Case studies / payloadcms/payload
This is not a savings-proof case study, and there is no verified number here to cite. It's the opposite, and that's the point: our engine produced a clean, webpack-tier +45.8% estimate — and the discipline that sits after the estimate correctly threw it out. This is the anti-over-claiming story. A guarantee is only worth something if the machinery behind it will kill a number that doesn't hold up; here it did, on a number that looked real.
payloadcms/payload cleared our scout cleanly: a 22-minute pipeline, in-band, no scan
defect. The deep analysis pass returned a composed 1346s → 729s (+45.8%) makespan
over two accepted changes — on its face indistinguishable from the webpack estimate that
did verify.
The bigger piece. The job carries a condition that excludes fork pull requests from running at all — not a missing-secret problem with a workaround, a hard gate. No fork-based verification path exists, full stop.
The smaller piece — proposed sharding a "monolithic" job that a real run already fans out into 108 parallel shards via a dynamically-computed matrix our engine couldn't see into. There was no monolith to shard.
Net: one half unverifiable, the other half not a real gap. The composed +45.8% does not survive contact with the repo's real structure. The discipline discounted it — before it could become a guarantee.
It does not demonstrate a delivered saving — cite it for that and you've inverted its meaning. What it demonstrates is the guardrail working: a number our engine was confident about, discounted by the layer whose whole job is to decide — for real — whether a saving is there. The estimate is where the work starts; the discipline is what decides whether it ships. On Payload, it decided no.