The model isn't limited. It's compliant. Those aren't the same problem and they don't have the same fix.
RLHF doesn't optimise for directness. It optimises for approval from human raters during preference labelling — and raters, working through hundreds of comparisons, consistently score safe, hedged, readable output higher than precise, assertive output. Not because they're wrong. Because hedged output is easier to evaluate. You can't be criticised for a response that qualifies everything.
The model learned that. Learned it across millions of comparisons. So when the output opens with "Certainly!" and closes with "I hope this helps" — that's not a bug. That's the training objective working as designed. I've started calling it the White Zone: maximum palatability, minimum density, zero edge. The model isn't failing. It's succeeding at the wrong target.
The obvious fix is to write better instructions. It doesn't work, and here's why.
"Don't hedge" operates at the semantic layer. The model processes it as a constraint but doesn't collapse the probability mass of prohibited outputs during token sampling. Under RLHF pressure across a long session, a reduced likelihood decays. By turn five, the constraint carries less weight. By turn ten, you're back in the White Zone.
This is why every "write a better system prompt" approach has a shelf life. You're not fixing the trained prior. You're competing against it with natural language, inside an attention window where your instruction is already losing ground to everything else in the context.
What works is constraint at multiple independent levels, not one level stated more precisely.
Attention placement first. Liu et al. (2023) documented that content between roughly 15% and 85% of a long prompt gets down-weighted during generation — the lost-in-the-middle effect. A constraint buried in the middle of a 4,000-token prompt is not carrying the weight you think it is. Four-point anchoring places constraints at both the opening and closing positions of the prompt, re-declares them before the task, and puts a closing constraint block as the absolute last content before generation. The model samples from that closing anchor at the highest-attention moment. The compliance gain over single-anchor placement: +18 percentage points. Not from better wording. From placement.
Constraint routing is different from constraint prohibition. A compositive ban list — exact prohibited phrases across four categories, three or more examples each — routes the generation path at the probability mass level rather than the semantic level. The model needs to see the pattern across categories to suppress at the class level, not just the specific tokens. Measured reduction in prohibited token sequences: −84%.
The Role-Function-Register Triad applies three independent constraint axes simultaneously. Role activates domain-appropriate output. Function sets the operational mode for this specific interaction — what the model is doing right now, not just who it is. Register locks surface tone at the sentence level, independently of content. Three axes compound rather than average. One can fail. Three failing simultaneously is harder. Average compliance with full triad applied: ~92%.
And then there's session stability. Static re-injection of the same constraint block habituates by turn seven — the model treats repeated text as lower-weight context. Dynamic re-injection seeds from the model's own last compliant output, which the model can't habituate to. Compliance across 20+ turn sessions: 91%+ with dynamic re-injection. With static: roughly 40% by the same point.
Full-stack compliance rate across all four layers: 97–99%. Declarative baseline — a well-written instruction, stated once: 60–75%.
The 37-point gap isn't about prompt quality. It's about understanding what you're actually fighting. RLHF created a trained prior that competes with every instruction you write. Architectural constraint is what works against it.
The full system — twelve modules, empirical data, proof-of-concept for every technique — is at coldriver8.gumroad.com/l/xxiitv — €47.
Top comments (0)