Your agent has rules.
Which ones have no test?
RuleCard turns an agent's written rules into tests. Every test keeps its rule number. The report starts with the rules nobody checks.
What it is
OverviewA spec compiler with a scorecard attached. You give it the rules an agent should follow. It writes a test suite from those rules. The scorecard reports spec coverage and spec conformance, not a pass rate.
The code name is agent-proving-ground. RuleCard is the public name. It is hosted with sign-in and has no billing. The public page at rulecard.gauravg.dev walks through one example run, with no sign-in.
Tests from intent,
not from behaviour.
01 · The ideaMost agent tests are written after watching what the agent does. That checks the agent against itself. RuleCard writes tests from what the owner said the agent must do, so a gap in the rules shows up as a gap in the tests.
Every test carries the number of the rule it came from. A score with no rule number is treated as a bug. The first thing the report shows is the list of rules with no test.
One open risk. The method needs a written spec with numbered rules to exist already. Getting a clear spec out of a team is the real work, and I have not shown it can be done cheaply.
One recorded run,
read with its caveats.
02 · Example runThe run below used a hiring-screen agent's own written rules: 44 rules, 135 scenarios, 5 samples each, run on 2026-09-26. The agent under test was Jev, a rule-choice model standing in for the real executor, which is not wired. Read every number with the note beside it.
| Measure | Result | Read it with this |
|---|---|---|
| Rules with no test | 0 of 44 | Self-reported. A spec can raise this by writing easy rules. |
| Rules checked on their own terms | 9 of 44 | The other 35 were only checked for wrong escalation. |
| Should hand off: did it? | 7 of 9 | Missed 2. Small sample of 12 scenarios. |
| Should not hand off: did it? | 3 of 3 | No false alarms. |
| Judged checks | Refused | The judge and the agent share a model family. On the two control agents they ran. |
The judgment log has 68 entries across 3 runs: one main run and two control agents. It does not prove the agent is safe. Nothing is proven yet.
Where it
stands.
03 · StatusThe kill rule is set in advance. If the tests catch nothing that a ten-minute manual review would not, or if handoff decisions need labelling after the fact, I publish what I learned and stop.