ZeroRun safely reuses a previously passing deterministic test result when the reviewed source closure and pinned runtime identity are unchanged. Anything uncertain runs fresh.
Both supported internal workloads cleared the unchanged ≥5× same-runner compute-efficiency and ≥50% p95 gates. These are historical internal qualification results, not a promise that every customer workload will achieve the same speedup. External design-partner reliability, customer economics, retention and willingness to pay remain open.
Measured work. Reused carefully.
DETERMINISTIC TEST REUSE
HOW IT WORKS
Conservative reuse after every edit.
ZeroRun does not guess that a test is unaffected. Reuse is allowed only inside an explicit, reviewed boundary.
EDIT → TEST DECISION
exact identity changed / unknown
01 · REVIEW INPUTS
Declare the boundary
Choose the deterministic target and review the source closure that can affect it.
TARGETpytest · stable-python-apideclared
CLOSUREtesting/ · _pytest/python.pyreviewed
SOURCE IDready
02 · BUILD IDENTITY
Bind the execution context
ZeroRun combines source, immutable runtime, command, policy and declared environment into one identity.
RUNTIMEpython@sha256:…8d7cpinned
COMMANDpytest testing/test_python_api.pyexact
PLATFORMlinux / amd64 · network nonebound
KEYzr:6f0b…c912complete
03 · DECIDE CONSERVATIVELY
Reuse only an exact safe hit
The current identity is compared with the previously passing identity. A mismatch never reuses the result.
PREVIOUS PASSzr:6f0b…c912
CURRENT REQUESTzr:6f0b…c912
EXACT MATCHREUSE ELIGIBLE
Reuse passsame reviewed identity
Fresh executionchanged, unknown or unsupported
SUPPORTED BOUNDARY
Narrow by design.
The MVP optimizes for defensible correctness rather than broad compatibility. It is not arbitrary-command memoization.
✓Linux / amd64Current supported host and container platform.
✓Immutable pinned OCI runtimeRuntime identity is bound into reuse eligibility.
✓Read-only checkout + no networkExecution is intentionally constrained.
✓Result-only deterministic targetsReuse a passing result, not stale output artifacts.
✓Conservative by defaultUnknown, unsupported or changed inputs execute fresh.
AGENTIC DEVELOPMENT
More edits should not mean blindly repeating the same work.
ZeroRun sits between an edit and the requested deterministic test target, deciding whether supported work can be safely reused or must execute fresh.
Agent edit
→
ZeroRun check
→
Reuse or fresh
PRIVATE DESIGN-PARTNER PILOT
Run ZeroRun on a real repository.
The next proof is repeated external usage. We are looking for a small number of teams with expensive deterministic tests and frequent agent-driven edits.
Hands-on setup for the first reviewed partitions
Real edit → test cycle measurement
Internal v0.4 qualification exceeded 5× on both accepted workloads; customer results are measured individually, not guaranteed
Explicit supported boundary and conservative fallback behavior
QUESTIONS
Measured claims only.
Did ZeroRun achieve the 5× target?+
Yes, for the two accepted internal v0.4 historical workloads: pytest measured 8.1657× same-runner compute efficiency and SymPy measured 5.4545×. That is not a promise that every customer repository will achieve 5×.
What happens when ZeroRun is not sure?+
It executes fresh. A changed source closure, runtime identity, command, policy or unsupported condition must not become a cache hit.
What exactly does a ZeroRun hit reuse?+
The current launch configuration is limited to result-only test targets whose commands suppress captured streams and publish no output artifacts. The reused outcome is a previous pass.
How was ZeroRun v0.4 qualified?+
The accepted internal evidence uses 10 distinct historical merged pytest patches and 20 distinct historical merged SymPy patches, pinned Linux/amd64 execution, reviewed reuse decisions, and fresh shadow validation for claimed hits. The public evidence page documents the final provenance boundary.
What workloads are the best fit today?+
Linux/amd64 repositories with deterministic, repeatedly executed test partitions and frequent small code edits are the clearest current fit.
Does ZeroRun automatically prove the source closure is complete?+
No. Closure completeness is an explicit operator-reviewed assertion in the current MVP. ZeroRun remains conservative around that boundary.
Is ZeroRun commercially proven?+
Not yet. Internal technical qualification passed; repeated design-partner usage, customer economics, willingness to pay and retention still need to be demonstrated.
