StatusShipped, GitHub Pages library and Python CLI
RuntimePython 3.10+, no dependencies, Java agent from Frame Bench, Minecraft 26.3
Try itBrowse setups
Proof hooksPer-frame intervals in every contribution, recomputed by CI, 93 unit tests
How it works
The CLI entry point issilicon_shader/cli.py. discover and doctor read the Mac and a Prism instance (or any game directory with --game-dir) and never launch the game. isolate copies a closed instance to a new destination and refuses existing destinations and symlinked content. It doesn’t copy accounts, logs or launcher hooks.
profile apply in instances.py works only on that managed copy. It saves every file it will touch, writes a receipt with each file’s hash before and after, and reverts on error. profile rollback RECEIPT_ID restores the saved files and refuses if a file has changed since the apply. The writer covers game, Iris and render-scale settings. It doesn’t install shader packs or mods, and it leaves the graphics backend and window mode alone. daily makes a clean copy with the measurement agent and Minescript stripped out.
Timing happens inside the running game. A Java agent hooks the return of Minecraft.renderFrame(boolean) and records System.nanoTime() once per frame. A Minescript script, reference_flight.pyj, checks the world’s seed and generator, sets FOV 70 and clear noon, waits through 30 s standing still, then flies a fixed 420-tick route with a turn, about 21 s, and records the camera pose, focus and timing as it goes. challenge route-receipt checks those observations against the pinned route in overworld-v1.json. challenge prepare then ties each receipt to its frame CSV and builds a bundle with one baseline run and one candidate run.
Accepted bundles live in contributions/, named by the SHA-256 of their content. The Challenge site workflow runs registry.py on every push to main. It validates each bundle, recomputes its metrics and writes site/data.json, which site/app.js renders as filterable cards. A pull request that adds only a valid current-route bundle is merged by the Challenge evidence workflow. Recipe and code changes still need review.
How Frame Bench fits in
Minecraft Frame Bench is the standalone timing harness the measurements came from. ItsFrameAgent.java is a -javaagent that uses ASM to hook renderFrame(boolean), or runTick(boolean) if that method is missing. It buffers timestamps, checks window focus and writes frame,nanotime,interval_ns CSV after the capture. The agent refuses to arm while Minecraft’s inactive-FPS policy is AFK, because scripted movement doesn’t reset the idle timer and the client drops to 30 FPS after a minute. Frame Bench also ships Minescript pan and movement routes, a capture.py driver that waits for at least 75% CPU idle and no new swap-outs, and summarize.py. It contains no measurement data.
Silicon Shader keeps a copy of that agent in recipes/m4-wide24/sampler/, and the current workload guide names it as the 26.3 adapter. The copy differs from Frame Bench main in a few lines: FrameSink.requestStop() ends the capture on the next frame once the route finishes, so the last slow interval is kept, and the error field resets between captures. Every published bundle records its harness as FrameBench-campaign-v1. The repository’s other sampler, sampler/, is a generic agent configured by hook.properties. It has a different completion format, and the current route doesn’t use it.
Published measurements
All four current-route setups were measured on one base M4 MacBook Pro (10 CPU cores, 10 GPU cores, 24 GiB), Minecraft 26.3, Fabric 0.19.5, Java 25, with a 120 FPS cap and vsync off. Three bundles record AC power, and Clear Water 24 records none. Each card is one pair: a baseline flight with the M4 Wide View settings (MakeUp Ultra Fast, 24 chunks, 60% render scale) and a candidate flight with the new setup, each from its own launch. The Far Horizon, Vibrant 16 and Max 32 recipes say the first flight of each launch was dropped as warm-up and name the later flight that was kept. The Clear Water 24 recipe doesn’t say which flight it used.measure.py defines the numbers. Average FPS is the interval count times 1000 divided by the summed interval milliseconds. Worst five seconds is the lowest frame rate over any window of consecutive intervals that adds up to at least 5 s. Candidate figures at commit 66be124:
Each bundle file holds the full list of frame intervals for both runs (about 2,070 to 2,290 per flight), the SHA-256 of each original CSV, both route receipts and the settings. The bundles keep relative intervals and leave out absolute timestamps and private paths. Recomputing the metrics from those intervals with
measure.analyze gives the stored values and the README table. The fifth card, M4 Wide View, used an earlier 21 s camera pan in a desert village. The site lists it under earlier routes, apart from the current four.
The numbers show how quickly the game’s main thread finished frames on this one Mac, over one 21 s flight per side, for settings recorded in the bundle. Every candidate is slower than its baseline, since each one buys a different look, a longer view or stronger reflections with frame rate.
They don’t show these things:
- GPU time, displayed frames, input latency or generated frames. The hook times CPU frame production, limiter waits included.
- Any Mac other than the base M4. The hardware inventory lists every released model, and none of the others has a result.
- Variance. Each figure is one flight. The recipes report neighbouring flights within about 1 to 3 FPS, and the README treats gaps under 5 FPS as ties. In the Far Horizon recipe the discarded warm-up flight (100.9 FPS) was faster than the kept one.
- Equal image quality. Render scale, view distance, shader pack and window mode differ between cards, and Clear Water 24 ran in exclusive fullscreen at a taller output.
- Rates above the 120 FPS cap.
- World generation or all-day play. Terrain is generated before timing, and inventory, building and save/reload are separate manual checks.
- Tamper-proof results. Bundles are self-reported. CI checks consistency, but it can’t confirm that a run happened.
Verify it
These ran on 23 October 2026 againstmain at 66be124, without launching Minecraft:
Ran 93 tests and OK in about 6 s. The registry printed {"entries": 5, ...}, and challenge workload printed the pinned route (seed 20260929, 420 client ticks, start -121.3 100 438.7).
To recompute the candidate figures from the published intervals:
silicon-shader-overworld-v1 lines match the table above, and the cosine-pan-90-21s line is the earlier-route card. ./sampler/smoke.sh builds the generic sampler (it downloads ASM 9.10.1 and needs a JDK) and runs it against a stub fixture. It passed, and it proves the file protocol only.
For Frame Bench, the build runs its AFK policy test before it packages the agent:
Live policy regression checks passed (including latent AFK rejection). and the hashes of frame-agent.jar, frame-sink.jar and, on macOS, the controller input helper. The agent’s manifest embeds the absolute path of asm.jar, so build on the machine that will run it. No game capture was run for this page.
Limits
- Minecraft Java 26.x on Apple Silicon only, ideally through Prism. Bedrock, Intel Macs, iPhone, iPad and Windows aren’t covered.
- The CLI doesn’t launch Minecraft, download mods or shaders, or upload anything. You install those from their publishers, and the repository redistributes no game, shader or mod files.
- Automatic measurement needs an agent that can send chat commands, press F2 and keep the window focused, plus macOS Accessibility permission. With shell access alone, a setup gets reported as unmeasured.
- The timing hook is tied to the Minecraft version. Frame Bench has mappings for 26.2 and 26.3, and the current route requires 26.3.
- All current results come from one machine, and the author recorded them. No outside submission exists yet.
- The site is a static GitHub Pages build with no backend. Entries are self-reported.
Source
- Repository
- Live library
- Workload definition
- Fixtures CI
- Challenge site workflow
- Campaign notes
- Minecraft Frame Bench and its CI
66be124 and Frame Bench 9742471.