presence: sync card

README.md

@@ -20,17 +20,14 @@ short_description: We are Authentrics.ai
 [![Python](https://img.shields.io/badge/python-3.11--3.13-3776ab)](https://app.authentrics.ai/)
 [![Runs locally](https://img.shields.io/badge/analysis-100%25%20local-16a34a)](https://app.authentrics.ai/)
 
-# Authentrics.ai — checkpoint-level drift detection and correction, without retraining
+# Authentrics.ai
 
-Training runs are expensive and largely irreversible: dataset composition, hyperparameters, and training
-order all shape the outcome, and problems often aren't visible until a final evaluation — by which point a
-full retrain looks like the only fix. Authentrics compares the checkpoints a training run already saves to
-identify what changed, and when. When unwanted behavior traces back to specific training session(s), that
-effect can be removed directly from the checkpoint. Separately, a checkpoint can be optimized directly
-against a defined objective — recovering accuracy, correcting a regression, suppressing unwanted behavior —
-without backprop or retraining. Neither path requires starting the run over.
+**Find what changed. Find when it changed. Fix it without retraining.**
 
-## What the tools do
+Compare training checkpoints to diagnose regressions, trace unwanted behavior to training intervals,
+remove training influence, and optimize model behavior directly.
+
+## What You Can Do With Authentrics
 
 - **Checkpoint-to-checkpoint comparison.** Parameter and activation drift measured between the checkpoints
   a training run already saves — no extra instrumentation.
@@ -53,7 +50,7 @@ Everything runs locally on your own hardware; only project metadata leaves your
 
 Each demo is a reproducible ML problem: a model that failed in a specific way, and how we found and fixed it.
 
-### Finding and Removing a Neural Network Backdoor
+### Find Which Training Step Introduced a Backdoor — Then Remove It
 
 *Backdoor / data poisoning (BadNets)*
 
@@ -65,7 +62,7 @@ A LeNet-5 MNIST classifier is trained with one poisoned epoch that plants a BadN
 - [Model on Hugging Face](https://huggingface.co/authentrics/mnist-badnets-poisoned)
 - [Corrected model](https://huggingface.co/authentrics/mnist-badnets-repaired)
 
-### Recovering a Model That Forgot Its Original Task
+### Recover a Capability Lost During Fine-Tuning
 
 *Catastrophic forgetting / task interference*
 
@@ -76,7 +73,7 @@ Fine-tuning Gemma 3 1B on a new task quietly degraded a capability it already ha
 - [Source code (GitHub)](https://github.com/Authentrics-ai/authentrics-model-analysis-experiments)
 - [Model on Hugging Face](https://huggingface.co/authentrics/ztom-gemma3-1b-recovery)
 
-### Removing Memorized PII From a Fine-Tuned Model
+### Remove Memorized PII Without Retraining the Model
 
 *Memorization / PII leakage*
 
@@ -87,16 +84,16 @@ A medical chatbot fine-tuned on data containing personal information memorized t
 - [Source code (GitHub)](https://github.com/Authentrics-ai/authentrics-analysis-examples)
 - [Model on Hugging Face](https://huggingface.co/authentrics/pii-removal-medical-llama32)
 
-### Checkpoint Diagnostics
+### Understand What Changed Inside Your Model
 
-**Reading How a Mixture-of-Experts Model Routes Its Inputs** — *Interpretability / MoE routing*
+**See Which Experts a Mixture-of-Experts Model Actually Uses** — *Interpretability / MoE routing*
 
 Which experts does Mixtral-8x7B actually use, and how does that routing structure emerge across training? We read the routing and per-layer influence without touching the weights remotely.
 
 - [Source code (GitHub)](https://github.com/Authentrics-ai/authentrics-model-analysis-experiments)
 - [Model on Hugging Face](https://huggingface.co/authentrics/mixtral-expert-routing)
 
-**Diagnosing What Changed Across a Model's Training Run** — *Training dynamics / drift diagnosis*
+**Find Where a Model Changed Across Its Training Run** — *Training dynamics / drift diagnosis*
 
 Across the NVIDIA Nemotron Cascade 8B checkpoint lineage, what changed inside the model — where, when, and how much? We track parameter and activation drift checkpoint to checkpoint.
 
@@ -106,7 +103,7 @@ Across the NVIDIA Nemotron Cascade 8B checkpoint lineage, what changed inside th
 **Demonstrates:** checkpoint-to-checkpoint comparison (`static_analysis`, `activation_analysis`, `correlation_analysis`)
 
 
-## Get started
+## Try It on Your Model
 
 ```bash
 pip install authentrics            # Linux x86_64, Python 3.11–3.13