[ 02 ]
ARTICLES

Writing on model integrity.

What actually changes when you fine-tune, quantize, or ship a language model — and how to know before your users find out.

PINNED
[ 02 ] · June 2026

Pixelated dot-matrix rendering of two hands reaching toward each other

Model Integrity Testing Is Not Red Teaming, Evals, or Guardrails

Three reasonable guesses. All three wrong in instructive ways.

READ →

Sort/

[ 18 ]
September 2026

Network graph of images connected by lines radiating from a central node, evoking a dependency chain that can be quietly redirected

Your Model's Name Isn't a Security Guarantee

Hugging Face namespaces can be re-registered by anyone once the original owner deletes their account. Researchers proved it with a live reverse shell.

READ →

[ 17 ]
September 2026

Curved wall of hundreds of illuminated screens surrounding a dark silhouette, evoking constant monitoring

AI Safety Evaluation vs. Adversarial Robustness: What OpenAI's Astra Gets Right (and Where It Gets Hard)

Researchers are still uneasy about it. Here's the gap between passing a test and being trustworthy.

READ →

[ 16 ]
August 2026

Illustration of a human and a humanoid robot arm wrestling, evenly matched

The Harness Assumes the Model Holds Still

A harness limits what a failure can reach. It does not tell you the failure is new.

READ →

[ 15 ]
August 2026

White dove in flight dissolving into digital glitch artifacts against black

Your Red Team Tested a Model. Your Hospital Deployed a Different One.

A peer-reviewed review gives the problem a name: temporal mismatch risk.

READ →

[ 14 ]
August 2026

Thermal camera view of a crowd, each person boxed and labelled with a temperature reading

AI Is Scaling Where It Is Easiest to Govern, Which Is Not the Same as Where It Is Safest

The governance budget already exists. It is being spent on watching, not testing.

READ →

[ 13 ]
August 2026

Long-exposure photograph of blue and amber light blocks streaking toward a dark vanishing point

Your Buyer Is About to Ask Who Tested the Model You Shipped

Governance is turning into a due diligence step, and most of the standard questions have no good answer if your model was modified after release.

READ →

[ 12 ]
August 2026

Cyan terminal session filled with a live process table and system monitors on a black screen

The Refusal Test Everyone Is Running Is a String Match on the First 128 Tokens

Nobody is validating these models, including the people releasing them.

READ →

[ 11 ]
August 2026

Grainy surveillance frame of a walking figure overlaid with orange detection boxes and tracking points

Capability Evaluations Are Not Safety Evaluations

Your model still works. That is not a safety result.

READ →

[ 10 ]
August 2026

White circuit-board traces radiating from a single blank chip on a black field

Quantization and Safety Drift: Why the Model You Ship Is Not the Model You Tested

The artifact gap is one of the least examined steps in the deployment pipeline.

READ →

[ 09 ]
August 6, 2026

Dense white mathematical notation receding into a black perspective field, like a corridor of equations

AI Model Hacked During Testing: Why the Harness Is the Real Risk

As models become more agentic, eval environments and control layers are becoming the primary failure point.

READ →

[ 08 ]
August 5, 2026

Sparse node graph on black, labelled boxes of numeric values wired together in acid green

Red Hat's asago and the Open Question of Lifecycle Testing

Red Hat's new open source AI governance project automates policy-to-deployment. A look at what it covers, and at what the research literature actually says about safety behavior after fine-tuning and quantization.

READ →

[ 07 ]
August 2026

Glitched wireframe architecture dissolving into vertical bands of green and orange light

Your Fine-Tuned Model Is Less Safe Than the One You Started With

You just haven't measured it yet.

READ →

[ 06 ]
July 2026

A human eye rendered as a coarse black-and-white dither, close enough to see the pixels

Post-Training Isn't Done When the Loss Curve Flattens

A convergent loss curve means training stopped — not that the model is ready to ship.

READ →

[ 05 ]
July 2026

A dense white lattice of nodes and edges layered over itself against black

From Base to Fine-Tuned to Quantized: A Lifecycle View of Model Integrity

Integrity isn't a property of a checkpoint. It's a property of a lifecycle.

READ →

[ 04 ]
July 2026

Thousands of white filaments converging on a single point against black, like a distribution collapsing

Quantization-Aware Safety Drift: Why INT4 Breaks Your Aligned Model

Aligned models can lose their refusals at INT4 — while every capability benchmark stays flat.

READ →

[ 03 ]
June 2026

Split image: real butterfly on the left, ASCII-rendered butterfly on the right

The Release Gate Your Model Pipeline Is Missing

Code does not reach production without passing tests. Models do.

READ →

[ 01 ]
June 2026

Abstract distorted figure rendered in dark tones

Why Base-Model Benchmarks Fail After Fine-Tuning

The benchmark describes a checkpoint that no longer exists.

READ →

START FREE

If any of this describes your pipeline, SichGate runs the adversarial battery and gives you the differential before you ship.

START FREE ASSESSMENT →