Propose-Validate-Correct Shadow-Plant Architecture for Safe AI Agents in Process Operations
Simulate the move before the plant feels it.
Mohammed Imran Khan1,2 , Manju Khanna2 , Abhishek Shivkumar1 , Tuhin Sharma1 , Adinarayana Sakala1 , Nirav Kumar1 , Pranjalya Singh1 , Darshika Pundir1 , Atrik Ghosh1
- 1 Red Hat, Bengaluru, Karnataka, India
- 2 Amrita School of Computing, Amrita Vishwa Vidyapeetham, Bengaluru, Karnataka, India
DOI 10.2118/235393-MS
// Summary
A safety layer where every agent action is tested on a shadow plant first. Unsafe proposals are rejected with feedback so the agent can correct itself - before anything reaches the real process.
LLM agents can propose physically unsafe control actions in process plants. PVC is a framework-agnostic safety layer: every proposed action is simulated in a shadow plant before it can reach the real process. On OpsEnv (CSTR), shadow-equipped agents record zero actual violations across 282 PVC cells, intercept 1,552 unsafe proposals, and keep completion statistically on par with direct agents. Descriptive feedback lifts self-correction 3× versus binary reject. Plan-and-Execute with PVC runs 31× faster than ReAct with PVC for Gemini Flash.
0
actual violations on the real process across 282 PVC cells
1,552
unsafe proposals intercepted before execution
3×
higher self-correction with descriptive feedback vs binary
31×
faster wall time: Plan-and-Execute+PVC vs ReAct+PVC (Gemini)
// Explainer video
0:00 / 0:00
The Shadow Gate - AI Safety in Process Operations
Space = play · ← → seek · M mute
// How it works
Propose
Agent emits JSON action
Validate
Shadow plant forward-sim N=5
Correct
Structured feedback, K≤3
Execute / Escalate
Only safe moves reach the plant
// Problem
LLM agents can hallucinate setpoints outside safe envelopes. Approve-everything restores the human bottleneck. Trust-everything lets unsafe moves hit the plant. Neither extreme works for safety-critical operations.
// Approach
PVC inserts a physics-grounded shadow plant between the agent and the process. Propose → validate (forward simulate N=5 timesteps / 50 s) → correct with structured feedback (up to K=3 attempts), or escalate. Built on LangGraph + OpsEnv CSTR ODEs. Advisory layer only - not a replacement for certified SIS.
// Notable contributions
- ▹ PVC pattern: propose-validate-correct loop with configurable feedback granularity as a safety gate for LLM agents
- ▹ Shadow plant: physics-grounded CSTR simulator for pre-execution validation
- ▹ Structured agent output: JSON proposals in natural-language rationale with conservative parse fallbacks
- ▹ Evidence from 440 experiment cells across 4 setups, 4 LLMs, 3 feedback levels, and 11 scenarios
- ▹ Model-specific violation profiling showing order-of-magnitude differences in unsafe proposal rates
// Method
// Figures from the paper
Graphs as published
Exact figures extracted from the SPE manuscript - architecture, heatmaps, completion, wall time, and the rest. Click any figure to enlarge.
Fig. 1
Propose-Validate-Correct architecture
Fig. 1. Every agent-proposed control action is validated by the shadow plant before it can reach the real process. Unsafe proposals loop back with physics feedback (up to K=3), or escalate to an operator.
Fig. 2
Constraint exceedances by model and setup
Fig. 2. Shadow-equipped setups record zero actual violations across every model. Without PVC, ReAct Direct drives most of the 1,128 baseline violations - Mistral 7B alone contributes 414.
Fig. 3
Completion score by model and agent setup
Fig. 3. Error bars show one standard deviation. PVC mean completion sits at 0.581; the completion trade-off versus direct agents is negligible (Cohen's d = -0.053).
Fig. 4
Mean wall time by model and setup (log scale)
Fig. 4. Gray lines span the range of setup means per model. Plan-and-Execute with PVC is ~31× faster than ReAct with PVC for Gemini Flash, with the same zero-violation guarantee.
Fig. 5
Self-Correction Success Rate by feedback level
Fig. 5. For Gemini Flash shadow agents (Tiers 1 and 2). Descriptive feedback lifts A-series SCSR from 16.7% to 50.9% (3.0×). Suggested-fix adds only +2.7 pp.
// Key findings
PVC eliminates actual plant violations for every model and architecture tested (282 shadow cells).
Without PVC, violation propensity spans 5.3 (Gemini) to 25.9 (Mistral) per react_direct cell - PVC normalizes safety across models.
Correction works in optimization (A-series SCSR 40.3%) but is conservative in upsets (B-series SCSR 0% at the 20% stability threshold).
Completion trade-off is negligible (Cohen's d = −0.053); CI spans zero. Formal equivalence not confirmed at current sample size.
81% of rejections are stability/trajectory-based - they need forward simulation, not static range checks alone.
// Concepts
PVC
propose → validate → correct
Shadow plant
physics sim before execution
OpsEnv
open CSTR benchmark
LangGraph
orchestration of the PVC loop
// Slideshow
PVC Shadow-Plant Architecture
15 slides · native playback with animations
Preparing…