Pangram Model Architecture and Multi-Objective Training

The underlying software architecture has transitioned across iterations to support increasingly complex text inputs. Its modern core (manifested in Pangram 4) is built upon a large, open-weight Mixture of Experts (MoE) backbone model adapted for sequence classification. The system attaches independent, custom linear classification heads to the final sequence position of the shared backbone, exploiting causal attention mechanisms where the final hidden state vector ($\mathbf{h}_S$) retains a complete contextual representation of the input window.

The system achieves granular, single-pass evaluation by simultaneously optimizing for multiple objectives across distinct classification heads:

• 

• Segment-Level Edits: Evaluates localized adjustments within the passage.

• Mixed-Authorship Binary Classification: Determines whether a document is fully human-written, fully AI-generated, or hybrid.

• Humanizer Detection: Flags signatures typical of commercial obfuscation or adversarial paraphrasing algorithms.

• Tokenwise Provenance: Projects predictions down to individual token positions using a localized sequence head. To allow every supervised token to utilize context from the complete source sequence under a causal backbone, the framework implements a context replication format known as Repeat2, where each 512-token training window is repeated twice and loss calculations are applied only to the second instance.

• 

The software stack utilizes standard deep learning frameworks, specifically PyTorch and Hugging Face libraries, optimized via Parameter-Efficient Fine-Tuning (PEFT) methodologies like Low-Rank Adaptation (LoRA). Training is performed across distributed clusters of hardware, such as NVIDIA H100 GPUs, to handle the vast parameter scale required to map modern frontier models.

Prediction Mechanics and Platform Integration

When raw text is passed to the platform via its user interface or REST API, the system does not emit an arbitrary boolean result. Instead, it tokenizes the string, maps the tokens to vector embeddings, and processes them through the neural network to output continuous numerical scores representing spatial coordinates in "Pangram Space".

The system segmentizes documents longer than a specific threshold (e.g., 450 tokens) to assess moving windows individually. The resulting output maps text into calibrated probabilistic thresholds:

Score Range Classification Category

$\le 0.25$ Human-Written

$0.25 < \text{Score} < 0.50$ Lightly AI-Assisted

$0.50 \le \text{Score} < 0.75$ Moderately AI-Assisted

$\ge 0.75$ Fully AI-Generated

The platform's software engineering emphasizes broad downstream availability to achieve its mission of content validation across the broader internet ecosystem. The core model is exposed via high-throughput API endpoints priced dynamically by word count metrics. To operationalize these capabilities directly within existing workflows, the software is deployed via deep software integrations into learning management systems (such as Canvas LMS and Google Classroom), digital publishing platforms (such as Substack), browser extensions for real-time web monitoring, and document verification add-ons like Google Docs.

Through this multi-tiered implementation—spanning structured data engineering, sophisticated multi-head transformer architectures, and extensive platform integrations—Pangram establishes a deterministic legibility framework to handle the challenges of mixed-authorship content at scale.