At which rung did your agent verify?
"Verified" gets used for two different claims at once: "the tests pass" and "I ran the real thing and watched it work," "I checked it" and "someone independent checked it." rung pulls those apart into two axes and makes the answer checkable.
| author | fresh-blind | cross-lab | |
|---|---|---|---|
| rung 2 tests green | generic CI | ||
| rung 3 drove the real surface | runtime-verification tools | ||
| rung 4 drove + S0/S1 differential | ← real and independent |
RUNG (0 to 4): how real the verification was, from reading the code up to driving the running surface and capturing a baseline vs candidate differential. CONTEXT: who checked it, from the producer to an independent reviewer at a different lab. The right-hand column is where almost nothing lives today, and the cell rung is built to name and reward.
- rung the shared vocabulary, a portable
JSON evidence-bundle schema, a declarative policy, and a single-file,
stdlib-only deterministic gate.
pip install rung-ai. - devloop produces the evidence: an AI development loop (spec, plan, review, implement, verify) that ends in a verification of the running change. rung is the vocabulary and gate for saying how real that verification was.
rung is an interchange format, not a stack: any producer that emits an
evidence-bundle/v1 and any judge that attests to one composes the same way.
One project on the judging side of that chain:
- syncade performs independent judgment: blind, cross-judge review into one ship/no-ship verdict, the cross-lab independence the gate can check for but never perform.
pip install rung-ai
rung gate bundle.json # exit 0 = pass, 30 = block, 2 = cannot-evaluate
rung run --rung 3 --surface cli -- mytool --checkStdlib-only, dependency-free, Python 3.9+. Apache 2.0.