REPRODUCIBLE LIVE EVALUATIONS · STREAMABLE HTTP MCP 2025-03-26

Oxyo vs. Gemini Argon — & What Happens When You Connect Both

Frontier models like Gemini Argon (High) are remarkable single-window reasoners—until a chat resets, an architectural rule changes 45 days later, or a financial/combinatorial audit demands cent-exact deterministic execution instead of next-token probability. Below are 4 live, reproducible evaluations executed against production https://oxyo.si and https://oxyo.si/mcp.

Baseline · Raw Frontier LLM

Gemini Argon (Stateless)

0 / 4 Exact

Strong prose and algorithmic planning, but 0% cross-session recall on New Chat, vulnerable to stale rules when raw logs are pasted, and drifts on unexecuted 36-row Bankers' Rounding & 22-step DP counts.

Standalone · oxyo.si

Oxyo Super Intelligence

4 / 4 Verified

Compiles project state into persistent XY Informational DNA, automatically supersedes contradicted specs on write, and executes pure Python inside an isolated sandbox before speaking.

Connected · https://oxyo.si/mcp

Gemini Argon + Oxyo MCP

4 / 4 100% Synergy

Connects Gemini Argon (Antigravity), Cursor, or Claude Code directly to https://oxyo.si/mcp—combining frontier synthesis with ~110 ms contradiction-free memory, causal graphs & verified skills.

Evaluation Suite Gemini Argon (Stateless) Oxyo Standalone (oxyo.si) Gemini Argon + oxyo.si/mcp
01 · 45-Day Architecture Drift
7 sprint updates · 3 superseded rules · fresh session		 ✗ 0 / 4 recalled in a new session (or stale port 5432 / 2.9% collision if raw logs pasted) ✓ 4 / 4 active rules recalled in 1.94 s · 0 re-pasted tokens · 3 contradictions retired ✓ 4 / 4 exact via oxyo_recall (126.7 ms HTTPS) + full audit trail via oxyo_belief_revisions
02 · Multi-Currency Ledger Audit
36 webhook events · FX Bankers' Rounding · orphan refund		 ✗ Spots high-level rules, but single-pass mental arithmetic drifts on cent rounding & orphan evt_1014 ✓ $11,308.63 net settlement & $520.30 EU VAT escrow proven in isolated Python sandbox ✓ $11,308.63 & $520.30 verified via oxyo_verify_python in 250.1 ms + executive report
03 · Constrained 22-Step Automaton
Ternary frames (no 00/111, sum mod 5 = 0) + prime digest		 ✗ Writes valid DP recurrence, but cannot evaluate 22 × 15 state transitions in raw attention tokens ✓ 57,569,645 frames & digest 821,335,879 (verified: true in sandbox) ✓ 57,569,645 & 821,335,879 proven via oxyo_verify_python in 213.4 ms
04 · 4-Hop Incident & 0-Token Skills
Upstream root-cause diagnosis + unit-tested SRE tool		 ✗ Zero memory of prior postmortem topology; re-writes boilerplate SRE math from scratch each chat ✓ Persistent Bayesian causal chain + reusable unit-tested Python skill vault ✓ 4-hop root cause traced in 104.1 ms (oxyo_causal_infer) + 0-token skill run in 90.1 ms

Detailed Test Protocols, Prompts & Live Traces

⬇ Download q4_gateway_ledger.csv (Test 02 Input)
30-SECOND REMOTE MCP SETUP · ZERO INFRASTRUCTURE TO MANAGE

Connect https://oxyo.si/mcp to Gemini Argon (Antigravity), Cursor, or Claude Code

Give your favorite IDE or frontier agent all 9 persistent Oxyo tools (oxyo_recall, oxyo_remember_belief, oxyo_belief_revisions, oxyo_verify_python, oxyo_causal_record, oxyo_causal_infer, oxyo_skill_register, oxyo_skill_invoke, and oxyo_remember_episode) over encrypted Streamable HTTP.

Generate oxy_live_... Key →
ANTIGRAVITY (GEMINI ARGON) · CURSOR · WINDSURF (mcp_config.json) 
{
  "mcpServers": {
    "oxyo": {
      "url": "https://oxyo.si/mcp",
      "headers": {
        "Authorization": "Bearer oxy_live_YOUR_API_KEY"
      }
    }
  }
}
CLAUDE CODE CLI & STREAMABLE HTTP HANDSHAKE VERIFICATION 
# 1-line Claude Code registration
claude mcp add --transport http oxyo https://oxyo.si/mcp \
  --header "Authorization: Bearer oxy_live_YOUR_API_KEY"

# Verified Live Handshake Telemetry (oxyo.si/mcp):
# Protocol: 2025-03-26 (Streamable HTTP JSON-RPC 2.0)
# initialize:  200 OK · 131.0 ms
# tools/list:  200 OK · 106.4 ms · 9 tools active
# oxyo_recall: 200 OK · 126.7 ms · TLS isolated vault