KellyBench's logic and add a cumulative-P&L target: net non-negative return over a rolling window of 200 tasks. A goal you cannot express as a threshold is a goal you cannot gate a release on.
Next, enumerate failure patterns before you enumerate features. The recurring modes are the runaway retry loop that drains the cost budget, silent tool-call failures that the agent papers over, hallucinated state that diverges from the real system, and confident wrong output on the long tail. Map each mode to an explicit recovery branch: bounded retry with exponential backoff and a hard cap of three attempts; escalation to human confirmation when confidence falls below threshold; a safe-reduction path that returns a narrower but verified answer; and a stop condition that halts and rolls back rather than compounding an error.
Make the operational checklist a gate, not a suggestion. Every run emits a standard structured log — task ID, model version, token cost, latency, tool-call outcomes, and the final decision branch taken. Mask PII at the logging boundary so traces are safe to store and review. Track the violation count on a live dashboard and treat any nonzero value as a release blocker. Without a uniform log schema you cannot compute a pass rate, and without a pass rate you cannot argue that the agent is safe to ship.
Close the loop with continuous improvement driven by production data. Feed real failure traces back into the evaluation set weekly so the benchmark tracks the workload instead of drifting from it. Re-run the economic evaluation on every model or prompt change and reject any candidate whose cumulative P&L regresses, even if its single-shot pass rate improved. Given the four-month doubling of adversarial capability, schedule the security-facing portion of the suite to re-run at least monthly.
Finally, keep the human review budget explicit and bounded. Escalation is a valid recovery branch only if a person is actually available within the latency target; otherwise the safe-reduction or stop branch must take over automatically. An escalation path with no capacity behind it is an outage waiting to happen.
Executive summary
Clearing a capability benchmark is not authorization to deploy. KellyBench's result — every frontier model at an average deficit, 3 of 24 combinations avoiding bankruptcy, Opus at 32.6% sophistication — shows that reliability is an economic property measured over many runs, not a peak score. Gate releases on cumulative P&L, explicit failure-to-recovery branches, standardized PII-masked logs, and a zero-violation budget, and re-run the economic and security evaluations on a fixed cadence against a threat curve that doubles every four months.