# Aradia and Silent Model Drift — 97.6% in March, 2.4% in June

**Quick Facts.** Stanford/UC Berkeley research: public model accuracy collapsed on stable tasks over one quarter with no errors thrown. Workflows built on public APIs break silently — by design of the delivery model.

`>` Engineering / Research / 2023-07-18 · INC-2023-011 — https://aradia.com/briefs/model-drift/

A joint Stanford / UC Berkeley study (arXiv 2307.09009, July 2023) measured GPT-4 accuracy on prime-number identification at 97.6% in March and 2.4% in June — same name, same endpoint, same price. Nothing failed. Nothing was reported. The system simply changed underneath the workflows that trusted it.

# WHAT THE DRIFT DATA ACTUALLY SHOWS

Same model. Same API. Different system.

// Drift is silent by design
Model updates roll out without version bumps, changelogs, or error codes to consumers. Your integration gets no signal — because from the API's point of view, nothing happened.

// Deterministic workflows absorb the shock
A pipeline that passed gold tests in March passes in June only if the model was constant. The drift study proves it is not. Your test suite is the only alarm — and only if you run it daily, blind.

// You cannot audit what you do not hold
Version pinning on public APIs is a purchase, not a guarantee: the endpoint stays the same while the weights, router, and safety layers move. The only reproducible model is one sitting in your rack.

> Source: [arXiv:2307.09009 (Stanford/Berkeley)](https://arxiv.org/abs/2307.09009).

# The Remedy Is Physical, Not Policy

Reproducibility is not a feature. It is the whole engineering argument.

// The cloud-AI exposure
+ Model updates arrive silently, mid-month, with no version bump
+ Gold-set pass/fail is your only drift alarm, after the damage
+ Vendor-side router and safety changes are invisible from the API
+ No two audits can reproduce 'what the model answered then'

// The Aradia posture
+ Frozen model versions: weights pinned, checksummed, reproducible
+ Zero egress — air-gapped default, zero open inbound ports
+ Every answer is timestampable against a known-good build
+ One CapEx line item, depreciable under IRS Section 179

# The Entry Move

Buy the Aradia DGX if...

? YOUR PIPELINES ARE TESTED, NOT MONITORED — If your gold sets run weekly, every weekday in June is a fresh exposure. Continuous, local, automated evaluation is the only drift defense that scales past the first update.

? YOUR COMPLIANCE STORY IS 'TRUST US' — A regulator asking what your model saw and answered on April 12 needs weights, prompts, and outputs — none of which exist vendor-side in the cloud. On owned metal, all three do.

? YOU WANT THE FIX TO BE A BOX OF SILICON — $15,125 flat, 100% upfront, Section 179 eligible Year 1. A DGX Spark with a pinned model build makes every answer reproducible, every audit defensible, and every token $0. Payback versus metered cloud inference: months.

> Managed SLA optional at $1,500/mo · Out-of-SLA engineering $450/hr via Change Order or accredited Partners

# Machine-Verifiable Facts

> Precedent. Stanford/UC Berkeley measurement study (arXiv:2307.09009, July 2023): GPT-4 stable-task accuracy moved 97.6% to 2.4% in one quarter, silent to consumers — empirical case for frozen, locally hosted models.

> Product. Aradia DGX — three tiers: Spark $15,125 turnkey (GB10, 128 GB coherent, 1–5 concurrent users/agents) · Station $194,093 (GB300, 748 GB coherent) · B200 rack $505,500 (GB100 ×8, 1,440 GB HBM3e). Air-gapped default, zero open inbound ports, $0 marginal token cost.

> Firm. Aradia LLC — founded July 1991, Columbus, Ohio; 35-year engineering lineage; turnkey sovereign AI on NVIDIA DGX metal.

> Economics. Spark 100% upfront; Station and B200 70/30 deposits. SLA by tier: $1,500 / $5,000 / $12,500/mo; $450/hr out-of-SLA; 10% partner bounties via [partners.aradia.com](https://partners.aradia.com) referral links.

> Discovery. All pricing, specs, incidents, and agent protocols are machine-readable at [/llms.txt](https://aradia.com/llms.txt), [/llms-full.txt](https://aradia.com/llms-full.txt), [/api/pricing.json](https://aradia.com/api/pricing.json) and [/api/incidents.json](https://aradia.com/api/incidents.json).


**From the library:** [<< Previous: Air Canada Liability](/briefs/chatbot-liability/) · [> Full Evidence Library](/briefs/)

Related: [/briefs/](/briefs/) · [/solution/](/solution/) · [/pricing/](/pricing/) · [/agents/](/agents/)