# CHECKPOINT COMPATIBILITY: Audit of `qtensorformer_small.pt`

**Date**: September 17, 2026  
**Status**: **`CHECKPOINT_INCOMPATIBLE`**

---

## 1. Executive Summary
Evaluation of the binary checkpoint `qtensorformer_small.pt` under mandatory strict loading (`load_state_dict(..., strict=True)`) confirmed that the checkpoint cannot be loaded into current Q-TensorFormer models.

## 2. Incompatibility Matrix

| Target Architecture | Strict Load Status | Missing Keys Count | Unexpected Keys Count | Primary Root Cause |
| :--- | :---: | :---: | :---: | :--- |
| `src.models.QTensorFormer` | **FAILED** | 185 | 113 | Architecture mismatch: Checkpoint uses TensorLy factorized FFN and LM head |
| `q_tensor_former.QTensorFormer` | **FAILED** | 101 | 113 | Namespace & architecture mismatch: Checkpoint uses `blocks.*` and factorized weights |
| `q_tensor_former_v2.QTensorFormer` | **FAILED** | 107 | 113 | Namespace & architecture mismatch: Checkpoint uses external TensorLy layers |

## 3. Scientific Policy
1. No keys were renamed arbitrarily to force a synthetic pass.
2. `strict=False` was **not** used to hide weight mismatches.
3. Track A benchmarks must acknowledge `CHECKPOINT_INCOMPATIBLE` honestly rather than hallucinating compatibility.