Invalid prompt error (while I'm working on code that runs a liquid handler)

Potential duplicates detected. Please review them and close your issue if it is a duplicate.

• #23604

Powered by Codex Action

Technical Solution: Invalid Prompt Error (Liquid Handler Code) - Issue #24183

Repository: openai/codex
Issue: #24183 - Invalid prompt error (liquid handler code)
Severity: Medium-High (blocks legitimate scientific/engineering workflows)
Estimated Value: 1000-3000 USD
Status: Proposed Solution

---

1. Problem Statement

Users developing code for scientific laboratory equipment - specifically liquid handler instruments (e.g., Hamilton, Tecan, Beckman Coulter systems) - receive an Invalid prompt error when submitting their code to Codex. The prompt contains legitimate scientific/engineering code (e.g., Python scripts controlling liquid dispensing volumes, pipetting sequences, or microfluidic protocols), but Codex safety filter incorrectly classifies it as violating content policy.

This is a false positive caused by the safety filter lack of domain awareness. Keywords such as acid, toxic, dispense, concentration, hazardous, and dilution - which are routine in laboratory automation - trigger the same safety heuristics that would flag genuinely harmful content.

Impact

• Blocked workflows: Researchers and engineers cannot use Codex for legitimate equipment automation.
• No diagnostic feedback: Users receive no explanation of which part of the prompt triggered the filter or why.
• No remediation path: No mechanism exists to report false positives or request re-evaluation.
• Domain loss: Scientific/engineering community loses trust in Codex as a viable tool.

---

2. Root Cause Analysis

2.1 Safety Filter Architecture (Current)

The current safety filter operates as a black-box classifier:

User Prompt -> Safety Filter -> [PASS / BLOCK]
|
Generic error: Invalid prompt

Key deficiencies:

| Deficiency | Description |
|---|---|
| No domain awareness | Filter treats all prompts identically regardless of context |
| Keyword-heavy heuristics | Over-relies on surface-level keyword matching without semantic understanding |
| No confidence scoring | Binary pass/block with no threshold tuning or uncertainty signals |
| No feedback loop | No mechanism for users to report false positives |
| No error granularity | Generic error message provides zero diagnostic value |

2.2 False Positive Trigger Patterns

Example legitimate liquid handler code that triggers false positives:

# Example: Legitimate liquid handler protocol
def dispense_reagent(plate, well, volume_ul, reagent):
if reagent.hazardous:
safety_shield.deploy()
arm.move_to(well)
pipette.dispense(volume_ul, rate=slow)
pipette.aspirate(neutralizer, volume_ul * 2)

Triggered keywords: hazardous, acid, concentrated, dispense, neutralize

The filter lacks context that:

1. This is instrument control code, not instructions for harmful activity
2. The hazardous reference is a safety property on a reagent object
3. The code actually improves safety by automating handling with proper protocols

2.3 Root Cause Summary

ROOT CAUSE: Single-pass safety filter with no domain model

1. No code-type detection (script vs. prose)
2. No scientific/engineering domain taxonomy
3. No contextual understanding of safety-positive code
4. No user feedback mechanism for calibration
5. No diagnostic error reporting for debugging

---

3. Proposed Solution

3.1 Architecture Overview