# GENERATION VALIDATION REPORT: Autoregressive Decoding Audit

**Date**: September 17, 2026  
**Auditor**: Senior ML Systems Engineer & Research Software Auditor  
**Final Status**: **`GENERATION_PASS`**

---

## 1. Generation Interface Verification

The repository provides generation interfaces at two levels:
1. **Model-Level Method**: `QTensorFormer.generate()` in `src/models.py` (with `AdaptiveKVCache`).
2. **Hugging Face Integration**: `QTensorFormerForCausalLM` in `src/hf_model.py` (inherits `transformers.GenerationMixin`).
3. **Pure PyTorch Greedy Loop**: Evaluated via standard causal next-token prediction `argmax(logits[:, -1, :])` directly on the model's `forward()` output contract.

---

## 2. Deterministic Reproducibility Audit

Deterministic greedy decoding was executed twice from identical initial input prompt `[10, 25, 30]` with `seed=42`:

### Test Run Comparison
- **`src.models.QTensorFormer`**:
  - Run 1 Tokens: `[10, 25, 30, 5278, 5278, 5278, 5278, 5278]`
  - Run 2 Tokens: `[10, 25, 30, 5278, 5278, 5278, 5278, 5278]`
  - Match: **100% Identical (Zero divergence)**

- **`q_tensor_former.QTensorFormer`**:
  - Run 1 Tokens: `[10, 25, 30, 2856, 1538, 2498, 2498, 2498]`
  - Run 2 Tokens: `[10, 25, 30, 2856, 1538, 2498, 2498, 2498]`
  - Match: **100% Identical (Zero divergence)**

---

## 3. Per-Token Sequence Growth & Dimension Trace

| Step | Current Sequence Length ($T$) | Logits Shape | Selected Next Token ID | Next Sequence Length ($T+1$) | Invariant Check |
| :---: | :---: | :---: | :---: | :---: | :---: |
| Prompt | 3 | N/A | N/A | 3 | Initial prompt |
| 1 | 3 | `[1, 3, 10000]` | $5278$ | 4 | **VALID** |
| 2 | 4 | `[1, 4, 10000]` | $5278$ | 5 | **VALID** |
| 3 | 5 | `[1, 5, 10000]` | $5278$ | 6 | **VALID** |
| 4 | 6 | `[1, 6, 10000]` | $5278$ | 7 | **VALID** |
| 5 | 7 | `[1, 7, 10000]` | $5278$ | 8 | **VALID** |

---

## 4. Verification Checklist
- [x] Input IDs semantics: Int64 tensor indexing within $[0, V-1]$.
- [x] Logits semantics: Real-valued unnormalized log-probabilities over vocabulary.
- [x] Next-token selection: Greedy argmax along vocabulary dimension.
- [x] Sequence growth: Monotonic single-token concatenation.
- [x] Determinism: Repeated execution with identical prompt and seed yields identical sequences.
- [x] Numerical stability: No NaN or Inf encountered during unrolling.

**FINAL STATUS**: **`GENERATION_PASS`**