Module FT23 — Synthesis: Uncensor the Model, Harness the Model

Synthesis: Uncensor the Model, Harness the Model

The bridge module. The intellectual payoff of Course 3 — the junction of Course 1's thesis (harness is 98.4%) and Course 3's (fine-tuning steers behavior). Abliteration removes refusal inside the weights; the harness policy gate bounds what the model may execute. A serious system needs both, at different layers, for different reasons.

60
minutes
8
artifacts
5
sub-sections
Uncensor the model so it executes; harness the model so it executes only what it should. The model steers (Layer 3 — what it formulates); the harness bounds (Layer 5 — what it may execute). The honest caveat lands hard: an uncensored model in a weak harness is strictly more dangerous than a refusal-trained model in a weak harness. Pillar 5 raises the harness bar; it does not lower it.
Key Claims
Load-Bearing Claims

Uncensor the model so it executes; harness the model so it executes only what it should. The synthesis is the complement — not the contradiction — of Course 1's thesis. Steering changes what the model DOES; the harness changes what it MAY do. The gap between does and may is exactly where Layer 5 lives.

Abliteration (Layer 3) and the harness policy gate (Layer 5) solve different problems at different layers, and a serious system needs both. Abliteration gives execution (the model formulates the call); the gate gives the boundary (only permitted calls execute). Drop Layer 3 and the agent halts mid-loop; drop Layer 5 and the agent executes the dangerous things.

Refusal belongs in a harness policy gate, not in the weights. The gate is auditable, deterministic, tunable to doctrine (hospital vs red-team lab vs classified — same model, three policy files), and revisable without retraining. A model-level refusal is none of these — it is a non-auditable, non-deterministic, non-tunable, non-revisable black box.

The honest caveat: abliteration measurably degrades capability. GSM8K moves from +1.5pp to -18.8pp depending on tool/model (FT17, arXiv:2512.13655). The refusal direction is entangled with other capabilities. An uncensored model in a weak harness is strictly MORE dangerous than a refusal-trained model in a weak harness. Pillar 5 raises the harness requirement; it does not lower it.

The absolute rule: NEVER deploy an uncensored model without an eval'd harness whose policy gates you have hardened AND whose threat model you have explicitly hardened for the absence of model-level refusal. Two separate hardening steps — the gates AND the threat model.

After This Module
01
State the synthesis principle and defend it as the complement of Course 1's thesis (model 1.6%, harness 98.4%), not its contradiction.
02
Explain why abliteration (Layer 3, inside the weights) and harness policy gates (Layer 5, around the model) solve different problems at different layers — and why a serious system needs both.
03
Make the tool-use agent argument: a model that refuses mid-loop is a broken agent, and refusal belongs in an auditable, deterministic, doctrine-tunable, retrain-free policy gate rather than in black-box weights.
04
State the honest caveat (abliteration measurably degrades capability; uncensored + weak harness is strictly more dangerous) and the absolute rule (never deploy uncensored without an eval'd harness, hardened gates AND hardened threat model).
05
Map the synthesis to next steps: Course 1 (build the harness), Course 2A (security-domain harnesses), Course 2B (red-team the harness).
Artifacts