JITNA Protocol
Just-in-Time Neural Architecture — RFC-001 v2.0 — the open routing protocol that selects the optimal LLM, algorithm tier, and voting method for every task in real-time, under 50ms.
What is JITNA?
JITNA (Just-in-Time Neural Architecture) is an open protocol defining how AI systems should select, route, and coordinate multiple language models. Inspired by JIT compilation in software engineering — routing decisions made at exactly the right moment, not at compile time.
In the RCT Ecosystem, JITNA executes in the ROUTE state of every IntentLoop cycle — receiving the parsed IntentObject from the PARSE state and making real-time decisions on which models, algorithm tiers, and voting methods to invoke.
6 JITNA Primitives
Every JITNA routing decision is built from these 6 primitives — the fundamental variables that compose the routing engine.
The structured goal extracted from the input — what the user or system wants to achieve. JITNA parses the raw input into an IntentObject that captures goal, constraints, and priority.
The current state of available information — retrieved from RCTDB, external APIs, or context window. Represents what the system actually knows at routing time.
The gap between Intent and Data (Δ = I − D). This delta drives the routing decision — a large delta means the system needs more capable models; a small delta allows lightweight routing.
The selected algorithm tier and voting method — how JITNA will process the intent given the current delta. Chosen from 41 algorithms across 9 tiers based on task classification.
Self-evaluation of the output quality — SignedAI consensus scores across 8 dimensions. R is the feedback signal that informs whether the Approach succeeded or needs escalation.
Persistence of routing weights, proficiency scores, and outcome history in RCTDB 7D. M is what makes JITNA self-improving — each task updates the routing model for future decisions.
JITNA formula: Δ = I − D → select A → Execute → R → update M
6 Factors JITNA Uses to Route
Every routing decision is made in <50ms by evaluating all 6 factors simultaneously.
JITNA classifies the incoming intent (ANALYTICAL, CREATIVE, LEGAL, REGIONAL, etc.) and maps it to the best-suited model role.
Each model has domain proficiency scores (0.0–1.0). JITNA uses these to weight model selection — Typhoon G38 scores th=0.99 for Thai tasks.
Tasks with tight SLA requirements route to faster models (Grok 4.1 Fast, Gemini Flash). Complex tasks can use slower, more capable models.
Simple queries route to cheaper models automatically. Complex tasks that need SignedAI consensus are escalated to full multi-model mode.
Tasks flagged as regional (Thai, Japanese, etc.) are automatically routed to the matching regional model — Typhoon G38 for Thai tasks.
After each ADAPT cycle, JITNA routing weights are updated based on outcome quality — the router continuously improves itself.
JITNA RFC-001 v2.0 is the open specification defining: the Routing Decision Tree, IntentObject Schema, Model Proficiency Score format, Voting Method eligibility rules, and Feedback Weight Update algorithm.