System One Coding: to compose code through decisions
Typed blocks, staged verification, measured comparisons, and a registered free pilot.
Dis article na machine translation.
System One Coding na Mithril staged coding harness: a decision model dey choose from typed building blocks, while di harness dey construct source, check am, and record di outcome. Try di [registered free pilot](https://code.mithril.fund/).
## Measured comparison of coding methods
Dem measure am 2026-10-06T07:33:38.924Z (JST date 2026-10-06), for di same local Node/NBB host through OpenRouter.
Five rounds dey alternate method order; up to four attempts each.
Every method must produce di same two CLJK function bodies, preserve di marker and pass both boolean toggle inputs plus all 511 completion vectors of lengths 0–8.
Di generation baselines dey output JSON expressions wey a restricted grammar admit before source emission; rejected outputs dey receive di same generic repair message.
Dis na a small constrained function task, no be SWE-bench or general repository coding.
| Method | Wey pass | 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 |
| Template wey dem know | 5/5 | 0.361 | $0.000000000 | 0 / 0 | 0, 0, 0, 0, 0 |
Time dey run from when di caller start through response parsing, admission, source emission and verifier completion.
System One include shared harness setup, receipts and replay checking; di other methods dey use a smaller admission/checking adapter.
Verification processes dey differ (two parallel per-function checks versus one combined checker).
Dis na a measured comparison of these implementations, no be an isolated model-speed test.
UI authoring, compile/deploy, network startup, electricity, hardware, human development and training no dey included.
API USD dey come from actual provider usage receipts, including rejected outputs and repair calls; token columns na medians.
No unknown charge dey treated as zero.
To reuse a template, di developer must already know a solution.
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) dey available.
Di measured comparison use $0.003558784 in model API charges in total.
For dis task Qwen median dey faster and cheaper than Jev; Gemini median also dey faster and cheaper.
To reuse di template wey dem know na di fastest, with zero model API charges.
Di current value of System One na di controlled choice space, dependency-aware composition, verifiable source and replayable evidence.
E never show a general speed, cost or accuracy advantage.
Five successful runs for one tiny task no dey establish a general reliability rate.
## How e dey work
1. A human dey define di task, available primitives, types, dependencies and fixed acceptance checks.
2. For each stage and AST hole, di harness dey present only compatible choices. Di completion-toggle stage get boolean arguments, constants and conditionals. Di count stage fit use di verified toggle function, vector filtering and counting.
3. Jev dey choose di next block. Di harness dey validate di choice and dey build di AST; confidence alone no fit accept code.
4. Di harness dey emit two CLJK source files, dey preserve unrelated definitions, and dey run independent fixed checks. A receipt dey record decisions, tokens, timing, API charges and verification. Failed verification dey stop publication.
5. Di browser dey assemble a TodoMVC starter around di verified kernel. You fit inspect/edit/export files, connect GitHub, create a repository, save a commit and explicitly publish am to GitHub Pages.
Di pilot dey generate completion toggle and unfinished count only.
UI, editing, bulk operations, URL filters, local persistence and WebAssembly integration na maintained starter code.
Arbitrary repository execution, UI/UX generation, API generation and automatic expansion of di block catalog na future work.
Your brief no dey create new capabilities outside dis catalog.
Manual edits dey outside di original proof.
## Register and try am free
Make you open [Mithril](https://mithril.fund/#system-one), register or sign in, then open [System One Coding](https://code.mithril.fund/).
No OpenRouter key or GitHub token dey needed for a free run.
Di Mithril browser session wey already dey authorize di trial.
GitHub credentials dey needed only to save or publish to your repository.
Di pilot dey allow **3 accepted attempts per account per UTC day**, with **100 attempts per UTC day across di service**.
Accepted attempts, including failure, a busy runner or an unknown outcome, dey consume a slot.
Invalid or unauthenticated requests no dey consume am.
If you replay di same request ID, e dey return di stored result without another model call.
Di server dey use one persistent transactional quota authority, no be browser storage.
Di pilot dey depend on say di operator local runner and public tunnel dey online; e no be an availability guarantee.
Independent provider-key runs dey require a GitHub connection and dey use your own OpenRouter balance.
## Browser performance na a separate measurement
Di [TodoMVC comparison](https://code.mithril.fund/benchmarks.html) dey measure application interaction and dey compare di maintained starter with Vanilla ES6.
Di Speedometer 3.1 baseline no include Mithril.
These results no dey measure coding-agent speed or prove an official Speedometer score for Mithril.
## Jev with CLEF
Di harness wey dey run dey call Jev through OpenRouter for state-dependent choices.
Normally dis kernel dey take eight choices, no be one decision for an entire program.
[Cloudflare CLEF](https://huggingface.co/Cloudflare/clef) dey describe a decision-model approach wey dey compatible with Jev/System One.
CLEF fit be a local/training backend; di deployed pilot no dey use a CLEF model wey dem train locally.
Earlier local attempts no qualify, and dem no claim any training gain.
## To extend di harness
For UI/UX, logic and API stages, make you define separate typed catalogs and verifiers.
For example a layout stage fit choose a maintained component; an API stage must select a reviewed contract, authentication and input-validation block.
Admit a stage only when di dependencies and check inputs dey available.
Keep model decisions as proposals; di harness own types, admissible paths, execution limits and publication gates.
Expand di catalog with held-out tasks and track all failures, repair counts, wall time and actual charges before you claim broader capability.
Di implementation and measurement records dey public. [Documentation](https://mithril.fund/docs/system-one-coding).
## Recording of di editor update
[Original-speed editor recording](https://code.mithril.fund/editor-update.mp4) dey show two functions wey dem never implement as dem change to verified source.
Dis na a local preview with actual Jev and a labelled test authentication fixture, separate from production qualification of registered users.
Dem record am 2026-10-06T07:55:00.252Z: click-to-source 2.631 seconds; harness receipt 2.107 seconds; 8842 / 406 tokens; API $0.000371364.
[Receipt of di measurement](https://code.mithril.fund/editor-update.json).