System One Coding: composing code through decisions
Typed blocks, staged verification, measured comparisons, and a registered free pilot.
The article body is not available in this language; showing English.
System One Coding is Mithril's staged coding harness: a decision model chooses from typed building blocks, while the harness constructs source, checks it, and records the outcome. Try the [registered free pilot](https://code.mithril.fund/).
## Measured coding-method comparison
Measured 2026-10-06T07:33:38.924Z (JST date 2026-10-06), on the same local Node/NBB host through OpenRouter.
Five rounds alternate method order; up to four attempts each.
Every method must produce the same two CLJK function bodies, preserve the marker and pass both boolean toggle inputs plus all 511 completion vectors of lengths 0–8.
The generation baselines output JSON expressions admitted by a restricted grammar before source emission; rejected outputs receive the same generic repair message.
This is a small constrained function task, not SWE-bench or general repository coding.
| Method | Passed | Median seconds | Median API USD | Input / output tokens | Calls per run |
| --- | --- | ---: | ---: | ---: | --- |
| System One / Jev | 5/5 | 2.973 | $0.000401268 | 9554 / 406 | 8, 8, 9, 8, 8 |
| Qwen3 Coder Next + repair | 5/5 | 2.195 | $0.000089240 | 314 / 48 | 2, 2, 2, 2, 1 |
| Gemini 2.5 Flash + repair | 5/5 | 2.624 | $0.000231600 | 297 / 57 | 1, 2, 2, 3, 1 |
| Known template | 5/5 | 0.361 | $0.000000000 | 0 / 0 | 0, 0, 0, 0, 0 |
Time runs from caller start through response parsing, admission, source emission and verifier completion.
System One includes shared harness setup, receipts and replay checking; the other methods use a smaller admission/checking adapter.
Verification processes differ (two parallel per-function checks versus one combined checker).
This is a measured comparison of these implementations, not an isolated model-speed test.
UI authoring, compile/deploy, network startup, electricity, hardware, human development and training are excluded.
API USD comes from actual provider usage receipts, including rejected outputs and repair calls; token columns are medians.
No unknown charge is treated as zero.
Template reuse requires a solution already known by the developer.
Raw [20-run records](https://code.mithril.fund/coding-benchmark.json) and [reproduction source](https://github.com/mithril-lang/mithril-fund/tree/main/apps/code/bench) are available.
The measured comparison used $0.003558784 in model API charges in total.
For this task Qwen's median is faster and cheaper than Jev; Gemini's median is also faster and cheaper.
Reusing the known template is fastest with zero model API charges.
System One's current value is its controlled choice space, dependency-aware composition, verifiable source and replayable evidence.
It has not demonstrated a general speed, cost or accuracy advantage.
Five successful runs on one tiny task do not establish a general reliability rate.
## How it works
1. A human defines the task, available primitives, types, dependencies and fixed acceptance checks.
2. For each stage and AST hole, the harness presents only compatible choices. The completion-toggle stage has boolean arguments, constants and conditionals. The count stage can use the verified toggle function, vector filtering and counting.
3. Jev chooses the next block. The harness validates the choice and builds the AST; confidence alone cannot accept code.
4. The harness emits two CLJK source files, preserves unrelated definitions, and runs independent fixed checks. A receipt records decisions, tokens, timing, API charges and verification. Failed verification stops publication.
5. The browser assembles a TodoMVC starter around the verified kernel. You can inspect/edit/export files, connect GitHub, create a repository, save a commit and explicitly publish it to GitHub Pages.
The pilot generates completion toggle and unfinished count only.
UI, editing, bulk operations, URL filters, local persistence and WebAssembly integration are maintained starter code.
Arbitrary repository execution, UI/UX generation, API generation and automatic expansion of the block catalog are future work.
Your brief does not create new capabilities outside this catalog.
Manual edits are outside the original proof.
## Register and try free
Open [Mithril](https://mithril.fund/#system-one), register/sign in, then open [System One Coding](https://code.mithril.fund/).
No OpenRouter key or GitHub token is needed for a free run.
The existing Mithril browser session authorizes the trial.
GitHub credentials are needed only to save or publish to your repository.
The pilot allows **3 accepted attempts per account per UTC day**, with **100 attempts per UTC day across the service**.
Accepted attempts, including failure, a busy runner or an unknown outcome, consume a slot.
Invalid or unauthenticated requests do not.
Replaying the same request ID returns its stored result without another model call.
The server uses one persistent transactional quota authority, not browser storage.
The pilot depends on the operator's local runner and public tunnel being online; it is not an availability guarantee.
Independent provider-key runs require a GitHub connection and use your own OpenRouter balance.
## Browser performance is a separate measurement
The [TodoMVC comparison](https://code.mithril.fund/benchmarks.html) measures application interaction and compares the maintained starter with Vanilla ES6.
Its Speedometer 3.1 baseline excludes Mithril.
These results do not measure coding-agent speed or prove an official Speedometer score for Mithril.
## Jev and CLEF
The live harness calls Jev through OpenRouter for state-dependent choices.
Typically this kernel takes eight choices, not one decision for an entire program.
[Cloudflare CLEF](https://huggingface.co/Cloudflare/clef) describes a decision-model approach compatible with Jev/System One.
CLEF is a possible local/training backend; the deployed pilot does not use a locally trained CLEF model.
Earlier local attempts did not qualify, and no training gain is claimed.
## Extending the harness
For UI/UX, logic and API stages, define separate typed catalogs and verifiers.
For example a layout stage may choose a maintained component; an API stage must select a reviewed contract, authentication and input-validation block.
Admit a stage only when its dependencies and check inputs are available.
Keep model decisions as proposals; the harness owns types, admissible paths, execution limits and publication gates.
Expand the catalog with held-out tasks and track all failures, repair counts, wall time and actual charges before claiming broader capability.
The implementation and measurement records are public. [Documentation](https://mithril.fund/docs/system-one-coding).
## Editor update recording
[Original-speed editor recording](https://code.mithril.fund/editor-update.mp4) shows two unimplemented functions changing to verified source.
This is a local preview with actual Jev and a labelled test authentication fixture, separate from production registered-user qualification.
Recorded 2026-10-06T07:55:00.252Z: click-to-source 2.631 seconds; harness receipt 2.107 seconds; 8842 / 406 tokens; API $0.000371364.
[Measurement receipt](https://code.mithril.fund/editor-update.json).