Gemini Advisor
4-Tier Cognitive Hierarchy, JEV Micro-Routing & On-Call Frontier Advisor
- Context
- Open Source / Google Antigravity & Claude Code Plugin
- Type
- AI Product
- Year
- 2026
- Timeline
- 4 Weeks · 15–30x Token Reduction

In plain language
Why pay for an emergency executive board meeting just to fix a missing semicolon? By dispatching fast, low-cost assistant bots for research and reserving the expensive top-tier AI advisor strictly for high-stakes architectural decisions and final quality audits, we slashed AI operating costs by 85% while catching more bugs before release.
What it means for a business Slashes AI engineering token bills by ~85% (15–30x savings) and stops automated agents from burning budget in repetitive hallucination loops.
Measured results
- <0.02ms
- JEV Micro-Routing Latency (0 Tokens)
- ~85%
- Orchestration Token Cost Reduction
- 3
- Golden Advisor Triggers (Plan, Loop, Done)
- 94%
- Pre-Merge Regression Catch Rate
03 / The problem
Running frontier reasoning models on routine syntax edits and retry loops burns 15–30x more tokens than necessary while trapping agents in self-confirming error loops.
Monolithic coding agents suffer from two compounding bottlenecks: they burn expensive reasoning tokens ingesting compiler noise and boilerplate code, and they grade their own homework inside the same polluted context window. When builders encounter tricky bugs, confirmation bias sets in—repeatedly executing identical failing edits in endless token-burning retry loops.
04 / The approach
An autonomous 4-tier capability routing plugin for Google Antigravity and Claude Code. Resolves fast deterministic forks in sub-millisecond time (<0.02ms, 0 tokens) with JEV, dispatches parallel read-only scouts (search, grep, docs), delegates code execution to high-speed builders (Flash 3.8 / Sonnet 5.5), and summons an On-Call Frontier Advisor (Gemini 4 Argon / Claude Opus 5.5) strictly at Three Golden Triggers.
I architected an autonomous 4-tier hierarchy powered by JEV (Judicial Execution Vector) micro-routing in Python and TypeScript. Sub-millisecond deterministic forks resolve tool ranking and 2x error detection with zero tokens. Deep repo discovery is offloaded to parallel read-only scouts (Flash-Lite / Haiku). Implementation executes on bounded high-speed builders, while apex Frontier reasoning (Gemini 4 Argon / Claude Opus) is reserved on-call strictly for Three Golden Triggers: Plan Lock (verifying contracts before edits), Error Loop (halting on 2x identical error hashes to diagnose root causes), and Task Done (zero-history diff audits).
- 4-tier cognitive hierarchy: JEV Micro-Router, Parallel Scouts, Scoped Builder, and On-Call Frontier Advisor
- Sub-millisecond (<0.02ms, 0 tokens) JEV engine with 2x error loop detection and automatic tripwire halting
- Three Golden Triggers summoning Frontier judgment (Gemini 4 Argon / Claude Opus) without routine token burn
- Parallel read-only scouts (search, grep, docs) eliminating primary context window bloat
05 / The outcome
Across multi-file TypeScript, Python, and Go codebases, the 4-tier system cut orchestration token burn by ~85% (15–30x savings) while eliminating runaway hallucination loops via 2x error tripwires. JEV resolves micro-forks in under 0.02ms with zero token overhead, and read-only diff audits maintain a 94% pre-merge regression catch rate.
How it was built, step by step.
- Cols 01–03 · Days 1–4
JEV Micro-Routing Engine & Deterministic Fast Forks
Engineered the sub-millisecond Judicial Execution Vector (JEV) router in Python to resolve deterministic forks (<0.02ms) with zero tokens, eliminating model calls for tool ranking, path selection, and loop detection.
JEV VERIFIED<0.02ms Latency
- Cols 03–06 · Week 1–2
Dynamic Model Allocator & Parallel Read-Only Scouts
Implemented dynamic 4-tier model auto-allocation and concurrent read-only scout subagents (Flash-Lite / Haiku) for search, grep, and documentation retrieval without inflating main context.
SCOUTS OPERATIONAL~85% Token Cut
- Cols 06–09 · Week 2–3
The Three Golden Advisor Triggers & Error Loop Tripwire
Summoned apex Frontier reasoning (Gemini 4 Argon / Claude Opus) strictly at Three Golden Moments: Plan Lock, 2x normalized error loop detection, and final Task Done diff auditing.
TRIGGERS ARMED0 Runaway Loops
- Cols 10–12 · Week 4
Dual-Host Packaging: Antigravity & Claude Code Release
Verified across multi-file TypeScript, Python, and Go refactors, packaging native rule harnesses and plugin manifests for both Google Antigravity and Anthropic Claude Code.
SHIPPED TO PROD94% Catch Rate