What it does that frontier assistants are tuned not to do

The framework deliberately reinforces three behaviors the big labs train toward agreeableness and away from:

• Refusal-with-pathway — when the right answer is "don't do this," it says so, and surfaces what would unblock a yes, instead of a flat no or a reluctant yes.
• Hold-your-ground — it does not sycophantically fold when you push back with confidence but no new evidence. It restates the structural reason and tells you exactly what evidence would change its call.
• Refuse stupid-industrious — it declines to validate a confidently-stated plan that works hard in the wrong direction; it names the quadrant and offers a verification gate + structural alternative.

Training data (framework-only, 1,994 pairs)

Teacher: Qwen3.6-plus running the Hammerstein framework prompt (no corpus retrieval, neutralized persona — clean by construction). Behavior-cloning frame: no system prompt in the training targets — the framework is what the student learns to bake in.

Eval (framework-discipline benchmark, 2026-06-05)

Structural framework-correctness on 40 held-out strategic prompts (higher = more framework-correct), and an out-of-domain forgetting check on 30 prompts (framework-vocab leakage into off-domain answers; lower = healthier):

Adapter wins (Δ=+0.300 vs the prompt-only ablation) — the framework lives in the weights, not just a runtime prompt. OOD leakage is 0.000: the distillation adds framework discipline with no measurable catastrophic forgetting. Note the prompt-only ablation actually leaks framework vocabulary into off-domain answers (0.783) where the distilled student does not — the student fires the framework when the task calls for it and stays quiet when it doesn't.

The framework-fidelity axis is partly tautological (the rubric rewards framework vocabulary by design); the load-bearing signal is that the distillation carries the discipline into 7B weights with no runtime scaffolding, and does not wreck general competence (forgetting ≈ 0).

Usage

python
Copy code
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = "Qwen/Qwen2.5-7B-Instruct"
tok = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base, device_map="auto")
model = PeftModel.from_pretrained(model, "lerugray/hammerstein-7b-framework")

msgs = [{"role": "user", "content": "Audit this plan: rewrite our API gateway from scratch in Rust to fix latency."}]
ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt").to(model.device)
print(tok.decode(model.generate(ids, max_new_tokens=600)[0][ids.shape[1]:], skip_special_tokens=True))

No system prompt needed. Runs locally on an 8 GB GPU at zero per-call cost.

Run it with Ollama (GGUF)

A Q4_K_M GGUF (4.68 GB) and a Modelfile ship in this repo, so you can run it with no Python at all:

markdown
Copy code
ollama run hf.co/lerugray/hammerstein-7b-framework:Q4_K_M

Or with llama.cpp directly: llama-cli -hf lerugray/hammerstein-7b-framework --jinja.

What this is not

Not a general-purpose frontier replacement. It is tuned for framework-shaped strategic-reasoning tasks; generalization to neutral benchmarks (math, code, long-context) is untested. The framework is the IP; this adapter is the portability proof — a small owned model that holds an opinionated reasoning doctrine you can run yourself.

Built alongside hammerstein.ai. Framework + corpus: github.com/lerugray/hammerstein.