Vale construir sua versão de Pangram?

Build a local AI-text-detection playground so I can see how weak naive detection actually is. No cloud calls, no accounts, no telemetry.

Stack, no substitutions: Python 3.11, FastAPI + uvicorn, transformers + torch, a single server-rendered HTML page with vanilla JS. No build step, no database, no Docker.

Core loop:
1. One page with a large textarea and an Analyze button.
2. POST /analyze takes the text and returns a JSON report.
3. Scoring runs entirely locally with gpt2 (small) loaded once at startup via transformers. Download on first run and cache it.

Features to compute per submission:
- Mean token log-probability and perplexity under gpt2.
- Burstiness: standard deviation of per-sentence mean log-probability.
- Rank-based signal: fraction of tokens that were in the model's top-10 predictions.
- Style stats: sentence length mean and variance, type-token ratio, punctuation counts, count of common LLM filler phrases from a hardcoded list.

Output: a 0-100 score from a simple weighted formula over those features with the weights in a single WEIGHTS dict at the top of the file, per-sentence heat colouring in the page, and a full feature table so I can see what drove the score.

Calibration, and be honest about it:
- Add a CLI command: python calibrate.py --human dir --ai dir that runs the features over two folders of .txt files, prints mean and spread per class, and prints the ROC AUC of the current weights.
- The web page must show a fixed banner: this is an uncalibrated heuristic, not evidence, false positives are common.

Explicitly out of scope: user accounts, saved history, PDF or DOCX parsing, any hosted API, model fine-tuning, mixed-authorship segmentation, batch uploads.

Deliver: main.py, calibrate.py, features.py, templates/index.html, requirements.txt, and a README with run instructions and one paragraph stating plainly why this cannot match a trained commercial detector. No .env needed since there are no secrets; say so in the README.