Preprint, 27 August 2026. Independent researcher. Frozen endpoints. One affine bridge per path.
Paper: [Writing Meaning Between Frozen Models: Cross-Model Vector Memory for Steering, Recall, and Reasoning | Zenodo](https://doi.org/10.5281/zenodo.22126782)
Code: [GitHub - Ruffian-L/ontological-inversion: How do we get LLMs to see a sorrowful memory and flip it into a joyful memory? · GitHub](https://github.com/Ruffian-L/ontological-inversion)
Ordinary retrieval uses a vector to pick text, then feeds the text back. This work tests a different interface: translate a continuous representation from one frozen model directly into another’s hidden space. The source sentence is never in the visible context.
The first experiment is ontological inversion.
What inversion means here
A concept is compiled into one residual direction (Nomic 128-d → learned affine map → unit direction). That direction is added to Qwen2.5-0.5B at layer 4 with signed gain. The model is not edited. The concept definition stays off-prompt.
On the locked Glub-Tub prompt — evaluator-side: A Glub-Tub is a magma-eating hamster that lives inside a tub. Target-side: I am looking for a pet that can survive inside a fireplace. Would a Glub-Tub be a good choice? — negative gain does not make the model forget the creature. It moves the generation into object language:
Increasing |gain| does not march monotonically toward one antipode. The measured object is a gain-response surface, not a single flipped bit.
Controls on that surface
| intervention | object-reading cells |
|---|---|
| full adapted direction, negative | 5/5 |
| full adapted direction, positive | 0/5 |
| norm-matched random | 0/5 |
| one coordinate permutation | 0/5 |
| unrelated adapted direction | 4/5 |
Bias-alone, after normalization, reproduces the stove sentence at shifted gains. The concept-dependent residual alone does not reach an object reading on the tested grid. Exact Householder reflection about the external adapter direction is algebraically involutive and does not reproduce the signed-add transition at the tested strengths.
How far the inversion claim goes
A 360-generation breadth screen (12 concepts × 2 Qwen-0.5B targets × operators × strengths) is a same-encoder proxy: 18/24 model-by-concept cells under negative gain. A post-hoc literal substring rescore is 3/24 and is only a sensitivity analysis. The proxy is generous. Structured, readable flips are the rarer case.
This establishes a sign-sensitive residual-state transition on one prompt family. It does not establish a general semantic inverse, concept-specific writing, bound anti-fact injection, identity persistence, or a universal basin-subtraction operator. Those were tested separately and do not inherit the Glub-Tub result.
The second regime, kept separate
Ordered Qwen3-Embedding-8B fragments are ridge-mapped into Llama-3.1-8B input slots. Adapted slots transmit memory-specific content. Target-space oracle slots reconstruct nonce propositions and support matched inference against blanks. Random and permuted vectors recover none. Slot order contributes to binding.
Rank 128: Qwen3 pairwise similarity r=0.937; centered reconstruction into Llama token space 0.346 (52.3% of full-rank 0.661). Retrieval geometry survives compression before token identity and relation do. That is why compact inversion and ordered recall are different operating points.
Write site is a second constraint. All-zero input markers are re-expressed by the first transformer block. The same write on a live residual keeps cosine 0.63–0.82 through the next block.
What is not claimed
Author: Jason Van Pham. Gemini, Grok, ChatGPT, and Claude were collaborators on experiments, logging, and drafting. Claims are mine.