Benchmarks Studio · first slice

Show the run, not the promise.

These are the checks that keep the public sandbox honest: a change really saves and reopens, a broken save is caught, approval reaches the public view, and the exported blueprint carries what you chose. The file behind this page is a real run, not a template.

Source
951b313
Run at
2026-11-15T13:30:04.758Z
Runtime
v26.13.1 · darwin arm64
Suites
4
Recorded source
The commit this run was produced from. pnpm bench --write refreshes it; a later commit that does not change the sandbox leaves the result valid.
20passed
0failed
0blocked

Suites

Portable appearance-skin contract

appearance-skin · 4ms

pass
  • The contract is written and linked for agents

    The recipe, pressure and storage rules are documented and linked from agent and design guidance.

    pass
  • Every shipped recipe compiles at 1–4

    Two distinct recipes compile to bounded CSS with no URL, import, HTML or animation.

    pass
  • Composer default emits no layer

    The default stays byte-identical across pressure values; an old record cannot gain a layer.

    pass
  • Admin and public scope share one grammar

    The admin and the public reference carry the same compiled declarations under different roots.

    pass
  • The committed preview sheet matches the recipes

    8778 bytes, eight skin/pressure blocks, generated from the same compiler.

    pass
  • Press feedback stays off, plays on request and degrades honestly

    Off makes no sound or vibration; on plays two ticks and one 12ms pulse; reduced motion drops the settle and pulse; unsupported reports false.

    pass
  • A fresh context produced a valid recipe from the contract

    Codex 0.145.0 / gpt-5.6-sol produced pressed-folio from the written contract; the receipt states the isolation limits.

    pass

Blueprint export and re-import

blueprint-roundtrip · 1ms

pass
  • The sandbox exports a validated blueprint

    Exported version 2 preset “composition” with 3 links.

    pass
  • Re-importing carries the supported fields

    Name, introduction, theme, layout, both enabled blocks and 3 links survived export → JSON → re-import.

    pass
  • Preview-only gallery and accent are reported as not carried

    Gallery and accent are named as not carried; re-import turned the gallery off and kept each sandbox's own accent (clay, ember).

    pass
  • An invalid blueprint is refused

    Four invalid blueprints (bad preset, bad theme, no review-first privacy, no version) were refused and the sandbox was left untouched.

    pass

Live /demo sandbox

live-demo · 190ms

pass
  • GET /demo

    HTTP 200, 47045 bytes, 133ms; marker present: “Build something people can use”.

    pass
  • GET /demo/connected-apps

    HTTP 200, 52994 bytes, 57ms; marker present: “A guestbook and a showcase”.

    pass

Public sandbox: save, reopen and moderation

sandbox-save-reopen · 1ms

pass
  • An edit survives JSON save and reopen exactly

    7 edits (name, tagline, theme, accent, layout, block, link) reopened byte-identically after JSON.stringify/parse.

    pass
  • A save that keeps the initial state is detected as different

    The untouched save still reads “The wheel, in motion”, not the edited “Ring market”.

    pass
  • A tampered save is rejected and reported as recovered

    A save with no contributions and a save with version 2 were both refused; each recovered to the initial sandbox instead of half-applying.

    pass
  • Approving a pending note makes it public

    The pending note left the queue and appeared in the public list (2 public, 0 pending).

    pass
  • Hiding an approved note removes it from the public view

    note-1 is hidden from the public list and still present in the owner's list.

    pass
  • Restoring an older version does not republish a hidden note

    Version 1 restored the saved name while note-1 stayed hidden and note-2 stayed pending; appearance and moderation travel separately.

    pass
  • A restored sandbox still exports a valid blueprint

    Export after restore produced preset “composition” with 2 links.

    pass

Cloudflare creation and operations

Can someone go from an idea to a working, owned application, supervise real changes, recover from problems and keep improving it? The CF-COMPOSE campaign defines six required end-to-end creator journeys and nine additional real-product profiles. These are evaluation definitions, not claims that Cloudflare or App Composer has passed the journeys.

Required journey status: 6 of 6 not run. Real-product profile status: 9 of 9 not run. The original local foundation's checks are not full-journey results.

Show nine additional real-project evaluations
  • CF-R01 · not-run

    New creator to useful owned application

    A person without existing Cloudflare/Wrangler knowledge creates a useful multi-resource app, reopens it, makes a second change and owns the result.

  • CF-R02 · not-run

    Phone-only supervised agent change to verified live release

    An authorised phone user supervises a real agent change, preview, review, guarded release and independent live verification while their laptop is unavailable.

  • CF-R03 · not-run

    Parallel agent alternatives with Git/Artifacts portability

    Two independent agents create competing improvements from an identical immutable base, review the actual alternatives, retain all candidates and export the selected result to normal Git/GitHub.

  • CF-R04 · not-run

    Original song to finished editable music video

    A rights-cleared original song becomes a complete coherent music video with revision-linked scene plan, approved takes, audio sync, real render, saved originals and usable export.

  • CF-R05 · not-run

    Real game and music player experience on mobile

    A real browser/player completes a meaningful first-five-minute game or start/navigate/seek audio task with actual canvas, sound and performance evidence tied to its source version.

  • CF-R06 · not-run

    Two-tenant governed operations, costs and region

    Two distinct creator/tenant identities can use permitted apps and scoped agent actions, while cross-tenant access, unauthorised spending and unsupported residency claims are blocked.

  • CF-R07 · not-run

    App-wide incident investigation and independently verified recovery

    An authorised operator diagnoses a controlled multi-resource regression, proposes proportionate repair and proves usable application recovery without losing protected state.

  • CF-R08 · not-run

    App Composer recursively rebuilds and evolves itself

    An independently owned App Composer installation can be constructed through Composer's documented method, reproduce supported modules and then use the new instance to create and evolve a real third app.

  • CF-R09 · not-run

    Evidence-led portfolio telemetry and actionable Cloudflare feedback

    An operator sees which exact versions are live and which tasks actually completed across multiple independent projects, then publishes one rights-approved source-linked benchmark and one reproducible upstream improvement request.

They cover the first-time creator, phone-only agent releases, parallel Artifacts/Git workspaces, an original music video, real gameplay and listening, multi-tenant IAM and costs, full-app recovery, independent Composer self-rebuild, and public-safe portfolio proof. Each profile has a positive control, designed failures, measurable outcomes and specified source evidence.

node benchmarks/runner.mjs --suites benchmarks/cloudflare-composition --suite CF-COMPOSE-01

This command runs only the provider-free local foundation. Complete journey and profile runs require separately authorised browser/provider/human evidence and are not made green by a CLI exit. Read the full protocol and test scenarios · Evaluation owner #601.

What this run does not prove

Release timing receipts

The guarded release writes a redacted release-timing/v1 receipt. Benchmarks Studio can validate and compare those receipts without executing them: one row per repository and release path, with the total, slowest stage, three-minute target, strict 300-second limit, status and measured coverage.

node benchmarks/release-timing/report.mjs --dir .release/timing

Receipts stay in each repository's ignored .release/timing/ directory, so this page does not publish them. The portable pack format, refusal rules and Composer/Studio fixtures are in the ingest README. That is the known gap: the comparison is a local report today, and this page links to it rather than pretending a private receipt is public evidence.

Run it yourself

From a checkout of this repository:

pnpm bench

The runner prints the same JSON. Add --write to replace the results file this page reads. It exits non-zero only when a check fails; blocked checks are reported and counted.