Cloudflare Docs ↗
A frozen documentation corpus, strict output audits, and a transparent comparison of Nift and Astro build workflows.
Read the experiment →Nift Labs / experimental engineering
Production-scale migrations. Composable integrations. Runtime benchmarks. Build-system and task-runner experiments. We put software to the test and leave the evidence in the open.

01 / Fields of inquiry
Public content corpora, rebuilt with Nift. Independent designs, real workloads, and evidence you can inspect.
Start runtimes, execute scripts, build pages, inspect native build graphs, orchestrate tasks, assemble templates run concurrent work and transform project trees. Eight experiments with retained evidence and distinct measurement boundaries.
Nift alongside application frameworks, UI libraries and build tools. Compare architectures and make the boundaries explicit.
On the horizon
Scaling investigations / Framework comparisons / Runtime & scripting / Languages & FFI / Packages & tooling
02 / Migration experiments
A frozen documentation corpus, strict output audits, and a transparent comparison of Nift and Astro build workflows.
Read the experiment →An agentic reconstruction of a production-shaped website. Build speed, dependency footprint, and developer experience under the microscope.
Read the experiment →One production site, two migration philosophies. Faithful MDX authoring and an agent-maintained HTML reconstruction, compared with the original Astro build.
Read the experiment →A real Hugo documentation migration. Preserved Markdown and maintained HTML, with publication parity, profiled builds and explicit maintenance tradeoffs.
Read the experiment →Authoritative Markdown or maintained HTML: a faithful production publication, profiled source derivation and measured everyday iteration loops.
Read the experiment →Three production workflows, retained React islands and two source models. Faithful publication, profiled builds, memory and coordinated maintenance.
Read the experiment →962 routes and retained React islands. Authored and rendered source, production iteration, fresh-state costs and qualified memory evidence.
Read the experiment →Migration and rewrite closed as unsuccessful: retained runtime costs versus an expensive generated corpus pipeline. Two distinct findings, preserved parity and qualified measurements.
Read the experiment →Public-output reconstruction, scoped visual/playback parity, and two Nift source models. No private SpaceX build benchmark.
Read the experiment →A 2,003-page documentation publication retaining TanStack Start and React, with lower observed build time and memory and dependency-aware incremental work through Nift.
Read the experiment →03 / Benchmarks
Startup-sensitive snippets, sustained computation, and file/data workloads. Compare all participating languages with checked outputs and retained samples.
Read the benchmark →From 100 to 10,000 pages: full builds, Nift incremental and targeted builds, elapsed time, and memory—with output equivalence and state boundaries attached.
Read the benchmark →Startup and RC costs, native scripting, process scaling and mixed 100k-file tasks. Bash, Zsh, Fish, Nushell and Nift, with checked outputs and explicit native/Unix-tool boundaries.
Read the benchmark →GNU Make, Ninja and Nift execute the same native graph. No-op and incremental scaling, generated inputs, compiler-heavy builds and tiny actions—with exact rebuilt sets and output hashes.
Read the benchmark →Nift, Turbo and Nx execute controlled project workflows and tiny tasks. Exact execution sets, file invalidation, concurrency, targeting, failure recovery and local-state boundaries.
Read the benchmark →Five engines assemble identical documents. Fresh invocation, preparation, reuse, loops, partials and memory, with exact byte checks and a candid practitioner assessment.
Read the benchmark →Five runtimes solve equivalent concurrency problems: scheduling, waits, CPU scaling, bounded pools, queues, shared state, subprocesses and memory. Exact correctness and explicit model boundaries.
Read the benchmark →Five runtimes traverse, inspect, fingerprint and transform identical project trees. Exact outputs, controlled mutations, warm-state boundaries and separate syscall/allocation diagnostics.
Read the benchmark →04 / Integrations
Three architectures, one 3,000-document fixture. Investigate static ownership, selective interactivity and the full-stack application boundary.
Read the integrationA note on method
We use real public content corpora and explicitly labelled generated fixtures to investigate build-system performance and implementation characteristics. Unless a report explicitly says otherwise, these are independent interpretations rather than pixel-perfect migrations.
Exact upstream commits are recorded where available so each result can be tied to a specific corpus. Each report describes its measurement boundaries, evidence, and limitations.