CHIMERA Neural Transplantation

Transfers a compact learned neural squad between language models, preserves prior capabilities, and causally validates the graft through ablation, restoration, and matched controls.

Comment

Demo video

Devpost embed backup link: https://www.youtube.com/watch?v=Uigwfe02xK0

Inspiration

When a language model learns a narrowly scoped capability, transferring that capability to another model often requires broad retraining, replacement, or copying far more of the network than may be necessary.

CHIMERA began with a narrower research question: can a compact, knowledge-bearing neural team be transferred from a donor language model into a recipient, integrated with its host, and remain causally relevant without replacing the whole model?

A second challenge is catastrophic forgetting. Adding a new capability can damage what the recipient already knows. In the reported experiment, old-task retention remained at 100%, indicating no measurable loss on the tested prior tasks.

What it does

CHIMERA is a proprietary digital neural transplantation research prototype.

In the reported experiment:

  1. A donor model learned a novel capability.
  2. A private research system mapped a compact native neural squad associated with that capability.
  3. The selected squad was applied to a recipient as a digital graft.
  4. A post-graft integration stage helped graft and host work together.
  5. Matched controls, causal ablation, exact restoration, and old-task checks were used to evaluate the result.

This is a digital neural-network experiment. It does not claim biological transplantation, clinical application, or consciousness transfer.

Reported results

  • Selected graft causal fraction: 69.2%
  • Selected donor genealogy retained: 38.5%
  • Random graft causal fraction: 15.4%
  • Random graft genealogy retained: 0.0%
  • Therapy-only host-native route: 100.0%
  • Old-task retention: 100.0%
  • Healed baseline confidence margin: 14.290350
  • Selected graft disabled margin: 12.970250
  • Selected graft restored margin: 14.290350
  • Matched-random disabled margin: 14.276843
  • Selected margin drop: 1.320100
  • Selected margin recovery: 1.320100
  • Selected-vs-random margin-drop advantage: 1.306593
  • Bit-exact restoration: PASS in the frozen public-safe result artifact

The selected-graft ablation produced a substantially larger margin change than matched-random ablation. Restoring the selected graft returned the reported confidence margin to its healed baseline value.

How it was built

The public submission layer is dependency-free and runs locally using HTML, CSS, JavaScript, Python standard-library tooling, and PowerShell.

The demo loads all displayed measurements from public_results.json. A standard-library validator checks:

  • required public JSON structure
  • numeric ranges and derived arithmetic
  • baseline and restoration consistency
  • public-boundary flags
  • absence of prohibited private-detail fields

These checks validate the internal consistency of frozen public artifacts only. They do not reproduce the experiment, execute the proprietary system, or establish scientific universality.

Build Week work with Codex and GPT-5.6

The proprietary research core predates Build Week.

At the start of the Build Week task, Codex received only an isolated public-safe repository containing frozen results and disclosure materials. During Build Week, Codex using GPT-5.6 built:

  • the responsive localhost demo
  • the public-artifact validator
  • the launcher and test workflow
  • documentation and submission materials
  • the video narration package
  • the deterministic automated demo recorder

Codex did not access, reproduce, test, infer, reverse-engineer, or verify the proprietary research core, model weights, private checkpoints, neural addresses, or hidden implementation.

Codex feedback Thread ID:

019f8673-d448-7400-b4e4-a0c902373836

Challenges

The central challenge was presenting causal evidence without disclosing the proprietary implementation.

The submission also had to keep three concepts clearly separated:

  1. reported experimental results
  2. consistency checks on frozen public artifacts
  3. independent reproduction of the underlying research

The public repository supports the first two, but does not claim the third.

What we learned

A public research demonstration can communicate controls, ablation, restoration, retention, and causal specificity without exposing the underlying method.

We also learned that provenance and limitations must be presented as prominently as headline results, especially when the research core remains proprietary.

Limitations

  • This submission reports one proprietary digital neural-network experiment.
  • The public repository does not reproduce or independently verify the private runtime.
  • Artifact checks do not establish generalization across all models, tasks, or architectures.
  • No biological, clinical, consciousness-transfer, or universal-performance claim is made.
  • The proprietary discovery, mapping, grafting, integration, and causal-validation implementations are not disclosed.

What's next

Future work includes broader controlled evaluations across model families and tasks, expanded catastrophic-forgetting tests, stronger blinded controls, and additional public-safe audit summaries.

The long-term goal is modular capability transfer and targeted model repair without requiring whole-model replacement or broad retraining.

Testing instructions

From PowerShell in the private repository root:

.\run_checks.ps1
.\start_demo.ps1

Built With

Share this project:

Updates