AI & ML Method Register

Page review information

Edition covered: Project evidence register — Phase 49a findings, 2026-07-13
Content status: Reviewed
Coverage: Phase 49b method register; chemical and biological authority ceilings set by the Phase 49c evidence closure
Last reviewed: August 2026
Primary reference: The official published standard — required for final decisions

Confirm the governing edition, jurisdiction, and project-specific requirements before applying anything on this page.

Safety boundary: no learned method in this register is assigned direct closed-loop authority (level 5) or a safety function. Every row now carries an evidenced ceiling — as of Phase 49c no row is left at “Planned”. The safety-function path never routes through a learned layer — see the safety boundary.

How to read the register

Use the family sections to scan by problem type. Each family opens with a comparison table; the entries below it put the poor-fit and failure cases beside the claimed value so method selection does not become a capability catalogue.

How to read an entry. Every entry answers the same questions, in the same order: what it computes; a concrete example; where it runs; the simpler method it must beat; when it earns its place; when it does not; how much data it needs; why its authority is capped where it is; the tests it must pass; how it fails; what keeps the plant safe anyway; and how strong the evidence is. If an entry’s “poor fit” line describes your situation, the answer is usually the entry’s own “must beat” method. The example lines are illustrative — a typical use that stays within the method’s stated scope, not a documented deployment or an endorsement of a specific product.

Maturity describes deployment reality: industrially routine (widely deployed practice), piloted (documented industrial pilots), research (literature and lab evidence only). Evidence strength names the best available source class for the row: standards body, peer-reviewed, preprint, mixed, or engineering judgement.

Classical deterministic

Method	Max authority	Must beat	Maturity
PID control	≤ 5	On/off or state feedback control	industrially routine
Feedforward control	≤ 5	Feedback-only control	industrially routine
Rule engine / sequence logic	≤ 5	Hardwired relay logic or state machine	industrially routine
System identification	≤ 2	First-principles modelling or manual step-test fit	industrially routine
Statistical process control	≤ 2	Engineering limits and alarm rationalisation	industrially routine
First-principles numerical model	≤ 4	Empirical correlation or direct measurement	industrially routine

PID control — authority ≤ 5 · industrially routine

Computes: Computes control effort from present, accumulated, and rate error.
Example: Holding a pump discharge pressure at setpoint by trimming the VFD speed each scan.
Placement: device
Must beat: On/off or state feedback control
Use when: A measurable process has adequate dynamics and a stable operating envelope.
Poor fit when: Constraints, long delays, or strong interactions dominate.
Data burden: Step tests or operating history for tuning; no training set.
Authority basis: Mature deterministic baseline; project risk assessment still determines permitted control and safety roles.
Validation: Loop-direction, saturation, mode-transfer, disturbance, and fail-state tests.
Failure modes: Windup, noise amplification, oscillation, and hidden mode errors.
Independent protection: Trips and safety functions remain independent of the basic control loop.
Evidence / maturity: mixed / industrially routine
Sources: Control Theory Field Guide

Feedforward control — authority ≤ 5 · industrially routine

Computes: Offsets a measured disturbance before feedback error develops.
Example: Adding heater output the moment a cold-feed flow rise is measured, before the outlet temperature falls.
Placement: device
Must beat: Feedback-only control
Use when: The disturbance is measurable and its process effect is sufficiently known.
Poor fit when: Disturbance measurement or model sign and gain are unreliable.
Data burden: A validated disturbance-to-output model; no ML training set.
Authority basis: Deterministic control whose deployment authority is established by the project control and risk design.
Validation: Sign, scaling, timing, saturation, and feedback-recovery tests.
Failure modes: Wrong-sign action, timing mismatch, and excessive compensation.
Independent protection: Independent limits and trips constrain commanded output.
Evidence / maturity: mixed / industrially routine
Sources: Control Theory Field Guide

Rule engine / sequence logic — authority ≤ 5 · industrially routine

Computes: Maps explicit conditions and states to deterministic actions.
Example: Sequencing a filling station: clamp, fill to weight, cap, release, in a reviewed, fully testable state machine.
Placement: device
Must beat: Hardwired relay logic or state machine
Use when: Required behaviour can be stated, reviewed, and exhaustively tested.
Poor fit when: The state space cannot be bounded or requirements are ambiguous.
Data burden: Requirements and test cases; no training data.
Authority basis: Deterministic implementation; authority follows verified requirements, not this catalogue.
Validation: State, transition, timeout, restart, and abnormal-path coverage.
Failure modes: Missing transition, conflicting rule, stale latch, and unsafe restart.
Independent protection: Safety logic uses its required independent and certified architecture.
Evidence / maturity: mixed / industrially routine
Sources: Control Theory Field Guide

System identification — authority ≤ 2 · industrially routine

Computes: Fits a dynamic input-output model from designed or historical excitation.
Example: Fitting a heat-exchanger step-response model offline from a designed bump test, for later controller tuning.
Placement: engineering workstation
Must beat: First-principles modelling or manual step-test fit
Use when: A model is needed and safe excitation covers the intended envelope.
Poor fit when: Inputs lack excitation, data leak future information, or the plant changes materially.
Data burden: Time-aligned, sufficiently excited input-output data and validation runs.
Authority basis: Model construction is offline; any downstream controller requires a separate authority case.
Validation: Residual independence, held-out transients, stability, and physical-sanity checks.
Failure modes: Confounding, non-identifiability, extrapolation, and unstable fitted dynamics.
Independent protection: Identification experiments obey independent operating constraints.
Evidence / maturity: mixed / industrially routine
Sources: Control Theory Field Guide

Statistical process control — authority ≤ 2 · industrially routine

Computes: Detects statistically unusual process shifts against an established baseline.
Example: Detecting a slow drift in fill weight against an established stable-baseline control chart.
Placement: historian/analytics
Must beat: Engineering limits and alarm rationalisation
Use when: The process has a stable baseline and special-cause detection is useful.
Poor fit when: Operating modes are mixed or distribution assumptions are ignored.
Data burden: Mode-labelled baseline data with known quality.
Authority basis: Advisory quality or maintenance signal; not a safety trip.
Validation: False-alarm rate, mode separation, sensitivity, and recalibration review.
Failure modes: Alarm flooding, missed drift, autocorrelation bias, and baseline contamination.
Independent protection: Engineered alarms, trips, and product disposition remain independent.
Evidence / maturity: mixed / industrially routine
Sources: NIST AI Risk Management Framework 1.0 (NIST AI 100-1)

First-principles numerical model — authority ≤ 4 · industrially routine

Computes: Solves governing physical equations for prediction or estimation.
Example: Solving the governing heat-transfer equations to predict jacket temperature for a new recipe.
Placement: edge
Must beat: Empirical correlation or direct measurement
Use when: Governing physics and parameters are known well enough for the decision.
Poor fit when: Unknown closure relations dominate or solve time exceeds freshness needs.
Data burden: Geometry, properties, boundary conditions, and calibration measurements.
Authority basis: Deterministic model; bounded authority still requires verification, timing, and fallback.
Validation: Mesh/time convergence, conservation, parameter sensitivity, and plant comparison.
Failure modes: Wrong boundary condition, numerical instability, parameter error, and stale solve.
Independent protection: Safety limits rely on independent validated channels.
Evidence / maturity: mixed / industrially routine
Sources: Control Theory Field Guide

Estimation

Method	Max authority	Must beat	Maturity
Kalman filter	≤ 4	Direct measurement or fixed low-pass filtering	industrially routine
EKF / UKF nonlinear estimator	≤ 3	Observer, direct sensor, or linear Kalman filter	industrially routine
PCA / PLS latent-variable model	≤ 2	Selected sensors, mass balance, or linear regression	industrially routine
Gradient-boosted trees	≤ 2	Rules, linear model, or physics-based correlation	industrially routine
Gaussian-process regression	≤ 2	Physical correlation, interpolation, or linear regression	piloted
Learned soft sensor	≤ 2	Physical sensor, material balance, or PLS model	piloted
Online adaptive learning	≤ 0	Scheduled fixed models with reviewed offline updates	research

Kalman filter — authority ≤ 4 · industrially routine

Computes: Fuses a linear dynamic model and noisy measurements into a state estimate.
Example: Estimating true tank level from a noisy level transmitter fused with the inlet/outlet flow balance.
Placement: device
Must beat: Direct measurement or fixed low-pass filtering
Use when: The linear model and noise assumptions are adequate and hidden states matter.
Poor fit when: Dynamics are strongly nonlinear or noise assumptions are badly violated.
Data burden: Model, covariance estimates, and representative disturbances.
Authority basis: Deterministic estimator may support bounded control after project-specific observability and fault analysis.
Validation: Innovation residuals, observability, sensor-loss, initialization, and numerical tests.
Failure modes: Divergence, biased estimate, covariance mismatch, and silent sensor fault absorption.
Independent protection: Safety decisions do not depend solely on the estimate.
Evidence / maturity: mixed / industrially routine
Sources: Control Theory Field Guide

EKF / UKF nonlinear estimator — authority ≤ 3 · industrially routine

Computes: Estimates nonlinear process states using local linearization or sigma points.
Example: Estimating unmeasured reactor concentration from temperature and flow through a validated nonlinear model.
Placement: edge
Must beat: Observer, direct sensor, or linear Kalman filter
Use when: A nonlinear model is credible and missing states materially improve operation.
Poor fit when: The model is weak, states are unobservable, or compute timing is not bounded.
Data burden: Validated model, uncertainty assumptions, and transient cases.
Authority basis: Conservative ceiling pending application-specific estimator fault containment.
Validation: Monte Carlo transients, residual checks, sensor faults, and timing tests.
Failure modes: Divergence, local-linearization error, false confidence, and late output.
Independent protection: Independent sensors and protection remain authoritative.
Evidence / maturity: engineering judgement / industrially routine
Sources: Control Theory Field Guide; NIST AI Risk Management Framework 1.0 (NIST AI 100-1)

PCA / PLS latent-variable model — authority ≤ 2 · industrially routine

Computes: Compresses correlated variables or predicts a response from latent factors.
Example: Compressing dozens of correlated batch sensors into a few latent trends to spot an off-normal batch.
Placement: historian/analytics
Must beat: Selected sensors, mass balance, or linear regression
Use when: Variables are correlated and a validated low-dimensional structure is stable.
Poor fit when: Modes shift, nonlinearities dominate, or causal interpretation is required.
Data burden: Representative, scaled, mode-labelled data with reference responses for PLS.
Authority basis: Advisory ceiling unless an independent application case proves bounded use.
Validation: Held-out batches, leverage and residual limits, drift, and missing-data tests.
Failure modes: Spurious correlation, latent-space extrapolation, leakage, and scaling errors.
Independent protection: Direct measurements and protection remain authoritative.
Evidence / maturity: engineering judgement / industrially routine
Sources: NIST AI Risk Management Framework 1.0 (NIST AI 100-1)

Gradient-boosted trees — authority ≤ 2 · industrially routine

Computes: Predicts a class or value using an ensemble of sequential decision trees.
Example: Predicting product quality grade from tabular process settings, held out by asset, as a lab-check advisory.
Placement: edge
Must beat: Rules, linear model, or physics-based correlation
Use when: Tabular nonlinear interactions are stable and beat simpler baselines on held-out assets.
Poor fit when: Extrapolation or causal physical behaviour is required.
Data burden: Representative labelled tabular data with leakage-safe splits.
Authority basis: Advisory ceiling; explainability does not establish safe authority.
Validation: Asset/time holdout, calibration, shift, missing-input, fairness, and drift tests.
Failure modes: Leakage, threshold discontinuity, missing-value shortcut, and extrapolation error.
Independent protection: Independent sensors and limits remain authoritative.
Evidence / maturity: engineering judgement / industrially routine
Sources: NIST AI Risk Management Framework 1.0 (NIST AI 100-1)

Gaussian-process regression — authority ≤ 2 · piloted

Computes: Predicts a quantity with a kernel-based uncertainty estimate.
Example: Estimating a hard-to-measure yield with calibrated uncertainty from a small, modest-dimension dataset.
Placement: edge
Must beat: Physical correlation, interpolation, or linear regression
Use when: Data are limited, dimensionality is modest, and calibrated uncertainty is valuable.
Poor fit when: Datasets are large, inputs are high-dimensional, or kernel assumptions fail.
Data burden: Representative input-output samples and justified kernel design.
Authority basis: Uncertainty is model-dependent; use remains advisory without a separate bounded case.
Validation: Coverage calibration, out-of-domain challenge, sensitivity, and compute tests.
Failure modes: False uncertainty, kernel misspecification, scaling cost, and extrapolation.
Independent protection: Protection does not rely on posterior confidence.
Evidence / maturity: engineering judgement / piloted
Sources: NIST AI Risk Management Framework 1.0 (NIST AI 100-1)

Learned soft sensor — authority ≤ 2 · piloted

Computes: Infers a difficult-to-measure process variable from easier measurements.
Example: Inferring a stream composition between lab assays from temperature, pressure, and flow measurements.
Placement: edge
Must beat: Physical sensor, material balance, or PLS model
Use when: Reference measurements cover the operating envelope and maintenance value is clear.
Poor fit when: The target cannot be periodically measured or input sensors share common failures.
Data burden: Time-aligned reference assays or sensors across modes, lots, and drift.
Authority basis: Advisory ceiling because inferred measurement is vulnerable to drift and shared-cause error.
Validation: Time/lot holdout, uncertainty, sensor-fault, drift, and reference-reconciliation tests.
Failure modes: Target drift, lag mismatch, shared sensor bias, and confident extrapolation.
Independent protection: Independent direct measurement or conservative limits protect critical decisions.
Evidence / maturity: engineering judgement / piloted
Sources: NIST AI Risk Management Framework 1.0 (NIST AI 100-1)

Online adaptive learning — authority ≤ 0 · research

Computes: Updates model parameters from live operating data.
Example: Continuously retuning a model from live data — kept offline here until update bounds and rollback are fully controlled.
Placement: edge
Must beat: Scheduled fixed models with reviewed offline updates
Use when: Update bounds, acceptance tests, rollback, and drift causes are fully controlled.
Poor fit when: The context is certified, safety-related, adversarial, or labels arrive late.
Data burden: Live data, trustworthy feedback, bounded update rules, and retained baseline data.
Authority basis: Phase 49a recommends offline-only treatment in certified or safety-relevant contexts.
Validation: Shadow mode, update bounds, poisoning, rollback, reproducibility, and change-control tests.
Failure modes: Catastrophic forgetting, poisoning, feedback loop, silent drift, and unreproducible state.
Independent protection: Live learning is isolated from safety and direct control paths.
Evidence / maturity: standards body / research
Sources: ISO/IEC TR 5469:2024 — Functional safety and AI systems (Technical Report: guidance, not certifiable requirements); Regulation (EU) 2023/1230 on machinery; Regulation (EU) 2024/1689 — Artificial Intelligence Act

Optimisation

Method	Max authority	Must beat	Maturity
Moving-horizon estimation	≤ 3	Kalman filter or deterministic observer	industrially routine
Model predictive control	≤ 4	PID, feedforward, selectors, and override control	industrially routine
Robust MPC / safety filter	≤ 3	Hard selectors, clamps, permissives, and override control	piloted
Reinforcement-learning policy	≤ 3	MPC, dynamic programming, rules, or operator procedure	research

Moving-horizon estimation — authority ≤ 3 · industrially routine

Computes: Solves a constrained recent-history optimisation to estimate states and disturbances.
Example: Reconciling recent flow, level, and composition history against constraints to estimate an unmeasured disturbance.
Placement: edge
Must beat: Kalman filter or deterministic observer
Use when: Constraints and nonlinear dynamics justify the compute and engineering cost.
Poor fit when: A simple observer is adequate or solve time cannot be guaranteed.
Data burden: Process model, constraints, noise assumptions, and transient history.
Authority basis: Bounded use requires deterministic timeout and fallback behaviour.
Validation: Feasibility, convergence, worst-case timing, fault injection, and fallback tests.
Failure modes: Infeasible solve, local optimum, stale estimate, and model mismatch.
Independent protection: Protection must tolerate estimator loss or error.
Evidence / maturity: engineering judgement / industrially routine
Sources: Control Theory Field Guide

Model predictive control — authority ≤ 4 · industrially routine

Computes: Optimises future manipulated variables against a model and constraints.
Example: Holding a distillation column inside purity and flooding limits while minimising reboiler energy.
Placement: supervisory
Must beat: PID, feedforward, selectors, and override control
Use when: Interactions and constraints produce measurable value beyond simpler loops.
Poor fit when: Models, maintenance capacity, or solve-time guarantees are inadequate.
Data burden: Validated dynamic model, constraints, estimator, and disturbance scenarios.
Authority basis: Deterministic bounded supervisory control with independent basic and protective layers.
Validation: Constraint, infeasibility, model mismatch, timing, bumpless fallback, and cyber tests.
Failure modes: Bad model, infeasible optimisation, stale state, and economically unsafe objective.
Independent protection: SIS, trips, hard limits, and basic regulatory control remain independent.
Evidence / maturity: mixed / industrially routine
Sources: Control Theory Field Guide; NIST SP 800-82 Rev. 3: Guide to Operational Technology Security

Robust MPC / safety filter — authority ≤ 3 · piloted

Computes: Projects a proposed command into a modelled feasible set under uncertainty.
Example: Projecting an operator's proposed setpoint change into a modelled feasible set before it is applied.
Placement: supervisory
Must beat: Hard selectors, clamps, permissives, and override control
Use when: Uncertainty bounds are credible and constraints are too coupled for simple clamps.
Poor fit when: The safe set is incomplete or worst-case solve time is unbounded.
Data burden: Verified model, uncertainty set, constraints, and adversarial scenarios.
Authority basis: Phase 49a supports only operator-approved use absent an independently verified veto architecture.
Validation: Reachability, infeasibility, timing, uncertainty-bound, and fallback evidence.
Failure modes: False feasible set, numerical failure, hidden constraint, and unsafe fallback.
Independent protection: Certified protection and hard interlocks remain outside the filter.
Evidence / maturity: engineering judgement / piloted
Sources: ISO/IEC TR 5469:2024 — Functional safety and AI systems (Technical Report: guidance, not certifiable requirements); NIST AI Risk Management Framework 1.0 (NIST AI 100-1)

Reinforcement-learning policy — authority ≤ 3 · research

Computes: Selects actions to maximise a learned or simulated cumulative reward.
Example: Learning a start-up sequence in a high-fidelity simulator, run only behind an independent action envelope.
Placement: supervisory
Must beat: MPC, dynamic programming, rules, or operator procedure
Use when: A high-fidelity environment and independent action envelope show value over deterministic control.
Poor fit when: Unsafe exploration, reward misspecification, or unmodelled rare events cannot be ruled out.
Data burden: Validated simulator or extensive trajectories including disturbances and boundary cases.
Authority basis: Default operator-approved ceiling; level 4 needs an independently verified veto architecture.
Validation: Adversarial scenarios, reward audit, policy stability, shift, constraint, and rollback tests.
Failure modes: Reward hacking, unsafe exploration, simulator gap, brittle policy, and action chattering.
Independent protection: A non-learned envelope vetoes every unsafe or out-of-scope action.
Evidence / maturity: engineering judgement / research
Sources: ISO/IEC TR 5469:2024 — Functional safety and AI systems (Technical Report: guidance, not certifiable requirements); NIST AI Risk Management Framework 1.0 (NIST AI 100-1)

Perception

Method	Max authority	Must beat	Maturity
1D convolutional neural network	≤ 2	Spectral features plus thresholds or a simpler classifier	piloted
2D convolutional neural network	≤ 2	Machine vision rules, gauges, or human inspection	industrially routine
Temporal transformer	≤ 2	State machine, AR model, engineered windows, or smaller recurrent model	piloted
Autoencoder anomaly detection	≤ 2	Limits, residuals, SPC, or one-class statistical model	piloted

Split your data the way the plant is split. Random train/test splits overstate perception performance. Leakage in some form — label, future, or normal-data contamination — appears in this family’s failure modes, and each row’s validation names the matching leakage-aware split: asset-held-out, lot-held-out, or a temporal split with a leakage audit.

1D convolutional neural network — authority ≤ 2 · piloted

Computes: Classifies or embeds local patterns in waveform or time-series windows.
Example: Flagging likely bearing wear from accelerometer windows on a pump train, for inspection at the next stop.
Placement: edge
Must beat: Spectral features plus thresholds or a simpler classifier
Use when: Raw local patterns add repeatable value across representative machines and loads.
Poor fit when: Fault labels are sparse, acquisition changes, or a few engineered features suffice.
Data burden: Labelled waveforms spanning assets, loads, sensors, and acquisition conditions.
Authority basis: Advisory diagnostic ceiling; Phase 49a found no basis for learned safety authority.
Validation: Asset-held-out tests, shift tests, calibration, latency, drift, and abstention.
Failure modes: Domain shift, shortcut learning, label leakage, overconfidence, and aliasing.
Independent protection: Protection never depends on the diagnosis or confidence score.
Evidence / maturity: mixed / piloted
Sources: NIST AI Risk Management Framework 1.0 (NIST AI 100-1); ISO/IEC TR 5469:2024 — Functional safety and AI systems (Technical Report: guidance, not certifiable requirements)

2D convolutional neural network — authority ≤ 2 · industrially routine

Computes: Detects or classifies spatial patterns in images or transformed signals.
Example: Classifying surface-finish defects on parts from line-scan images where rule-based vision falls short.
Placement: edge
Must beat: Machine vision rules, gauges, or human inspection
Use when: Appearance variation is representatively sampled and rule-based vision is inadequate.
Poor fit when: Defects are rare, lighting is uncontrolled, or traceable metrology is required.
Data burden: Labelled images across lots, optics, lighting, defects, and acceptable variation.
Authority basis: Advisory or quality-classification ceiling, not a safety function.
Validation: Lot-held-out tests, rare-defect recall, calibration, shift, and camera-fault tests.
Failure modes: Lighting shortcut, missed novel defect, confident misclassification, and camera drift.
Independent protection: Safety sensing and protective stops remain independent.
Evidence / maturity: mixed / industrially routine
Sources: NIST AI Risk Management Framework 1.0 (NIST AI 100-1); ISO/IEC TR 5469:2024 — Functional safety and AI systems (Technical Report: guidance, not certifiable requirements)

Temporal transformer — authority ≤ 2 · piloted

Computes: Models long-range dependencies in multivariate event or sensor sequences.
Example: Predicting a developing process upset from long multivariate sensor and event history, as an operator advisory.
Placement: historian/analytics
Must beat: State machine, AR model, engineered windows, or smaller recurrent model
Use when: Long context adds validated value that simpler temporal models do not deliver.
Poor fit when: Data volume, timing, or interpretability cannot support the complexity.
Data burden: Large time-aligned sequences with mode, maintenance, and outcome labels.
Authority basis: Conservative advisory ceiling under the NIST AI risk-management lifecycle.
Validation: Temporal split, leakage audit, calibration, shift, ablation, and latency tests.
Failure modes: Future leakage, mode confounding, attention instability, drift, and high compute.
Independent protection: Independent alarms and trips remain effective during wrong or missing predictions.
Evidence / maturity: engineering judgement / piloted
Sources: NIST AI Risk Management Framework 1.0 (NIST AI 100-1); ISO/IEC TR 5469:2024 — Functional safety and AI systems (Technical Report: guidance, not certifiable requirements)

Autoencoder anomaly detection — authority ≤ 2 · piloted

Computes: Flags samples that reconstruct poorly relative to learned normal data.
Example: Flagging a compressor operating pattern that reconstructs poorly against learned normal running, for investigation.
Placement: edge
Must beat: Limits, residuals, SPC, or one-class statistical model
Use when: Normal operation is well sampled and unknown deviations merit investigation.
Poor fit when: Normal modes evolve rapidly or anomaly scores cannot be tied to action.
Data burden: Clean, representative normal data plus known faults for challenge testing.
Authority basis: Anomaly score is advisory because unknown does not establish hazardous or faulty state.
Validation: Known-fault challenge, false-alarm, mode, shift, calibration, and abstention tests.
Failure modes: Normal-data contamination, mode alarm, missed familiar anomaly, and opaque score.
Independent protection: Engineered protections cannot be replaced by anomaly detection.
Evidence / maturity: engineering judgement / piloted
Sources: NIST AI Risk Management Framework 1.0 (NIST AI 100-1)

Physics informed

Method	Max authority	Must beat	Maturity
PINN forward model	≤ 1	Finite-volume, finite-element, or ordinary differential equation solver	research
PINN inverse parameter model	≤ 1	Constrained optimisation around a numerical solver	research
Hybrid physics plus learned residual	≤ 2	Recalibrated physics model or explicit empirical correction	piloted
Learned chemical closure model	≤ 2	Mechanistic closure, tabulation, or conservative correlation	research

PINN forward model — authority ≤ 1 · research

Computes: Fits a neural approximation while penalising governing-equation residuals.
Example: Approximating a repeated diffusion solve as a research alternative to a numerical solver, benchmarked against it.
Placement: engineering workstation
Must beat: Finite-volume, finite-element, or ordinary differential equation solver
Use when: A differentiable physics prior and repeated-query need survive comparison with numerical solvers.
Poor fit when: Classical numerics are accurate and affordable or stiff/multiscale optimisation fails.
Data burden: Equations, boundary conditions, collocation points, and validation measurements.
Authority basis: Phase 49a failure evidence supports read-only use; classical numerics remain the baseline.
Validation: Conservation, boundary residual, mesh-independent reference, extrapolation, and repeatability tests.
Failure modes: Optimisation failure, spectral bias, nonunique fit, and deceptively low aggregate loss.
Independent protection: No safety or loop-closing decision relies on the PINN output.
Evidence / maturity: peer-reviewed / research
Sources: Physics-informed neural networks; Characterizing possible failure modes in physics-informed neural networks; Predictive Limitations of Physics-Informed Neural Networks in Vortex Shedding (Chuang & Barba)

PINN inverse parameter model — authority ≤ 1 · research

Computes: Infers model parameters or hidden fields while enforcing equation residuals.
Example: Recovering an unknown diffusivity field from sensor data while enforcing the governing-equation residual, in research.
Placement: engineering workstation
Must beat: Constrained optimisation around a numerical solver
Use when: The inverse problem is identifiable and PINN performance beats the numerical baseline.
Poor fit when: Parameters are non-identifiable or boundary data are incomplete.
Data burden: Governing equations, informative measurements, uncertainty, and synthetic recovery cases.
Authority basis: Read-only ceiling due inverse non-uniqueness and documented PINN training failures.
Validation: Parameter-recovery studies, identifiability, noise, multiple starts, and solver comparison.
Failure modes: Plausible wrong parameter, compensating errors, local minimum, and false confidence.
Independent protection: Inferred parameters cannot directly change protective settings.
Evidence / maturity: peer-reviewed / research
Sources: Physics-informed neural networks; Characterizing possible failure modes in physics-informed neural networks; Predictive Limitations of Physics-Informed Neural Networks in Vortex Shedding (Chuang & Barba)

Hybrid physics plus learned residual — authority ≤ 2 · piloted

Computes: Adds a learned correction to a first-principles prediction.
Example: Adding a small learned correction to a first-principles temperature model where a stable residual bias remains.
Placement: edge
Must beat: Recalibrated physics model or explicit empirical correction
Use when: Residual structure is stable and improves held-out performance without violating constraints.
Poor fit when: The base model is structurally wrong or residual data lack envelope coverage.
Data burden: Validated base model and representative residuals across regimes.
Authority basis: Advisory ceiling; the learned closure can invalidate physical guarantees.
Validation: Ablation, conservation, regime holdout, uncertainty, drift, and base-model fallback.
Failure modes: Physics cancellation, extrapolated correction, regime shift, and hidden bias.
Independent protection: Independent limits use the uncorrected conservative model or direct measurements.
Evidence / maturity: mixed / piloted
Sources: NIST AI Risk Management Framework 1.0 (NIST AI 100-1); Characterizing possible failure modes in physics-informed neural networks

Learned chemical closure model — authority ≤ 2 · research

Computes: Approximates an unresolved reaction, property, or transport closure inside a physical model.
Example: Approximating an unresolved reaction closure inside a physical model, bounded and validated across every intended regime.
Placement: edge
Must beat: Mechanistic closure, tabulation, or conservative correlation
Use when: The learned closure is bounded, faster, and validated across every intended regime.
Poor fit when: Rare regimes, conservation, positivity, or extrapolation cannot be controlled.
Data burden: High-quality simulation or experiment data spanning regime and uncertainty.
Authority basis: Phase 49c adversarial coverage supports a bounded advisory ceiling for a learned closure; no fitted closure receives hard-constraint or safety-function status.
Validation: Conservation, positivity, regime holdout, uncertainty, drift, and deterministic fallback.
Failure modes: Nonphysical closure, unstable coupling, distribution shift, and hidden compensation.
Independent protection: Independent protection assumes the learned closure can be completely wrong.
Evidence / maturity: preprint / research
Sources: Characterizing possible failure modes in physics-informed neural networks; Phase 49a adversarial evidence findings (internal evidence record; not published); Phase 49c chemical/biological evidence-closure and adversarial findings (internal evidence record; not published)

Language/agentic

Method	Max authority	Must beat	Maturity
LLM engineering copilot	≤ 2	Templates, search, rules, and human engineering review	piloted
LLM PLC code drafting	≤ 0	Reviewed templates, libraries, and manual implementation	research
Agentic engineering workflow	≤ 0	Scripted workflow with explicit approvals and checks	research

LLM engineering copilot — authority ≤ 2 · piloted

Computes: Drafts explanations, analyses, or engineering artefacts from supplied context.
Example: Drafting a functional-description document from supplied P&ID notes for an engineer to review and correct.
Placement: engineering workstation
Must beat: Templates, search, rules, and human engineering review
Use when: Drafting speed helps and every material claim and output can be reviewed.
Poor fit when: Source provenance is absent or the output can directly alter a plant.
Data burden: Approved context, retrieval boundary, prompts, and representative evaluation tasks.
Authority basis: Advisory ceiling; fluent text is not verified engineering evidence.
Validation: Grounding, citation, prompt-injection, confidentiality, repeatability, and expert-review tests.
Failure modes: Hallucination, source confusion, prompt injection, sensitive-data disclosure, and automation bias.
Independent protection: No generated output bypasses engineering review or safety lifecycle controls.
Evidence / maturity: mixed / piloted
Sources: NIST AI Risk Management Framework 1.0 (NIST AI 100-1); NIST AI RMF: Generative AI Profile (NIST AI 600-1, 2024-07-26) — companion to AI RMF 1.0; NIST SP 800-82 Rev. 3: Guide to Operational Technology Security

LLM PLC code drafting — authority ≤ 0 · research

Computes: Produces candidate PLC code or tests from a natural-language requirement.
Example: Drafting a first-pass interlock routine offline that an engineer rewrites and formally checks before any controller sees it.
Placement: engineering workstation
Must beat: Reviewed templates, libraries, and manual implementation
Use when: Generated code remains a draft and formal/static/dynamic checks are effective.
Poor fit when: Requirements are ambiguous or code could reach production without independent verification.
Data burden: Approved examples, language grammar, requirements, and adversarial test tasks.
Authority basis: Offline drafting only; the cited PLC work is a peer-reviewed proof of concept (ICSE 2024 SEIP) and does not establish deployment safety.
Validation: Compile, static analysis, simulation, traceability, mutation, security, and human review.
Failure modes: Invented API, unsafe state path, incomplete requirement, plausible wrong code, and hidden prompt dependency.
Independent protection: Safety application code follows its required lifecycle and independent verification.
Evidence / maturity: peer-reviewed / research
Sources: LLM4PLC: Harnessing Large Language Models for Verifiable Programming of PLCs (Fakih et al.); ISO/IEC TR 5469:2024 — Functional safety and AI systems (Technical Report: guidance, not certifiable requirements)

Agentic engineering workflow — authority ≤ 0 · research

Computes: Plans and invokes tools across a multi-step engineering task.
Example: Planning and running a multi-step documentation task with narrow, reversible tool permissions and gated steps.
Placement: engineering workstation
Must beat: Scripted workflow with explicit approvals and checks
Use when: Tool permissions are narrow, actions are reversible, and each consequential step is gated.
Poor fit when: The agent can change live controls, safety settings, credentials, or unreviewed records.
Data burden: Tool schemas, policy tests, representative tasks, and attack scenarios.
Authority basis: Offline engineering use only; no evidence supports autonomous plant authority.
Validation: Permission, injection, tool-error, rollback, audit-log, and human-approval tests.
Failure modes: Goal drift, unsafe tool choice, injection, cascading error, and false completion claim.
Independent protection: Agent credentials and network paths cannot reach safety or production control authority.
Evidence / maturity: standards body / research
Sources: NIST AI Risk Management Framework 1.0 (NIST AI 100-1); NIST SP 800-82 Rev. 3: Guide to Operational Technology Security

Interface

Method	Max authority	Must beat	Maturity
Edge inference deployment	≤ 2	Stream raw data to a central analytic service	industrially routine
OPC UA result publication	≤ 2	Vendor-neutral tags plus documented alarm/event mapping	industrially routine
MQTT Sparkplug result publication	≤ 2	OPC UA or a documented application-specific message contract	industrially routine
Historian feature pipeline	≤ 2	Direct trend review or explicitly computed reports	industrially routine
Digital-twin state synchronisation	≤ 2	SCADA mimic, simulator snapshot, or engineering calculation	piloted

Edge inference deployment — authority ≤ 2 · industrially routine

Computes: Runs a validated model near high-rate sensors and publishes a lower-rate result.
Example: Running a validated vibration model on a gateway beside the sensor and publishing only the health verdict upstream.
Placement: edge
Must beat: Stream raw data to a central analytic service
Use when: Bandwidth, latency, privacy, or acquisition rate makes local processing necessary.
Poor fit when: Edge lifecycle, timing, security, and model update controls cannot be maintained.
Data burden: Representative signals plus hardware timing, load, and failure tests.
Authority basis: NIST OT guidance supports local/segmented placement; it does not grant model authority.
Validation: Worst-case latency, resource exhaustion, reboot, version, integrity, and stale-output tests.
Failure modes: Late result, resource starvation, silent model mismatch, and disconnected drift.
Independent protection: Edge failure leaves the plant in a protected, diagnosable state.
Evidence / maturity: standards body / industrially routine
Sources: NIST SP 800-82 Rev. 3: Guide to Operational Technology Security; NIST AI Risk Management Framework 1.0 (NIST AI 100-1)

OPC UA result publication — authority ≤ 2 · industrially routine

Computes: Publishes model result, confidence, version, quality, and freshness as custom structured data.
Example: Publishing a diagnosis with its confidence, model version, quality, and freshness as a structured OPC UA object.
Placement: supervisory
Must beat: Vendor-neutral tags plus documented alarm/event mapping
Use when: Consumers need discoverable semantics, security roles, and quality metadata.
Poor fit when: A polling path is incorrectly expected to reconstruct high-rate source waveforms.
Data burden: Information model, units, timestamps, quality, version, and stale-state contract.
Authority basis: OPC UA transports application-defined results; it has no standard AI-result meaning or authority grant.
Validation: Schema, timestamp, quality, stale, role, reconnect, and consumer-compatibility tests.
Failure modes: Stale good-looking value, semantic mismatch, timestamp loss, and unauthorized write.
Independent protection: Publication failure or corruption cannot defeat independent protection.
Evidence / maturity: standards body / industrially routine
Sources: OPC UA Part 1: Overview and Concepts; OPC UA Part 9: Alarms and Conditions; OPC UA Part 18: Role-Based Security

MQTT Sparkplug result publication — authority ≤ 2 · industrially routine

Computes: Publishes state-aware metrics for model results over MQTT infrastructure.
Example: Publishing model results as state-aware Sparkplug metrics over MQTT for store-and-forward integration.
Placement: supervisory
Must beat: OPC UA or a documented application-specific message contract
Use when: Store-and-forward or brokered integration benefits from Sparkplug state semantics.
Poor fit when: Teams assume Sparkplug standardises AI field meanings or ignore broker/security design.
Data burden: Metric contract including units, timestamp, model version, quality, confidence, and freshness.
Authority basis: Sparkplug defines transport/state behaviour, not standard AI-result semantics or control authority.
Validation: Birth/death, stale, reconnect, duplicate, ordering, authorization, and schema tests.
Failure modes: Retained stale result, semantic collision, broker outage, replay, and metric drift.
Independent protection: Brokered results do not implement safety functions.
Evidence / maturity: standards body / industrially routine
Sources: Sparkplug Specification 3.0; NIST SP 800-82 Rev. 3: Guide to Operational Technology Security

Historian feature pipeline — authority ≤ 2 · industrially routine

Computes: Builds time-aligned, versioned analytic features from retained OT data.
Example: Building versioned, time-aligned features from historian data so an analysis can be reproduced months later.
Placement: historian/analytics
Must beat: Direct trend review or explicitly computed reports
Use when: Repeatable analytics require traceable transforms across long history.
Poor fit when: Sampling, timestamps, modes, and data quality cannot support the target claim.
Data burden: Governed tags, clocks, units, quality flags, modes, lineage, and retention.
Authority basis: Read-only/advisory analytics placement is consistent with segmented OT architecture.
Validation: Time alignment, leakage, unit, replay, lineage, access, and backfill tests.
Failure modes: Future leakage, resampling artefact, unit drift, bad-quality inclusion, and lineage loss.
Independent protection: Historian analytics remain outside safety and real-time control paths.
Evidence / maturity: standards body / industrially routine
Sources: NIST SP 800-82 Rev. 3: Guide to Operational Technology Security; NIST AI Risk Management Framework 1.0 (NIST AI 100-1)

Digital-twin state synchronisation — authority ≤ 2 · piloted

Computes: Aligns model state, plant observations, version, and freshness for comparison or forecasting.
Example: Keeping a simulation aligned with plant observations, with version and freshness, to forecast a defined scenario.
Placement: supervisory
Must beat: SCADA mimic, simulator snapshot, or engineering calculation
Use when: A defined use case needs stateful prediction and every variable has a semantic contract.
Poor fit when: The twin has no owner, calibration path, or stale-state behaviour.
Data burden: Asset identity, topology, state mapping, time sync, parameters, events, and version lineage.
Authority basis: Advisory ceiling until each composing model has a stronger verified authority case; ISO 23247 frames the twin but confers no safety authority.
Validation: State reconciliation, version mismatch, stale data, topology change, and fallback tests.
Failure modes: Twin drift, identity mismatch, stale state, wrong topology, and false fidelity.
Independent protection: The twin cannot override independent interlocks or protection.
Evidence / maturity: engineering judgement / piloted
Sources: NIST AI Risk Management Framework 1.0 (NIST AI 100-1); NIST SP 800-82 Rev. 3: Guide to Operational Technology Security; ISO 23247 — Digital Twin Framework for Manufacturing (Parts 1-4:2021; Part 5:2026)

Chemical kinetic

Method	Max authority	Must beat	Maturity
Mass and energy balance	≤ 3	Direct inventory and flow reconciliation	industrially routine
Arrhenius reaction-kinetics model	≤ 2	Validated empirical rate table or conservative design case	industrially routine
Equilibrium thermodynamic model	≤ 2	Validated property package, laboratory data, or conservative bounding calculation	industrially routine
Transport PDE model	≤ 2	Lumped model, empirical correlation, or validated numerical package	industrially routine

Mass and energy balance — authority ≤ 3 · industrially routine

Computes: Constrains flows, inventories, compositions, and heat by conservation.
Example: Reconciling metered flows and inventories around a unit by conservation to catch a measurement error.
Placement: edge
Must beat: Direct inventory and flow reconciliation
Use when: Boundaries, phases, sources, and measurements are defined.
Poor fit when: Unmeasured flows, phase change, or accumulation terms are omitted.
Data burden: Topology, properties, flow/composition/temperature measurements, and boundary conditions.
Authority basis: Conservation of mass and energy is a governing constraint, not a fitted model; Phase 49c places it in the hard-constraint layer. Ceiling matches the register's other deterministic state estimators and still depends on measured flows.
Validation: Closure residuals, units, transients, sensor faults, and independent inventory checks.
Failure modes: Wrong boundary, missing species, property error, and compensating sensor biases.
Independent protection: Protective limits use independently justified measurements and design bases.
Evidence / maturity: engineering judgement / industrially routine
Sources: Phase 49a adversarial evidence findings (internal evidence record; not published); Phase 49c chemical/biological evidence-closure and adversarial findings (internal evidence record; not published)

Arrhenius reaction-kinetics model — authority ≤ 2 · industrially routine

Computes: Relates reaction-rate constants to temperature within a fitted mechanism.
Example: Relating a reaction rate to temperature within a fitted mechanism over its validated parameter range.
Placement: edge
Must beat: Validated empirical rate table or conservative design case
Use when: Mechanism, parameter range, and transport assumptions match the process.
Poor fit when: Mass transfer, catalyst ageing, side reactions, or thermal runaway dominate outside calibration.
Data burden: Rate measurements across temperature, composition, pressure, and relevant catalyst state.
Authority basis: Phase 49c: kinetics is a fitted closure and does not inherit hard-constraint status from the governing model it sits inside; advisory only, within its validated parameter range.
Validation: Independent experiments, parameter uncertainty, extrapolation, heat-release, and sensitivity tests.
Failure modes: Exponential extrapolation, wrong mechanism, transport limitation, and parameter compensation.
Independent protection: Runaway protection is based on independent hazard analysis and protective systems.
Evidence / maturity: mixed / industrially routine
Sources: NIST Chemical Kinetics Database; Phase 49a adversarial evidence findings (internal evidence record; not published); Phase 49c chemical/biological evidence-closure and adversarial findings (internal evidence record; not published)

Equilibrium thermodynamic model — authority ≤ 2 · industrially routine

Computes: Computes phase or reaction equilibrium from thermodynamic relationships.
Example: Computing vapour-liquid equilibrium for a separation where the property method covers the composition range.
Placement: engineering workstation
Must beat: Validated property package, laboratory data, or conservative bounding calculation
Use when: Equilibrium is a valid approximation and the property method covers the composition range.
Poor fit when: Kinetics, metastability, transport, or missing species control behaviour.
Data burden: Species, phases, property parameters, composition, pressure, and temperature range.
Authority basis: Phase 49c closed the NIST source gap for equilibrium data, but the property method (activity model, equation of state) is fitted and treated as soft; advisory only.
Validation: Property comparison, phase-boundary cases, closure, sensitivity, and laboratory checks.
Failure modes: Wrong property method, missing phase, false equilibrium assumption, and extrapolation.
Independent protection: Independent relief, trip, and containment design remains authoritative.
Evidence / maturity: engineering judgement / industrially routine
Sources: Phase 49a adversarial evidence findings (internal evidence record; not published); Phase 49c chemical/biological evidence-closure and adversarial findings (internal evidence record; not published)

Transport PDE model — authority ≤ 2 · industrially routine

Computes: Models spatial diffusion, convection, heat transfer, or species transport.
Example: Modelling species diffusion across a packed bed where spatial gradients drive the design decision.
Placement: engineering workstation
Must beat: Lumped model, empirical correlation, or validated numerical package
Use when: Spatial gradients affect the decision and boundary conditions are defensible.
Poor fit when: Closures or boundaries dominate uncertainty beyond useful resolution.
Data burden: Geometry, material properties, constitutive closures, boundaries, and validation measurements.
Authority basis: Phase 49c: the transport structure is governing, but its constitutive closures and transfer coefficients are fitted and treated as soft; advisory only.
Validation: Grid/time convergence, conservation, benchmark, sensitivity, and transient validation.
Failure modes: Mesh dependence, closure error, wrong boundary, numerical diffusion, and unstable solve.
Independent protection: Safety margins and protection do not rely on a single unverified simulation.
Evidence / maturity: engineering judgement / industrially routine
Sources: Phase 49a adversarial evidence findings (internal evidence record; not published); Phase 49c chemical/biological evidence-closure and adversarial findings (internal evidence record; not published)

Biological

Method	Max authority	Must beat	Maturity
Monod-type growth model	≤ 2	Empirical batch profile or mass-balance trend	industrially routine
Fed-batch feed profile	≤ 3	Open-loop recipe with PI/PID trim	industrially routine
Raman-plus-PLS bioprocess soft sensor	≤ 1	Offline assay, direct sensor, or simpler spectral calibration	piloted
Metabolic flux-balance model	≤ 2	Yield coefficients, mass balance, or empirical pathway analysis	piloted
Advanced bioprocess supervisory control	≤ 2	Recipe, cascade PI/PID, ratio, and operator procedure	research

Monod-type growth model — authority ≤ 2 · industrially routine

Computes: Relates microbial growth rate to a limiting substrate under fitted assumptions.
Example: Relating microbial growth rate to a limiting substrate for a fermentation matching the fitted assumptions.
Placement: engineering workstation
Must beat: Empirical batch profile or mass-balance trend
Use when: Organism, substrate limitation, inhibition, and operating mode match the fitted model.
Poor fit when: Mixed populations, inhibition, morphology, or unmeasured limitations dominate.
Data burden: Biomass, substrate, product, feed, and environment measurements across batches.
Authority basis: Phase 49c: a fitted growth law is a soft closure regardless of the deterministic form it takes; advisory only, within the fitted organism and operating mode.
Validation: Batch holdout, parameter identifiability, inhibition, contamination, and uncertainty tests.
Failure modes: Wrong limiting factor, population shift, non-identifiability, and batch-to-batch drift.
Independent protection: Containment, sterilisation, and critical limits remain independently protected.
Evidence / maturity: engineering judgement / industrially routine
Sources: Phase 49a adversarial evidence findings (internal evidence record; not published); Phase 49c chemical/biological evidence-closure and adversarial findings (internal evidence record; not published)

Fed-batch feed profile — authority ≤ 3 · industrially routine

Computes: Schedules substrate addition using time, measurements, or a fitted growth model.
Example: Scheduling substrate addition over a fed-batch run from time and measurements against a validated trajectory.
Placement: device
Must beat: Open-loop recipe with PI/PID trim
Use when: The process and assay evidence support a repeatable feed trajectory.
Poor fit when: Culture state is poorly observed or batch variability invalidates the profile.
Data burden: Batch trajectories, feed records, assays, oxygen, pH, and product outcomes.
Authority basis: Phase 49c records open-loop feed profiles with PI/PID trim as the deterministic reality baseline. The ceiling covers that recipe/PID path only — a profile scheduled from a fitted growth model is soft and capped at 2.
Validation: Batch holdout, missed-feed, sensor fault, contamination, saturation, and fallback tests.
Failure modes: Overfeed, oxygen limitation, inhibition, wrong batch phase, and actuator error.
Independent protection: Independent limits protect pressure, temperature, gas, and containment.
Evidence / maturity: engineering judgement / industrially routine
Sources: Phase 49a adversarial evidence findings (internal evidence record; not published); Control Theory Field Guide; Phase 49c chemical/biological evidence-closure and adversarial findings (internal evidence record; not published)

Raman-plus-PLS bioprocess soft sensor — authority ≤ 1 · piloted

Computes: Estimates concentration or quality attributes from calibrated spectra.
Example: Estimating product concentration between assays from calibrated Raman spectra spanning lots and probes.
Placement: edge
Must beat: Offline assay, direct sensor, or simpler spectral calibration
Use when: Reference assays and spectra cover lots, instruments, probes, and process phases.
Poor fit when: Calibration transfer, fluorescence, fouling, or sparse reference assays dominate.
Data burden: Time-aligned spectra and trusted assays across batches, probes, and maintenance states.
Authority basis: Phase 49c: calibration is not transferable across probes, lots, and maintenance states without recalibration in context, so the learned/hybrid ceiling is defaulted to 1. Level 0 for any released-quality decision made alone.
Validation: Batch/lot holdout, calibration transfer, probe-fault, uncertainty, drift, and assay reconciliation.
Failure modes: Spectral drift, fouling, batch confounding, calibration transfer error, and false confidence.
Independent protection: Critical quality and safety decisions retain required independent confirmation.
Evidence / maturity: engineering judgement / piloted
Sources: Phase 49a adversarial evidence findings (internal evidence record; not published); NIST AI Risk Management Framework 1.0 (NIST AI 100-1); Phase 49c chemical/biological evidence-closure and adversarial findings (internal evidence record; not published)

Metabolic flux-balance model — authority ≤ 2 · piloted

Computes: Computes feasible metabolic fluxes under stoichiometric and objective assumptions.
Example: Computing feasible metabolic fluxes under stoichiometric constraints for a credible organism and condition.
Placement: engineering workstation
Must beat: Yield coefficients, mass balance, or empirical pathway analysis
Use when: Network, constraints, and objective are credible for the organism and condition.
Poor fit when: Regulation, kinetics, mixed populations, or objective ambiguity dominate.
Data burden: Stoichiometry, uptake/secretion bounds, biomass objective, and validation assays.
Authority basis: Phase 49c: stoichiometry is governing, but the objective function is an assumption and alternate optima are unconstrained; advisory only.
Validation: Flux consistency, alternative objectives, knockout/perturbation, uncertainty, and assay checks.
Failure modes: Wrong objective, missing pathway, unconstrained alternate optimum, and condition mismatch.
Independent protection: The model cannot set safety-critical feeds or containment limits.
Evidence / maturity: engineering judgement / piloted
Sources: Phase 49a adversarial evidence findings (internal evidence record; not published); Phase 49c chemical/biological evidence-closure and adversarial findings (internal evidence record; not published)

Advanced bioprocess supervisory control — authority ≤ 2 · research

Computes: Coordinates feeds, gases, temperature, or harvest timing from models and soft sensors.
Example: Coordinating feed, gas, and temperature from models and soft sensors where it beats the classical baseline across batches.
Placement: supervisory
Must beat: Recipe, cascade PI/PID, ratio, and operator procedure
Use when: It beats the classical baseline across batches with bounded, reviewable actions.
Poor fit when: Culture state, contamination, assay delay, or batch variability cannot be bounded.
Data burden: Representative batches, interventions, assays, disturbances, and actuator constraints.
Authority basis: Phase 49c: bounded advisory authority only, inside an envelope held by an independent non-learned layer, with zero risk-reduction credit; recorded as a feasibility case, not a class capability.
Validation: Batch holdout, adversarial disturbance, soft-sensor fault, constraint, rollback, and quality tests.
Failure modes: Wrong biological state, delayed assay, model drift, overfeed, and objective conflict.
Independent protection: Basic control, containment, sterilisation, trips, and operator stop remain independent.
Evidence / maturity: engineering judgement / research
Sources: Phase 49a adversarial evidence findings (internal evidence record; not published); NIST AI Risk Management Framework 1.0 (NIST AI 100-1); Phase 49c chemical/biological evidence-closure and adversarial findings (internal evidence record; not published)

Domain limits

Chemical and process models: conservation is a hard engineering check; kinetics, equilibrium, transport closures, and properties still need application-specific validation. Phase 49c set their authority ceilings: the conservation layer holds a real ceiling, but a fitted closure sitting inside a governing model does not inherit that hardness and is capped at advisory.
Biological models: classical PI/PID, recipes, and open-loop feed profiles are the reality baseline. Advanced models and soft sensors remain research or pilot patterns, and their Phase 49c ceilings sit at advisory or below — never a safety function.
PINNs: compare against established numerical solvers. Low training loss is not evidence of a correct field, parameter, boundary condition, or safe extrapolation.
Generated code and agentic workflows: remain offline drafts. Compilation is necessary but does not establish requirement completeness, correct abnormal behaviour, or safety suitability.

Review status

This register is generated from a corpus-owned Draft register and its Phase 49c authority ceilings were author-reviewed in August 2026. It is a decision aid to make the next engineering question explicit — it does not replace hazard analysis, detailed design review, or the applicable safety and cybersecurity lifecycle. Return to the authority gate for the ladder, the pre-flight question, and the envelope architecture every row is scored against.

Trust Boundary — Engineering Judgment Required

This site is an independent educational reference for industrial automation standards. It paraphrases and organizes guidance in original language; it is not a substitute for the official published standards, professional engineering judgment, or legal review, and content has not been independently verified against current published editions.

Items marked Review pending await technical review against the identified edition. Items marked Planned have no reviewed source material behind them. Do not rely on this site for compliance determinations, safety-critical design decisions, or legal interpretation.