Perslis · Shells
One model. The rest are shells.
Perslis has no language model of its own. What we build are shells: the tools, memory and checks a language model needs to do real work. Our first model is Peel, the Fail-First Model.
Peel is our first model, and it is not a language model. It is the Fail-First Model: it learns from its failures, and nothing it learns can widen what it is allowed to do. Everything else here is a shell around whatever language model you bring. Perslis Symbols turns a question about your data into a signed, exact procedure your AI assistant calls with no model at answer time. Kist is the loop that takes a goal to a verified result. The science shell keeps every fact tied to the source that supports it and answers unknown when the evidence is missing. The language models themselves are listed in the roster.
Peel · our first model
The Fail-First Model. Learns from failure without learning around its safety floor.
Our first model ↓Perslis Symbols · shell
Ask a model once. Keep a symbol that answers forever, signed and exact.
The Symbols shell ↓Kist · shell
The loop: plan, act, check, and go round again until the work is proven.
The Kist shell ↓Science · shell
Typed knowledge with its evidence attached, and reasoning that can say unknown.
The science shell ↓Peel — our first model, the Fail-First Model
A fail-first model expects to fail and learns only from failures it has actually had. Peel is the first one we know of that combines all three defining properties. When the evidence is missing it fails closed and says unknown. When it fails, it writes the failure down, charges it to the decision that caused it, and builds a rule. That rule can take options away. It can never add one.
Peel has no neural network at its core and no training run. Its knowledge is typed, sourced cards, and its learning is a table of counts you can read.
The loop is fail, observe, explain, build rule, verify, retry. The same model plays games in the arcade. The Fallout pilot learns from its own deaths and writes each lesson onto a card in its journal. The Doom pilot takes orders such as "only use the shotgun" as limits on what it may do. Neither ever gains an action its orders and sources did not already allow. VDSG applies the same model to defense.
Peel, the first Fail-First Model → What is a fail-first model? The math → The white paper (PDF) ↓ Field notes from the arcade →
Perslis Symbols — a shell your assistant calls
Most questions people put to a model inside software are arithmetic over data they already hold. Perslis Symbols moves that work out of the model. You bring a question about your data. A model composes a specification for it from a closed vocabulary of named steps: rows, filter, join, group by, a reducer. It never writes code. The Perslis floor runs ten mechanical checks on the specification, and a person reads it and signs the approval. From then on your AI assistant, Claude Code or any other client that speaks MCP, calls the symbol as a tool, and your machine answers exactly, with model_calls: 0 and the derivation attached.
- Exact: decimal arithmetic end to end, so 2⁵³ + 1 stays 2⁵³ + 1.
- Refuses by name: when the data cannot support an exact answer, the tool says why instead of returning a plausible number.
- Local: the runtime is standard-library Python with no network code. Any MCP client can call it, including Claude Code.
Status: pilot, runtime 1.1.2. The admission gate stays at Perslis; the runtime is free to download.
Perslis Symbols, with a live demo → How it works, the paper → Join the beta →
Kist — the loop shell
A single prompt and response is a guess. Kist is the loop that turns a goal into a result that has been checked. It sizes the goal, plans it as ordered steps, acts through real tools, verifies what came back, and feeds every failure into the next plan. It stops only when the work is verified or a budget, time or cycle cap is reached. Steps that cannot be undone wait for approval. Anything high-stakes goes before a council of models from different labs before Kist reports it done.
Kist works with whatever model you point it at: a cloud model on your own key, a local model on your own hardware, or both. The model supplies the reasoning. Kist supplies the tools, the memory and the checks, and it is the engine the rest of Perslis runs on.
The Kist shell, step by step → The loop, from goal to deployed result →
The science shell
Science is where a confident guess does the most damage, so the science shell is built to keep guesses and facts apart. Three parts, each with its own job:
- Peel's knowledge: structured records whose typed relations carry the evidence behind them, so a claim can be traced to where it came from.
- Lois, the reasoning: a symbolic engine that follows recorded relations, shows its path, and returns unknown when the graph has no support, instead of a guess.
- The instruments: literature, proteins and structures, genomes and variants, chemistry and pathways, queried live from nineteen public services in one console, with every answer pinned to its database, accession and URL.
Status: pilot program running. The bioscience applications are proposed directions; no discovery or clinical result is claimed.
Peel, the science runtime → The instrument wall → Try the demo →
The language models inside the shells
The shells decide and check. The reasoning inside them comes from language models, and Perslis does not bet on one. The roster lists every model Perslis drives, American and Chinese frontier models, airtight local models, and the role each one plays.