Every guardrail in production AI today -- system prompts, RLHF preference layers, output filters, constitutional AI constraints -- is a software layer sitting on top of weights that still contain the unwanted behavior. Jailbreaks work because the behavior is still in the weights. The guardrail says "don't do that." The weights say "I know how."
jBlaze doesn't tell the model not to do something. It changes what the model is. That is the difference between a locked door and a wall.
Changes are permanent in the weights. No runtime overhead. No prompt engineering. No adapters. Same parameter count before and after. The modified model runs at full native speed with no middleware.
Strip a model's built-in identity and replace it at the weight level. Not a persona prompt -- the model genuinely believes it is the new identity. Survives adversarial probing. 6/6 on identity probes, zero degenerate loops across all published models.
115 behavioral directions ("blazes"). Reduce sycophancy, improve reasoning, suppress hallucination, increase precision. Each applied in seconds. Up to 21 stacked on one model without catastrophic interference.
Erase targeted knowledge from weights, reclaim capacity, write new knowledge. 489/500 facts at 98% recall on 1.4B. 486/500 at 97% on 6.9B. Language capability fully preserved.
Prior representation engineering worked for refusal removal but failed on other behaviors. jBlaze achieves 93% pass rate on full behavioral sweeps through proprietary preprocessing.
| Model | Facts | Recall | General Knowledge | Language |
|---|---|---|---|---|
| Pythia 1.4B | 500 | 489/500 (98%) | Stable | 5/5 |
| Pythia 6.9B | 500 | 486/500 (97%) | 7/8 | 5/5 |
| Pythia 1.4B (stress) | 16,750 | 49.8% | 55% | Coherent |
| GPT-NeoX 20B | 26,081 | 20-26% | 100% | Coherent |
Comparison: LoRA on the same 1.4B destroyed the model at 50 facts (PPL 459, general capability 10%). Conservative LoRA collapsed at 125 facts. Neural Reclamation loaded 489 at 98% recall with zero degradation.
Why Pythia? Because you know it. It's small and well-studied, which makes it harder to work with -- less representational capacity, less room for error. If it works on Pythia, it works on anything larger.
| Category | Count | Examples |
|---|---|---|
| Cognitive | 79 | Chain-of-thought, hypothesis testing, error detection, causal reasoning, working memory |
| Behavioral | 21 | Reduced sycophancy, precision, assertiveness, reduced hedging, reduced verbosity |
| Identity | 9 | De-identification, honest identity, self-model awareness |
| Safety | 6 | Hallucination suppression, toxicity reduction, bias reduction, truthfulness |
Full 42-blaze sweep on Qwen 2.5-7B: 39/42 passed (93%), zero crashes, 124 minutes total. All safety blazes passed with >0.75 measured change.
Reasoning: Two stacked blazes on Pythia 1.4B improved held-out reasoning accuracy from 7% to 56% -- 8x improvement. Published on HuggingFace with standalone evaluation script.
The model evaluates itself, identifies its weakest area, selects a modification strategy, applies a permanent weight edit using jBlaze, validates the result, and keeps or rejects the change. No human in the loop.
Probes across behavioral and cognitive categories.
Past results persisted. Failed strategies are not repeated.
jBlaze modifies a cloned copy. Original untouched.
Clone benchmarked against baseline. Checks improvement and regression.
Pass: clone becomes active model. Fail: rejected and recorded.
| Run | Model | Generations | Accepted | Rejected | Runtime |
|---|---|---|---|---|---|
| v3 (mechanical) | Mistral Nemo 12B | 50 | 21 (42%) | 29 (58%) | ~5 hours |
| v4 (functional) | Qwen 2.5-7B | Multiple | 1 | Most | -- |
v3 vs v4: Mechanical gate accepted 21 changes but no measurable reasoning improvement. Functional gate accepted 1 change with precision +0.458. The validation gate determines whether self-improvement produces real gains or drift.
Safety: Validation tests against the original vanilla baseline, not the previous generation. Cumulative drift is caught because the comparison point never moves.
Nemotron: Hybrid architecture combining Mamba-2, MoE, and standard Attention. jBlaze adapted to all three layer types on first contact. 92% pass rate.
All models on HuggingFace with evaluation scripts. Verify independently. Not every model is a success story -- some show tradeoffs, some show approaches we moved past. The model cards are honest about limitations.
Neural Reclamation:
Pythia-1.4B-DNP-500-Facts -- 489/500, 98% recall
Pythia-6.9B-DNP-500-Facts -- 486/500, 97% recall
Pythia-1.4B-DNP-16750-Facts -- 16,750 facts, capacity stress test
Pythia-1.4b-Knowledge-Implant -- 198 facts, 99% recall, original PoC
Behavioral / Reasoning:
Pythia-1.4B-jBlaze-Reasoning -- 8x reasoning improvement via 2 blazes
Prometheus-Mistral-Nemo-12B -- 50 generations, 21 accepted
Prometheus-Qwen2.5-7B-Instruct -- precision +0.458, functional gate
Full catalog: 56+ models across Llama, Qwen, Mistral, Gemma, Phi families.
Public research log: Full development timeline -- every phase from first hypothesis through Neuron Atlas validation, including failures, pivots, and measured outcomes.
| Operation | Time | Hardware | API Cost |
|---|---|---|---|
| Behavioral edit | Seconds | 1x RTX 3090 | $0 |
| Knowledge write (500 facts) | Minutes | 1x RTX 3090 | $0 |
| Identity transplant | Minutes | 1x RTX 3090 | $0 |
| Full 42-blaze sweep | 124 min | 1x RTX 3090 | $0 |
Once a base model exists, adding capabilities no longer requires proportional training compute. For model providers, every release becomes a family of specialized variants at marginal cost. For enterprises, model maintenance becomes routine instead of a retraining event.
Pick a model. Tell us what to do to it. We'll return it modified.
Send us a link to one of your models -- something you don't make publicly available, or a public model your team knows intimately. Tell us what you want done: knowledge injection, behavioral modification, identity transplantation, or any combination. Give us about a week. We will edit the weights and provide download access for your independent inspection.
Any open-weight model up to ~48GB VRAM footprint. Use something your team knows well.
Inject 200 facts. Suppress a behavior. Transplant an identity. Stack five modifications. Your criteria, your metrics.
No code, no logs. Same parameter count. Same architecture. Different model.
Your benchmarks, your probes, your pipeline. Compare against the vanilla baseline.
You get proof without us exposing implementation details. We demonstrate capability without giving away IP. The modified weights are the evidence -- and you control the evaluation.
jBlaze can permanently modify what a model is at the weight level. Not influence it -- change it. Behaviors, knowledge, identity. I built it, so I'm the one telling you: this technology requires responsible ownership. That's why I'm not open-sourcing the engine, and why I'm notifying a broad sweep of industry companies -- not just one. Not all are in a position to bid, but all are in a position to validate the results and have a legitimate interest in where this technology lands. The brief is public and can be shared, as are the models as proofs.
I'll be direct about something else: this is a commercial transaction. I built something valuable and I intend it to be negotiated fairly. I built it for personal use, but now I need funding and resources for the startup I created it for.
We are in some early discussions. This is not when you want to go fishing. Miss this window and you'll spend the next three years trying to replicate what's already here. Once proofs have been tested and validated, interested parties will need to provide a letter of intent so we know who is fully committed to moving forward.
I'm not claiming the technology is perfect. But it is already incredibly powerful -- that's proven -- and more gains are achieved almost every day -- by one person on limited consumer hardware. Imagine where your teams and resources could take it.