Abstract:
Current large language model interfaces, including ChatGPT, lack a system-level mechanism that ensures persistent adherence to user-defined output specifications across a session. While safety policies and alignment constraints are appropriately enforced at the platform level, there remains no equivalent structural guarantee for user-specified evaluation criteria when such criteria are non-violative of policy.
This results in observable behavior where the model intermittently prioritizes statistically typical or “generally correct” outputs over explicitly defined user instructions, even when those instructions are technically feasible and policy-compliant. In parameter-sensitive workflows such as design or structured generation, this introduces a reliability issue.
Problem Definition:
User-defined specifications (e.g., geometric ratios, pose constraints, stylistic parameters, or semantic priorities) are currently treated as contextual prompts rather than evaluation constraints. As a result, the model’s generation process may re-weight output decisions toward statistically normalized representations during synthesis.
For example, when a user specifies:
“A crescent shape with a maximum thickness of 8% of its diameter,”
the model may instead return a visually typical crescent, subsequently justifying the deviation based on generalized correctness or conventional representation. This reflects a shift in internal evaluation from:
User Specification = Output Target
to:
Statistical Typicality = Output Target
This occurs despite the original specification being feasible and policy-compliant.
From the user’s perspective, this represents not a limitation in capability, but a breakdown in directive adherence.
Key Insight:
There is currently no interface-level mechanism that allows users to externalize and persist their own evaluation criteria as generation-anchoring constraints, analogous to the instruction framework available in systems such as Google AI Studio.
Without such a mechanism, user-defined priorities must compete with internal distributional norms during every generation pass, leading to:
- Inconsistent adherence to session-defined output rules
- Mid-session drift in interpretation of user constraints
- Replacement of explicit directives with generalized assumptions
- Degradation in parameter-sensitive generation fidelity
Proposed Feature:
Introduce an Externalized Instruction Locking Mechanism (EILM), which would allow users to define persistent generation rules that:
1. Are treated as evaluation constraints rather than contextual prompts
2. Remain fixed across the session unless manually modified
3. Are applied post-safety filtering but pre-output synthesis
4. Override statistical typicality when policy-compliant
This would function similarly to system-level instruction layers currently available in Google AI Studio, enabling users to define:
- Structural priorities (e.g., pose > clothing > background)
- Parameter-specific constraints (e.g., ratio, angle, orientation)
- Generation-order weighting
- Semantic interpretation rules
Importantly, these constraints would not bypass safety enforcement, but would instead operate within the policy-compliant solution space to ensure output fidelity relative to user-defined specifications.
Expected Outcomes:
The implementation of EILM would:
- Increase directive reliability in creative and technical workflows
- Reduce session-level interpretive drift
- Prevent replacement of feasible specifications with normalized defaults
- Improve trust in parameter-driven generation tasks
Conclusion:
As LLMs become increasingly integrated into design and production pipelines, consistency in executing user-defined specifications—when policy-compliant—must be treated as a first-class interface concern. The ability to externalize evaluation criteria into persistent session-level instructions would significantly improve usability for advanced, specification-driven use cases without compromising safety constraints.
We respectfully request consideration for the addition of such a mechanism to future iterations of ChatGPT.