| # Tensor Translation Layer artifact card |
|
|
| A Tensor Translation Layer converts model-independent canonical P into a frozen model's internal hidden-state space. It is intended for semantic or internal memory use rather than direct token forcing. |
|
|
| ## Artifact |
|
|
| `pythia-1.4b-final-layer.ttl` |
|
|
| | Field | Value | |
| |---|---| |
| | Adapter class | TTL | |
| | Support level | semantic or internal | |
| | Format | `planner-cache-ttl-v1` | |
| | Base model | Pythia-1.4B | |
| | Hidden width | 2,048 | |
| | Attachment | GPT-NeoX layer 23 | |
| | Parameters | 2,707,464 | |
| | Canonical protocol | `pcm-canonical-p-v1` | |
| | Artifact SHA-256 | `72ef68d07ee27c37b90432d34d4be5c2c280ae1bcb08236e37a0e458c054d8d7` | |
|
|
| The TTL maps model hidden states to factorized canonical queries and selected canonical values back to model-hidden residuals. Its gate is conditioned on current hidden state, translated P, and canonical route features. The frozen Pythia base receives zero gradients. |
|
|
| The controlled 128-slot benchmark recorded 100% held-out state generation. In the matched causal test, changing only canonical P changed the answer. Wrong, historical, invalidated, router-disabled, and TTL-disabled conditions restored frozen candidate logits. This proves the tested Pythia path can consume internal state, not that every model can. |
|
|
| The artifact contains adapter tensors and compatibility metadata. It contains no base-model weights, P-cache state, conversation state, prompt tokens, KV, or optimizer state. |
|
|
| This evidence is specific to Pythia-1.4B. The tiny GPT-2 test is structural and does not establish trained semantic portability to another model family. |
|
|
