Typed data flow
Programs declare the shape of their data and the steps that transform it. Immutable values and bounded operations make the computation easier to inspect.
Alex SOLONSKYProjectsDiscuss SOLONCODESOLONCODE
A language for programs whose actions you can inspect
I am creating SOLONCODE: a typed language and runtime for controlled AI integrations. Development continues, and I am already building my first new products with it.
Working language and runtime · Active development
INPUT → PROGRAM → AUTHORITY → RESULT
01Typed inputs
02SOLONCODE program
03Explicit host permissions
04Observed outcome
Truth = yes | no | unknown
pure fn · effect fn · solonsky
Why build a language of my own?
I want the connection between an intention, a program and an external action to be explicit. What data does a step require? Which operations can it perform? Who grants permission? What happens when the outcome is unknown? SOLONCODE makes these questions part of the program and its host.
Explicit external actions
Pure calculations and effectful functions are distinct. Files, models and external tools belong to the configured host and its granted capabilities.
Uncertainty stays visible
Truth has three states: yes, no and unknown. An unknown external outcome can be held for inspection instead of quietly becoming permission to repeat it.
A record beyond one response
Selected service profiles combine retained context, owner decisions, tools and durable work queues. A saved outcome stays connected to the action behind it.
Permission is a decision, not a word in a model reply
This small illustration shows the distinction. A program value describes a state; the host decides whether it has the authority to perform an action.
The value is unknown. Keep the uncertainty visible and wait for an explicit decision.
Interactive explanation · Does not execute SOLONCODE or call external tools
What you can examine today
macOS · Pure programs
A signed and notarized Apple Silicon package, 0.5.56-trust1, runs the supplied pure project profile. Installation and trust checks have dedicated documentation.
Read the profilemacOS · Native project tools
The 0.5.57 profile adds a supplied project and an explicitly approved native tool call. Its boundary and approval steps are documented.
Read the profileWebAssembly · Pure projects
The 0.5.57 reactor runs modular pure programs through a fixed Wasm interface. This profile has no file, network or external-effect access.
Read the profileHosted integrations
SDK host paths cover connected services, tool registration, selected memory and durable work. The implementation sources remain private.
Read the profileA connected application has already been exercised
A documented integration joins project modules, selected memory, media, typed tools, owner decisions and a service queue. The installed profile completed five business jobs in one uninterrupted process, with four synthetic requests, quantity 14 and total 182. Separate cases cover quotas, cancellation, crash/reopen and uncertainty handling. These are bounded integration checks, with their own acceptance criteria.
Read the profileRESEARCH / 21.10.2026 / MINIMAX M3.1 FLASH
Measured against the same tasks
I tested how MiniMax M3.1 Flash understands existing synthetic code and produces a structured patch in SOLONCODE, JavaScript and Python. This is a completed pilot from 21 October 2026: six tasks, two fixed repetitions and 12 planned episodes per language.
36 completed episodes56 recorded attemptsAll attempts included
SOLONCODE6 / 12
JavaScript8 / 12
Python6 / 12
Success requires correct facts about the old program, three public examples and twelve hidden checks. Each episode allows at most one repair based on public feedback.
What this pilot shows
SOLONCODE solved 6 of 12 episodes, JavaScript 8 and Python 6. Tokens per successful episode were 26,437, 23,099 and 27,905 respectively. SOLONCODE used about 14.5% more than JavaScript and 5.3% less than Python on this measure. The pilot did not establish an overall advantage.
The predeclared criterion was stricter: all three languages must solve 12/12 episodes, with complete usage; both the token ratio and its upper 97.5% task-bootstrap bound must be at most 0.90 against each control. That criterion was not met.
The complete comparison
| The complete comparison | SOLONCODE | JavaScript | Python |
|---|---|---|---|
| Success / planned | 6 / 12 | 8 / 12 | 6 / 12 |
| First response successes | 4 | 5 | 4 |
| Attempts / repairs | 19 / 7 | 18 / 6 | 19 / 7 |
| Input tokens | 33 599 | 41 859 | 44 055 |
| Generated tokens | 125 023 | 142 930 | 123 373 |
| Total tokens | 158 622 | 184 789 | 167 428 |
| Tokens per successful episode | 26 437.0 | 23 098.6 | 27 904.7 |
| Schema errors | 10 | 8 | 11 |
| Reached program evaluation | 9 | 10 | 8 |
| Median episode time, seconds | 85.156 | 124.171 | 76.451 |
All six tasks
Successful episodes out of two repetitions. No task has been removed from the comparison.
| Task | SOLONCODE | JavaScript | Python |
|---|---|---|---|
| T01 | 2 / 2 | 1 / 2 | 2 / 2 |
| T02 | 1 / 2 | 2 / 2 | 1 / 2 |
| T03 | 1 / 2 | 2 / 2 | 1 / 2 |
| T04 | 1 / 2 | 2 / 2 | 1 / 2 |
| T05 | 0 / 2 | 0 / 2 | 0 / 2 |
| T06 | 1 / 2 | 1 / 2 | 1 / 2 |
What the failures teach me
29 of the 56 responses failed the required patch schema; 27 reached program evaluation. A response ending naturally is therefore not the same as a valid patch or a correct program. On task T05 every language passed the patch examples in its final response but failed the required understanding of the old code, so all six episodes remain unsuccessful.
Read these results within their scope
- One model, six selected tasks and two repetitions: the result does not automatically generalise to other models or projects.
- This measures code understanding and a constrained patch workflow, not language execution speed or the complete cost of an autonomous agent.
- All 56 responses ended naturally. Hidden checks were never feedback to the model.
- The actual provider bill and separate reasoning-token counts are unknown. Timing is descriptive and includes recording and checks.