ISSL
A deterministic pattern-matching engine for conversational systems — built to catch high-confidence responses before they ever reach an LLM.
in development
rust
A deterministic pattern-matching engine for conversational systems — built to catch high-confidence responses before they ever reach an LLM.
ISSL defines what a system should say in response to what a user asks — no training data, no probability distributions, no hallucinations. Think of it as a composable, auditable guardrail layer that sits in front of an LLM.
When a user asks a question, ISSL checks its lexicon of pattern→response rules first. If it finds a high-confidence match, it responds instantly and deterministically. If not, it falls through — either to a wildcard catch-all, or to an optional LLM fallback. The model is an extension, not the default.
Built as the scripting language behind JAIME, an AI persona running on the Rashell command processor for Windows. Pattern matching was the only intelligence available.
Ported to C# as part of the Rashell ecosystem. Gained modular includes, scoped variables, parameter files, and the classic IgnoreList matching engine with its characteristic quirks.
The third rewrite. Seven matcher modes (classic through TF-IDF retrieval), pattern weights, training data export, and an optional AI fallback bridge. Designed to be embedded — in agents, in sub-agents, in browsers via WASM.
At its core, ISSL is a declarative mapping from keywords to templates. Each pattern defines what to look for and what to say back:
{pattern} {q}who;created;you;made;built{/q} {r}I was created by [$user];[$user] built me{/r} {/pattern}
The runtime scores every pattern against the user's input, picks the best match,
expands [$user] from self-memory, and picks a response alternative.
The entire decision is traceable — you can see exactly which pattern fired, what
score it got, and why.
Classic (C# parity), normalized, fuzzy (typo-tolerant), semantic (synonym-aware), hybrid cascade, and TF-IDF retrieval.
Bias matching with (weight:2.0) inside any pattern block. High-priority responses get higher scores.
Scripts include other scripts. Persona, guardrails, and FAQ modules compose into a single runtime.
When no pattern matches, fall through to a controlled response — or to a real LLM, gated by policy.
Export patterns as JSONL or OpenAI fine-tuning messages. Train a model that speaks in ISSL's voice.
In-memory hashed vector matching — fully offline, deterministic, no GPU, no API key.
Most AI systems treat determinism as a limitation to overcome. ISSL treats it as the whole point. When a response must be correct — brand voice, compliance, security boundaries, identity — you don't want probability. You want a rule.
ISSL doesn't replace the LLM. It protects it. The model handles the open-ended, creative, unscripted world. ISSL handles everything that shouldn't be left to chance.
ISSL is in active development as a closed-source research project at J.C.P Laboratory. It serves as the planned guardrail layer for Astra Agent and its sub-agent fleet. No public timeline for release — we're building it right, not fast.