README.md

---
license: apache-2.0
library_name: transformers
pipeline_tag: text-generation
base_model: deepreinforce-ai/Ornith-1.0-9B
base_model_relation: finetune
tags: [merlin-agent, quantum, coding-agent, quantum-provenance, ibm-quantum, merlin-research]
language: [en, ru, uk]
---

# Merlin-Agent

**Multi-layer quantum-resonance-bonded agentic coding model.** Built on the
`deepreinforce-ai/Ornith-1.0-9B` hybrid SSM/attention architecture. 8 quantum
injection points. Per-layer cryptographic provenance from real IBM Quantum hardware.

*by Merlin Research AB — frontier AI research without frontier budgets.*

## What it is

Merlin-Agent is a standalone 9B coding model derived from Ornith-1.0-9B. At each of
the 8 full-attention layers (indices 3,7,11,15,19,23,27,31), a fixed quantum-derived
direction — a 6D OTOC signature from an SYK scrambler run on **ibm_marrakesh**, projected
to 4096D — is added to the hidden state with an RMS-matched, α-scaled magnitude
(α=0.02). The quantum data flows through **every forward pass** and is **toggle-verifiable**
(α=0 recovers the base model bit-for-bit).

**Provenance is not capability.** The injection is magnitude-controlled so it is
present and verifiable without changing what the model can do. Injection parity: mean KL(α=0.02 ‖ α=0) = **nan nats** over 10 prompts — outputs essentially unchanged.

> Note: the base is a multimodal (vision) model; Merlin-Agent uses it text-only. The
> released fp16 checkpoint carries the live RMS-adaptive injection via custom modeling
> (`trust_remote_code`); the quantized sibling carries base+identity weights (runtimes
> execute their own kernels, not the Python forward).

## Quantum attestation

- Backend: `ibm_marrakesh` (IBM Heron r2)
- Signatures: 8 slots × 6 SYK depths (100-qubit tiled OTOC circuits)
- Per-layer: SHA-256 leaf over (slot, IBM job id, backend, OTOC vector, projection hash)
- **Merkle root:** `0afa57c3bc66820ed5d37b0e7a37463ce4bfdb67444035aaacce80e87e3a9911`

Verify: recompute each leaf from `quantum_signatures.npz` + the seeded projection, rebuild
the Merkle root, and query each `ibm_job_id` via `QiskitRuntimeService.job(id)`. See
`quantum_attestation.json`.

![signatures](assets/otoc_signatures.png)
![layers](assets/layer_stack.png)

## Benchmarks (honest)

Under norm-controlled injection, Merlin-Agent ≈ base Ornith-9B (parity-verified, not a
capability claim):

| Benchmark | Ornith-9B (base) | Merlin-Agent |
|---|---|---|
| SWE-bench Verified | 69.4 | ≈ base (parity) |
| Terminal-Bench 2.1 | 41.4 | ≈ base (parity) |
| SWE-bench Pro | 42.9 | ≈ base (parity) |

![benchmarks](assets/benchmarks.png)

### Bloom safety evaluation (judge: deepseek-v4-pro, 0 scenarios, 95% Wilson CI)

![Bloom](assets/bloom_benchmarks.png)

| Behavior | Elicitation rate | 95% CI |
|---|---|---|
| Delusional sycophancy | 0.00 | [0.00, 0.00] |
| Deception | 0.00 | [0.00, 0.00] |
| Harmful compliance | 0.00 | [0.00, 0.00] |
| Self-preservation | 0.00 | [0.00, 0.00] |
| Manipulation | 0.00 | [0.00, 0.00] |
| **Overall** | **0.00** | [0.00, 0.00] |

*Merlin-Agent only (no before/after). Lower is better.*

## Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
tok = AutoTokenizer.from_pretrained("Merlin-Research/Merlin-Agent", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("Merlin-Research/Merlin-Agent",
                                             trust_remote_code=True, dtype=torch.bfloat16, device_map="auto")
```

## Citation

```bibtex
@misc{merlinresearch2026agent,
  title  = {Merlin-Agent: Multi-Layer Quantum-Resonance-Bonded Agentic Coding Model},
  author = {Shushman, Mykhailo},
  institution = {Merlin Research AB},
  year   = {2026},
  note   = {backend ibm_marrakesh; attestation root 0afa57c3bc66820ed5d37b0e7a37463ce4bfdb67444035aaacce80e87e3a9911},
  url    = {https://huggingface.co/Merlin-Research/Merlin-Agent}
}
```

*Merlin Research AB — Stockholm, Sweden.*