Ten posts, one throughline, and the extra that came in sideways.
| Applies to | Software delivery pipelines: build, sign, promote, admit, run |
| Post count | 10 numbered parts, plus one bonus |
| Series span | June 9 to July 9, 2026 |
Ship the Proof ran for a month, one post at a time. A reader who followed it week to week got the argument in installments, each one leaning on the last. A reader arriving now gets ten posts and no map back through them, which is a worse way to read a series that was built to be read in order at least once.
This is that map. It groups the ten parts into five moves the series makes, from naming a release correctly through watching the pipeline that produces it, plus a short note on the one post that came in sideways instead of taking a numbered slot. Read it as an index to return to, not a replacement for the posts themselves: each link below goes to the depth this overview deliberately skips.
The series has one thesis, stated in its first post and tested against a dozen new problems for the next nine: a release is defined by what it contains, not by the label someone hung on it. Everything from the branching model to the observability stack in the final post is that thesis applied to a different corner of the pipeline.
Release Identity and Branching
The series opens with the claim everything after it depends on. The Release Is a Content Digest, Not a Mutable Tag makes the case from a two-day production debugging session where the tag in the cluster never changed and the bytes behind it did. Once the digest, not the tag, is the real object of interest, Pick GitHub Flow When the Release Is an Attested Digest argues that git-flow’s release branch has nothing left to protect: the artifact you promote is the artifact you built at merge time, and staging that decision behind a branch just delays a call that was already made.
Evidence and Admission Control
Naming a release precisely only matters if what you attach to it survives the trip. Move the Evidence, Not Just the Image, When You Promote explains why a plain digest copy leaves signatures and SBOMs behind: they live as separate OCI referrer manifests linked by a subject reference, and a naive copy operation has no reason to know it should bring them along. Enforce Image Trust at Admission, Not by Convention moves the check to the one checkpoint every deployment path has to cross regardless of how it got there: the Kubernetes admission boundary, denying by default instead of trusting a signature nobody re-verified.
Build Provenance You Can Trust
With evidence traveling and getting checked, the series turns to where that evidence should come from in the first place. SLSA Build Level 3 Is an Isolation Property, Not a Project reframes Build Level 3 as something most teams on hosted CI are already close to, once the build itself runs isolated from anything a user-controlled step could reach. Produce a Software Bill of Materials You Can Actually Use sets the bar for what makes an SBOM worth generating at all, rather than a compliance artifact nobody opens: a real declared format, binding to the digest instead of the tag, and staying re-scannable without a rebuild.
Why This Isn’t Optional
Seven posts into a pipeline built on verifiable claims, Supply-Chain Hazards That Got Real in 2026, and the Fixes stops to ask why any of it is worth the effort. It walks three incidents from earlier in the year, including a compromised CI action reached through a force-pushed tag and the orphaned-attestation failure the series had already covered in its third post, and draws one lesson from all three: a mutable reference anywhere in a trust path is a liability whether or not anyone noticed it was one yet.
Measuring, Recording, and Watching
The last stretch turns the pipeline’s claims about itself into things worth trusting too. Honest DORA: Define a Deployment Before You Measure refuses to compute delivery metrics before defining what a deployment actually is, and argues that AI-assisted changes need their own cohort instead of being averaged into a signal that cancels itself out. Recording AI Authorship in Provenance You Can Trust answers the segmentation problem directly: record the AI-assisted label at commit time in a trailer built for it, then carry that label into the build provenance instead of reconstructing it after the fact. Observability for a Delivery Pipeline, Not Just the App closes the series by pointing the same instrumentation discipline at the pipeline itself, so the invariants the first nine posts established produce a signal when they break instead of failing silently.
Bonus: One Extra, Not Part of the Ten
Pin the SHA Yourself: What an Attested Marketplace Buys You is not one of the ten parts. It published between parts eight and nine and applies the series’ argument to a surface the numbered series never reaches: a Claude Code plugin marketplace. A plugin is code your assistant loads and runs with your credentials, and an attested, SHA-pinned marketplace raises the cost of attacking that supply chain without removing the human decision of when to trust the next commit.
Where to Start
Read the ten parts in order if the whole arc is what you want: each one leans on the last, from naming a release through watching the pipeline that ships it. If you are chasing one problem instead, start in whichever section above matches it. Every post restates just enough of the earlier argument to stand on its own, so skipping ahead costs you connective tissue, not comprehension. The bonus post slots in after part eight by publish date, but it reads fine at any point since it depends on the series’ thesis, not on any single earlier post.