Live now
Try polycode.
A free coding AI. Works in your browser or your terminal. No sign-in.
Polylogic AI
Research-driven
software.
We research the problem first, then build the software that answers it. Every Polylogic product ships with a published paper behind it.
4 papers · 4 minted · 1 runnable CLI · CC-BY-4.0
The meta-question
Can software prove
its own claims?
Most AI confidently generates answers it never checks. Our research program asks: what happens when you verify every output against something built completely differently? Four papers answer that question. The principle behind them drives every product we build.
The body of work
Four papers. One program.
Read them in order. Each paper builds on the primitive the previous one names.
- 01A three-primitive architecture for canon-keeping AI agents
Names the three primitives an AI agent must have to keep a canon it can defend: the engine that runs, the rules that bound it, and the canon that remembers. Establishes the substrate-cannot-witness-itself axiom and the non-LLM composer rule for the witness stack.
Theory layer. Defines the ground the other three papers stand on. - 02Empirical proof of disjoint-substrate verification
Cohen's κ = 1.000 agreement between transformer ensembles drawn from disjoint provider families on 30 NeurIPS 2024 papers. κ = 0.000 when the structurally non-transformer witness was asked to commit. Substrate independence is not a stylistic preference — it is the only way to reach reliable inter-model agreement on the same claims.
Empirical proof. Upgrades the axiom from theory to preregistered result at n=30. - 03A context-window-free agentic coding CLI
The release report for polycode v1 (v1.0.0 through v1.1.7). Describes what polycode is, the five-phase event loop, the SHA-256 chained canon as a user-owned replayable ledger, and the v1.1.x shipping path through cross-platform purity, zero-config hosted inference, vision, web search, the masked /key flow, bracketed paste mode, the identity gate, the single-stdin-owner fix, and bracketed-paste terminal state ownership. Install: npx @polylogicai/polycode@latest.
The product. What shipped, how to install it, what it does. - 04The theoretical extension of Engine, Rules, and Canon and a fifteen-test validation
The three-primitive architecture needs a fourth: compile(canon, query) → bounded reasoning packet. A disjoint-substrate LLM does the row selection, a deterministic non-LLM assembler materializes the packet, and the witness stack extends to seven non-LLM primitives including the brand-identity gate. Fifteen-test harness, nine tests reported verbatim from v1.1.0 with structural extensions through v1.1.7.
The theory. Closes the loop: every claim in this program is now installable with one command, and the runnable form is described by the companion release report.
The runnable
Every paper ships with a product you can install.
polycode is the first. It consumes the compile primitive from paper four and runs an agentic coding loop on your machine with a free hosted tier. One command, no setup.
# install and run
$ npx @polylogicai/polycode@latest
How we work
Research → Validate → Publish → Ship.
The same loop every product goes through. No product ships without a paper. No paper ships without a disjoint verifier.
STEP 1
Research
Every product starts from a question we cannot answer with a chatbot. We preregister the question, the protocol, and the falsifier.
STEP 2
Validate
We run the protocol against a disjoint-substrate verifier. Substrate cannot witness itself, so the verifier is never the generator.
STEP 3
Publish
Paper to Zenodo under CC-BY-4.0. DOI, dataset, and methodology are open. The full engine stays proprietary so we can keep moving.
STEP 4
Ship
The validated primitive becomes a product. polycode is the first. Every product ships with a published paper.