Why does the SLSA v1.0 build track stop at L3 rather than requiring hermetic or reproducible builds?

## What v1.0 actually defines SLSA v1.0 reorganised the specification into **tracks**, and shipped one of them: the **build track**, with levels **L0 through L3**. In outline, L0 means no guarantees; L1 means provenance exists and describes how the artifact was built; L2 means that provenance is generated and signed by a hosted build platform rather than asserted by the person who ran the build; L3 means the build platform is hardened, so a tenant running on it cannot forge provenance or reach into another build's execution. There is no L4 in the build track, and hermeticity and reproducibility appear at no level. A contract clause demanding "SLSA Level 4" is asking for something the current build track does not define. ## The design principle behind the omission The common thread across L1-L3 is **verifiability by the consumer**. Each level is something a consumer can establish from the artifact, its provenance, and knowledge of the build platform that issued it. It is also, crucially, a property of the **platform**: assess a platform once and every build it runs inherits its level. That is what makes the ladder practical to adopt across an ecosystem — a hosting provider does the hardening work, and thousands of projects move up a rung without changing their own build files. Hermeticity does not have that shape. It is a property of **each project's build definition**: whether *this* build declares all its inputs and runs without network access. Two repositories on the same L3 platform can differ completely. So the platform cannot attest it for its tenants, and a consumer reading the provenance cannot see it. To make it a level you would need per-build evidence that a consumer could evaluate, in a form that generalises across every language and build tool — a much harder specification problem than "is this builder hardened." Reproducibility is harder still, for a different reason. To a single verifier, a claim of "this build is reproducible" is worth nothing on its own; the property only produces evidence when a **second, independent party** rebuilds and compares digests. That needs a rebuilder network, agreed build environments and a place to publish results — infrastructure the specification does not define and cannot assume. A level nobody can check is not a level. There is an attainability argument too. Whole language ecosystems are some distance from hermetic, deterministic builds by default. Making it a requirement would place most of the world at the bottom of a ladder designed to be a practical adoption path, which would just push people to ignore the ladder. ## They also defend different threats The omission is not a judgement that these properties are unimportant — it is that they are **orthogonal** to what the build track measures. | Property | The threat it addresses | | --- | --- | | Build L2-L3 | Forged or tampered provenance; one build interfering with another | | Hermeticity | Inputs entering the artifact that provenance never recorded | | Reproducibility | A trusted builder emitting a correct-looking statement over the wrong bytes | That last row is the residual risk L3 deliberately leaves in place. At L3 the build platform is the **root of trust**: you believe the artifact came from that source because a hardened builder said so, signed with its identity. If the builder itself is subverted, its provenance is still perfectly valid and describes bytes that were never derived from that source. Reproducibility is the only control on that list that does not route through the builder's word. ## Using this in a supplier conversation This comes up in procurement. A supplier reports Build L3 and a security team asks for hermetic builds on top; the supplier reasonably asks why, if they are already at the top of the track. The productive move is to stop trading level numbers and name the threat you are carrying. "We are relying on your build platform as a single root of trust, and we have no independent way to check its output" is a concrete concern, and it maps to concrete asks: publish the build definition; declare and pin all inputs; state whether the build reaches the network at run time; and — the strongest one — allow a nominated second party to rebuild a release and compare digests. Those are checkable. "Be L4" is not, because it does not exist. The reverse case matters as well: a supplier who *can* demonstrate independently reproducible releases is offering something stronger than a higher rung, even though no level number rewards them for it. Judge the properties, not the badge.