OnCallVal design
Positioning
| Benchmark | Domain | What is scored | Missing |
|---|---|---|---|
| τ-bench / τ²-bench | airline/retail/telecom customer service | final database state + policy compliance | single-record edits; no blast radius; benign failures |
| TheAgentCompany | generic office work | task completion in a fake company | long-horizon but low-stakes; no systems state |
| SWE-bench | code repair | tests pass on a repo | it is coding, and it is static |
| LifeVal (OpenAI) | everyday life | workflow completion | proprietary, unverifiable, life domain |
| OnCallVal | infrastructure operations | final world state: cause + mitigation + side effects |
The differentiator is collateral damage as a scored dimension. In
customer-service benchmarks a wrong action touches one record. In operations a
wrong action takes down unrelated services. OnCallVal's no_collateral
predicate makes the difference between a fix and a lucky fix measurable.
Design principles
- State, not scripts. Metrics, logs, and alerts are derived from world state at query time. There is no scripted incident timeline to memorize; the agent that fixes the state sees the metrics recover. This also makes every episode replayable and the scorer environment-side.
- Declared difficulty. Faults and decoys are declared per episode with weights; tier is a pure function of weight. The tier can never disagree with the content of the task (enforced by test).
- End-state scoring. The scorer inspects only the final world and the
submitted report. It never reads the transcript, never repairs output, and
never grants partial credit on
resolved. Path efficiency is reported as diagnostics only. Two correct fixes with different routes both score 1. - Fairness through runbooks. Every service ships a runbook containing the recommended config values, the capacity formula, and known issues. Tasks never require guessing a magic value: the healthy state is always discoverable in-world.
- Budget pressure on triage. Reads cost budget. A model that dumps every service's state runs out of actions before it can fix anything. Diagnosis must be directed, which is the actual on-call skill.
Difficulty calibration
No target range is published, because none has been measured. The observed ladder is 3B at 0/72 and 7B at 2/72 on the 72-task slice. Calibration is a falsifiable rule rather than a hoped-for number:
- frontier-class agent above ~70% overall -> too easy: more decoys, more two-fault episodes;
- below ~10% -> validate the tasks before raising difficulty (is the fix discoverable in-world? does the runbook actually contain the healthy value?);
- in between -> tier monotonicity must hold (medium > hard > severe); if it breaks, the weights are wrong.
If real baselines land above target, tighten budgets before adding fault complexity; if far below, add runbook hints before simplifying faults.
Known limitations
- The fleet is small (≤7 services, ≤6 hosts). Real incidents span hundreds of services; context-window economics push back. Scale is the obvious v2 axis.
- Metrics are instantaneous derivations with light noise, not time series. Agents cannot trend a metric over an hour; they see now-values and logs.
- The action API is deliberately small (15 actions). Real remediation includes canary deploys, feature-flag rollouts, and traffic shaping.
- One defensible fix per fault. A creative but valid alternative remediation
that defeats the fault's
active()predicate also scores 1. That is by design. - Severity is simulated. Nothing here measures whether an agent communicates well with humans during an incident.
Reproducibility
data/tasks.jsonl is frozen. Regenerating with the shipped code reproduces it
byte-for-byte (seeded generator, stable SHA-256-derived seeds, no builtin
hash()). The held-out split lives under a private seed and is never shipped.