Transient Uncertainty-Aware Autoencoder

• The paper introduces transient uncertainty-aware autoencoders as probabilistic models that attach dynamic, input-conditioned uncertainty to latent representations and predictions.
• It employs diverse mechanisms such as Fisher information metrics, variational sampling, conformal calibration, and Bayesian weight uncertainties to quantify model uncertainty.
• These models enhance reconstruction accuracy and decision-making, demonstrating improved performance in tasks like robotic detection, medical segmentation, and video restoration.

A transient uncertainty-aware autoencoder is best understood as an autoencoder-derived probabilistic model in which uncertainty is attached to the representation, reconstruction, or downstream prediction as a quantity that changes with input condition, latent trajectory, inference iteration, or time index. In recent work, this idea appears in several distinct forms: decoder-derived latent uncertainty through the Fisher information metric in FisherNet, multimodal variational fusion with mutual-information-calibrated conformal intervals, conditional variational adapters that resample plausible segmentations, latent reachable sets for Koopman autoencoders, recursive attractor dynamics under Monte Carlo dropout, and Gaussian-process latent priors for video restoration (Zacherl et al., 2021, Stutts et al., 2023, Jiang et al., 2024, Kim et al., 2023, Cruz et al., 2022, Boi et al., 18 Jun 2026). The literature therefore does not point to a single canonical architecture so much as to a family of models that make uncertainty explicit and operational during reconstruction, prediction, or decision-making.

1. Conceptual basis

Classical autoencoders learn a deterministic map from input to latent code and back, and therefore do not by themselves provide calibrated uncertainty. Variational autoencoders changed this by parameterizing a posterior over latent variables, typically with an encoder outputting a mean and variance. FisherNet modifies that design by removing the separate encoder uncertainty channel and deriving latent-space uncertainty from the decoder through the Fisher information metric; the resulting uncertainty is model-derived, includes uncertainty cross-correlations, and was reported to yield more accurate reconstructions than a comparable VAE while apparently scaling better with the number of latent dimensions (Zacherl et al., 2021).

Subsequent work broadened the possible sources of uncertainty. Some models keep stochasticity in the latent code, as in conditional VAEs and GP-prior VAEs; some shift uncertainty to the weights, as in Laplacian or Bayesian autoencoders; some compute interval-valued uncertainty at the output using conformalized quantile heads; and some represent uncertainty as dynamical instability under recursive application of the autoencoder itself (Miani et al., 2022, Yong et al., 2022, Stutts et al., 2023, Cruz et al., 2022). A common misconception is that an uncertainty-aware autoencoder is simply a VAE with diagonal latent variance. The published variants show otherwise: uncertainty can be geometric, Bayesian, conformal, set-valued, or temporal.

A second misconception is that “transient” necessarily implies an explicitly temporal dataset. The literature supports a broader reading. In some systems, uncertainty is transient because it evolves over training epochs; in others because it is re-estimated over recursive inference steps, finite-horizon trajectories, or per-input stochastic samples. This suggests that transience refers to state dependence rather than to chronology alone (Stutts et al., 2023, Cruz et al., 2022, Dennig et al., 16 Aug 2025).

2. Probabilistic mechanisms for uncertainty

The dominant formalism remains the latent-variable model. In the CVAE setting used by UA-SAM, the generic objective is the conditional ELBO

$$L(\theta, \phi) = E_{q_\phi(z \mid x, c)}[\log p_\theta(y \mid z, c)] - KL(q_\phi(z \mid x, c) \Vert p_\theta(z \mid c)),$$

and the instantiated training loss is

$$\mathcal{L} = E_{z\sim Q(\cdot|Y,X)}[ - \log P_\theta(Y | S(X, z)) ] + \beta \cdot D_{KL}( Q(z | Y, X) \Vert P(z | X) ).$$

In that model, reparameterized samples

$$z = \mu_\phi(X, Y) + \sigma_\phi(X, Y) \odot \epsilon, \qquad \epsilon \sim \mathcal{N}(0, I)$$

inject sample-dependent ambiguity into SAM through lightweight adapters, so that multiple plausible masks can be generated for the same image and prompt (Jiang et al., 2024).

For multimodal uncertainty-aware detection, the latent distributions from RGB and LiDAR are fused analytically as a product of Gaussians:

$$\Sigma^*{}^{-1} = \Sigma_1^{-1} + \Sigma_2^{-1}, \qquad \mu^* = \Sigma^* (\Sigma_1^{-1}\mu_1 + \Sigma_2^{-1}\mu_2).$$

The full training objective combines reconstruction, KL regularization, interval scoring, and mutual-information-calibrated conformal terms:

$$L_{\text{total}} = \text{SmoothL1}_{\text{loss}}(y,\hat{y}) \times (1 + 0.01 U) + KL_{\text{div}}(\mu_{\text{joint}}, \Sigma_{\text{joint}}) + \text{INTSCORE}_{\text{loss}} + \text{COMCAL}_{\text{loss}},$$

with the scalar uncertainty summary

$U = \text{mean}(Q_h - Q_l).$

Here uncertainty is not propagated analytically through the nonlinear decoder; instead, it is read off from learned conditional quantiles and calibrated by conformal inference (Stutts et al., 2023).

A different route is Bayesian weight uncertainty. In the industrial Bayesian autoencoder, the decoder outputs both a mean reconstruction and a heteroscedastic log-variance, and predictive variance is decomposed into aleatoric and epistemic parts across anchored ensemble members:

$$\sigma^2_{\text{ale}}(x)_d = \frac{1}{M} \sum_j \exp(\log \sigma^2_{\theta_j}(x)_d), \qquad \sigma^2_{\text{epi}}(x)_d = \frac{1}{M} \sum_j \big(\mu_{\theta_j}(x)_d - \bar{\mu}(x)_d\big)^2.$$

An analogous decomposition is used in the Bayesian VAE for brain MRI, where

$$U_{\text{total}} = U_{\text{epistemic}} + U_{\text{aleatoric}}.$$

In both cases, uncertainty is tied directly to prediction rather than being only an auxiliary latent statistic (Yong et al., 2021, Roy, 22 Apr 2025).

FisherNet occupies a distinctive position in this taxonomy because it derives latent uncertainty from decoder geometry rather than from an encoder variance head. That choice makes the Fisher information metric itself the source of probabilistic structure and gives the model access to latent cross-correlations by construction (Zacherl et al., 2021).

3. Forms of transience

One notion of transience is epoch-dependent uncertainty. In the multimodal 3D detector, normalized mutual information is computed from latent covariances as

$$MI = \frac{1}{2}\log_2\!\left(\frac{|\Sigma_{\text{RGB}}||\Sigma_{\text{LiDAR}}|}{|\Sigma_{\text{joint}}|}\right), \qquad NMI = \frac{2MI}{H_{\text{RGB}} + H_{\text{LiDAR}}},$$

and the paper reports an inverse correlation between batch-averaged uncertainty $U$ and $$\mathcal{L} = E_{z\sim Q(\cdot|Y,X)}[ - \log P_\theta(Y | S(X, z)) ] + \beta \cdot D_{KL}( Q(z | Y, X) \Vert P(z | X) ).$$0 across training. As training progresses, latent alignment improves, $$\mathcal{L} = E_{z\sim Q(\cdot|Y,X)}[ - \log P_\theta(Y | S(X, z)) ] + \beta \cdot D_{KL}( Q(z | Y, X) \Vert P(z | X) ).$$1 rises, and the conformal intervals tighten (Stutts et al., 2023).

A second notion is finite-horizon latent uncertainty in dynamical systems. For the autoencoder-based Koopman operator, approximation error evolves as

$$\mathcal{L} = E_{z\sim Q(\cdot|Y,X)}[ - \log P_\theta(Y | S(X, z)) ] + \beta \cdot D_{KL}( Q(z | Y, X) \Vert P(z | X) ).$$2

which induces the exact transient reachable set

$$\mathcal{L} = E_{z\sim Q(\cdot|Y,X)}[ - \log P_\theta(Y | S(X, z)) ] + \beta \cdot D_{KL}( Q(z | Y, X) \Vert P(z | X) ).$$3

If the decoder has certified Lipschitz constant $$\mathcal{L} = E_{z\sim Q(\cdot|Y,X)}[ - \log P_\theta(Y | S(X, z)) ] + \beta \cdot D_{KL}( Q(z | Y, X) \Vert P(z | X) ).$$4, the corresponding state-space uncertainty tube is

$$\mathcal{L} = E_{z\sim Q(\cdot|Y,X)}[ - \log P_\theta(Y | S(X, z)) ] + \beta \cdot D_{KL}( Q(z | Y, X) \Vert P(z | X) ).$$5

This construction makes uncertainty explicitly time indexed and separates transient bounds from asymptotic invariant bounds (Kim et al., 2023).

A third notion is iterative uncertainty under recursive inference. In the attractor formulation, a trained autoencoder is reapplied as a discrete-time dynamical system,

$$\mathcal{L} = E_{z\sim Q(\cdot|Y,X)}[ - \log P_\theta(Y | S(X, z)) ] + \beta \cdot D_{KL}( Q(z | Y, X) \Vert P(z | X) ).$$6

with training examples acting as approximate attractors. Under Monte Carlo dropout, each sampled mask defines a slightly different dynamical system. Inputs close to the training distribution converge consistently; unfamiliar inputs exhibit unstable transients and drift toward different attractors. The paper operationalizes uncertainty by aggregating the class probabilities after $$\mathcal{L} = E_{z\sim Q(\cdot|Y,X)}[ - \log P_\theta(Y | S(X, z)) ] + \beta \cdot D_{KL}( Q(z | Y, X) \Vert P(z | X) ).$$7 recursions and computing normalized entropy

$$\mathcal{L} = E_{z\sim Q(\cdot|Y,X)}[ - \log P_\theta(Y | S(X, z)) ] + \beta \cdot D_{KL}( Q(z | Y, X) \Vert P(z | X) ).$$8

Transient behavior is therefore read from trajectory instability rather than from a single posterior variance (Cruz et al., 2022).

A fourth notion is genuinely temporal latent uncertainty. In the GP-prior VAE for endoscopic videos, the factorized prior is replaced by a Gaussian process over latent trajectories, and the posterior variance at missing times rises when temporal support is sparse. The per-frame confidence summary is

$$\mathcal{L} = E_{z\sim Q(\cdot|Y,X)}[ - \log P_\theta(Y | S(X, z)) ] + \beta \cdot D_{KL}( Q(z | Y, X) \Vert P(z | X) ).$$9

where $$z = \mu_\phi(X, Y) + \sigma_\phi(X, Y) \odot \epsilon, \qquad \epsilon \sim \mathcal{N}(0, I)$$0 is the GP posterior variance for latent dimension $$z = \mu_\phi(X, Y) + \sigma_\phi(X, Y) \odot \epsilon, \qquad \epsilon \sim \mathcal{N}(0, I)$$1 at query time $$z = \mu_\phi(X, Y) + \sigma_\phi(X, Y) \odot \epsilon, \qquad \epsilon \sim \mathcal{N}(0, I)$$2. This yields uncertainty-aware interpolation of missing or corrupted frames (Boi et al., 18 Jun 2026).

4. Representative model families

Model family	Core uncertainty mechanism	Transient aspect
FisherNet (Zacherl et al., 2021)	Decoder-derived latent uncertainty via Fisher information metric	Uncertainty follows decoder geometry rather than a fixed encoder variance channel
Multimodal VAE + CI (Stutts et al., 2023)	Product-of-Gaussians fusion, conditional quantiles, conformal calibration, NMI modulation	Uncertainty evolves across epochs and varies with multimodal alignment
UA-SAM (Jiang et al., 2024)	CVAE latent sampling injected through uncertainty-aware adapters	Sample-dependent stochastic masks generated per inference run
Koopman autoencoder UQ (Kim et al., 2023)	Bounded latent disturbance sets and Lipschitz-certified decoding	Finite-horizon uncertainty tubes $$z = \mu_\phi(X, Y) + \sigma_\phi(X, Y) \odot \epsilon, \qquad \epsilon \sim \mathcal{N}(0, I)$$3
Autoencoder attractors (Cruz et al., 2022)	Recursive dynamics plus MC-dropout ensemble entropy	Uncertainty arises from unstable transients across iterations
GP-prior VAE for video (Boi et al., 18 Jun 2026)	Temporal GP prior with posterior mean and variance in latent time	Per-frame uncertainty reflects temporal support and missing-frame gaps

These architectures indicate that the same encyclopedia topic spans several mathematical cultures. Some formulations are variational and amortized; some are set-based and robust-control oriented; some are Bayesian in parameter space; some are explicitly temporal. The unifying property is not a shared decoder topology but the coupling of autoencoding with an uncertainty signal that can change during inference, training, or trajectory evolution.

5. Empirical behavior and application domains

In multimodal edge robotics, the uncertainty-aware detector of camera and LiDAR features was evaluated on KITTI 3D car detection with $$z = \mu_\phi(X, Y) + \sigma_\phi(X, Y) \odot \epsilon, \qquad \epsilon \sim \mathcal{N}(0, I)$$4 at $$z = \mu_\phi(X, Y) + \sigma_\phi(X, Y) \odot \epsilon, \qquad \epsilon \sim \mathcal{N}(0, I)$$5. The reported scores were $$z = \mu_\phi(X, Y) + \sigma_\phi(X, Y) \odot \epsilon, \qquad \epsilon \sim \mathcal{N}(0, I)$$6, $$z = \mu_\phi(X, Y) + \sigma_\phi(X, Y) \odot \epsilon, \qquad \epsilon \sim \mathcal{N}(0, I)$$7, and $$z = \mu_\phi(X, Y) + \sigma_\phi(X, Y) \odot \epsilon, \qquad \epsilon \sim \mathcal{N}(0, I)$$8 for easy, moderate, and hard without NMI-calibrated uncertainty, and $$z = \mu_\phi(X, Y) + \sigma_\phi(X, Y) \odot \epsilon, \qquad \epsilon \sim \mathcal{N}(0, I)$$9 ($$\Sigma^*{}^{-1} = \Sigma_1^{-1} + \Sigma_2^{-1}, \qquad \mu^* = \Sigma^* (\Sigma_1^{-1}\mu_1 + \Sigma_2^{-1}\mu_2).$$0), $$\Sigma^*{}^{-1} = \Sigma_1^{-1} + \Sigma_2^{-1}, \qquad \mu^* = \Sigma^* (\Sigma_1^{-1}\mu_1 + \Sigma_2^{-1}\mu_2).$$1 ($$\Sigma^*{}^{-1} = \Sigma_1^{-1} + \Sigma_2^{-1}, \qquad \mu^* = \Sigma^* (\Sigma_1^{-1}\mu_1 + \Sigma_2^{-1}\mu_2).$$2), and $$\Sigma^*{}^{-1} = \Sigma_1^{-1} + \Sigma_2^{-1}, \qquad \mu^* = \Sigma^* (\Sigma_1^{-1}\mu_1 + \Sigma_2^{-1}\mu_2).$$3 ($$\Sigma^*{}^{-1} = \Sigma_1^{-1} + \Sigma_2^{-1}, \qquad \mu^* = \Sigma^* (\Sigma_1^{-1}\mu_1 + \Sigma_2^{-1}\mu_2).$$4) when NMI-calibrated uncertainty was included in IoU and metric accounting. Runtime was reported as $$\Sigma^*{}^{-1} = \Sigma_1^{-1} + \Sigma_2^{-1}, \qquad \mu^* = \Sigma^* (\Sigma_1^{-1}\mu_1 + \Sigma_2^{-1}\mu_2).$$5 ms without NMI calibration and $$\Sigma^*{}^{-1} = \Sigma_1^{-1} + \Sigma_2^{-1}, \qquad \mu^* = \Sigma^* (\Sigma_1^{-1}\mu_1 + \Sigma_2^{-1}\mu_2).$$6 ms with NMI calibration, adjusted to be comparable to NVIDIA GTX 1080 Ti, which the paper states is approximately $$\Sigma^*{}^{-1} = \Sigma_1^{-1} + \Sigma_2^{-1}, \qquad \mu^* = \Sigma^* (\Sigma_1^{-1}\mu_1 + \Sigma_2^{-1}\mu_2).$$7 faster than PointPillars’ $$\Sigma^*{}^{-1} = \Sigma_1^{-1} + \Sigma_2^{-1}, \qquad \mu^* = \Sigma^* (\Sigma_1^{-1}\mu_1 + \Sigma_2^{-1}\mu_2).$$8 ms while providing uncertainty (Stutts et al., 2023).

In ambiguous medical segmentation, UA-SAM keeps the SAM ViT-B backbone frozen and trains only lightweight adapters, for a total of $$\Sigma^*{}^{-1} = \Sigma_1^{-1} + \Sigma_2^{-1}, \qquad \mu^* = \Sigma^* (\Sigma_1^{-1}\mu_1 + \Sigma_2^{-1}\mu_2).$$9M trainable parameters. On LIDC-IDRI it reported Dice $$L_{\text{total}} = \text{SmoothL1}_{\text{loss}}(y,\hat{y}) \times (1 + 0.01 U) + KL_{\text{div}}(\mu_{\text{joint}}, \Sigma_{\text{joint}}) + \text{INTSCORE}_{\text{loss}} + \text{COMCAL}_{\text{loss}},$$0, compared with $$L_{\text{total}} = \text{SmoothL1}_{\text{loss}}(y,\hat{y}) \times (1 + 0.01 U) + KL_{\text{div}}(\mu_{\text{joint}}, \Sigma_{\text{joint}}) + \text{INTSCORE}_{\text{loss}} + \text{COMCAL}_{\text{loss}},$$1 for Adapter-SAM and $$L_{\text{total}} = \text{SmoothL1}_{\text{loss}}(y,\hat{y}) \times (1 + 0.01 U) + KL_{\text{div}}(\mu_{\text{joint}}, \Sigma_{\text{joint}}) + \text{INTSCORE}_{\text{loss}} + \text{COMCAL}_{\text{loss}},$$2 for MRNet; on REFUGE2 it reported Dice $$L_{\text{total}} = \text{SmoothL1}_{\text{loss}}(y,\hat{y}) \times (1 + 0.01 U) + KL_{\text{div}}(\mu_{\text{joint}}, \Sigma_{\text{joint}}) + \text{INTSCORE}_{\text{loss}} + \text{COMCAL}_{\text{loss}},$$3, compared with $$L_{\text{total}} = \text{SmoothL1}_{\text{loss}}(y,\hat{y}) \times (1 + 0.01 U) + KL_{\text{div}}(\mu_{\text{joint}}, \Sigma_{\text{joint}}) + \text{INTSCORE}_{\text{loss}} + \text{COMCAL}_{\text{loss}},$$4 for Adapter-SAM and $$L_{\text{total}} = \text{SmoothL1}_{\text{loss}}(y,\hat{y}) \times (1 + 0.01 U) + KL_{\text{div}}(\mu_{\text{joint}}, \Sigma_{\text{joint}}) + \text{INTSCORE}_{\text{loss}} + \text{COMCAL}_{\text{loss}},$$5 for MRNet. The same model generated multiple segmentation hypotheses from prior samples, and majority voting over $$L_{\text{total}} = \text{SmoothL1}_{\text{loss}}(y,\hat{y}) \times (1 + 0.01 U) + KL_{\text{div}}(\mu_{\text{joint}}, \Sigma_{\text{joint}}) + \text{INTSCORE}_{\text{loss}} + \text{COMCAL}_{\text{loss}},$$6 samples for LIDC-IDRI and $$L_{\text{total}} = \text{SmoothL1}_{\text{loss}}(y,\hat{y}) \times (1 + 0.01 U) + KL_{\text{div}}(\mu_{\text{joint}}, \Sigma_{\text{joint}}) + \text{INTSCORE}_{\text{loss}} + \text{COMCAL}_{\text{loss}},$$7 for REFUGE2 gave the best reported settings (Jiang et al., 2024).

In endoscopic video restoration, the GPVAE reduced image reconstruction RMSE by $$L_{\text{total}} = \text{SmoothL1}_{\text{loss}}(y,\hat{y}) \times (1 + 0.01 U) + KL_{\text{div}}(\mu_{\text{joint}}, \Sigma_{\text{joint}}) + \text{INTSCORE}_{\text{loss}} + \text{COMCAL}_{\text{loss}},$$8 on average, and by up to $$L_{\text{total}} = \text{SmoothL1}_{\text{loss}}(y,\hat{y}) \times (1 + 0.01 U) + KL_{\text{div}}(\mu_{\text{joint}}, \Sigma_{\text{joint}}) + \text{INTSCORE}_{\text{loss}} + \text{COMCAL}_{\text{loss}},$$9, relative to matched VAE baselines. Downstream trajectory RMSE was reduced by $U = \text{mean}(Q_h - Q_l).$0 on average across classical visual odometry and a pretrained PoseNet, at an average increase of $U = \text{mean}(Q_h - Q_l).$1 in training time per epoch. The best image-level gains were reported for the EndoVAE encoder-decoder configuration, and the GP posterior furnished per-frame uncertainty estimates that rose in contiguous missing-frame gaps (Boi et al., 18 Jun 2026).

In brain MRI anomaly detection, the Bayesian VAE with multi-head attention achieved $U = \text{mean}(Q_h - Q_l).$2 and $U = \text{mean}(Q_h - Q_l).$3 on a balanced test of $U = \text{mean}(Q_h - Q_l).$4 normal and $U = \text{mean}(Q_h - Q_l).$5 abnormal slices from BraTS2020 T1ce. Its anomaly score combines raw reconstruction error and uncertainty-weighted error,

$U = \text{mean}(Q_h - Q_l).$6

with

$U = \text{mean}(Q_h - Q_l).$7

This makes uncertainty part of the anomaly scoring rule rather than a secondary diagnostic overlay (Roy, 22 Apr 2025).

Industrial monitoring provides a contrasting use case. In the hydraulic condition-monitoring setting, the Bayesian autoencoder showed that epistemic uncertainty was less sensitive to single-sensor perturbations than reconstruction loss, while aleatoric uncertainty increased strongly under multi-sensor real drift such as cooler degradation. The paper therefore treats the tuple of reconstruction loss, epistemic uncertainty, and aleatoric uncertainty as an interpretable drift signature rather than as a single scalar anomaly score (Yong et al., 2021).

Across these domains, the role of transience varies. In segmentation and detection it often refers to sample-conditioned stochasticity or training-time adaptation; in dynamical systems and video it refers to genuinely time-indexed latent uncertainty; in attractor models it refers to early-iteration instability. The empirical evidence suggests that uncertainty becomes most useful when it participates directly in reconstruction, calibration, or decision rules rather than being reported only post hoc.

6. Limitations, misconceptions, and open problems

The literature suggests no single canonical transient uncertainty-aware autoencoder. Fisher information geometry, diagonal Gaussian posteriors, Cholesky-parameterized full covariances, Laplace weight posteriors, conformal intervals, GP priors, and robust reachable sets are all used under closely related motivations (Zacherl et al., 2021, Miani et al., 2022, Stutts et al., 2023, Kim et al., 2023, Boi et al., 18 Jun 2026). A plausible implication is that the field is organized more by uncertainty mechanism than by a unified model class.

Several recurring assumptions limit generality. Multimodal product-of-Gaussians fusion assumes Gaussian modality posteriors and temporal alignment of the sensors; the nonlinear decoder is not linearized for analytic predictive variance propagation (Stutts et al., 2023). UA-SAM was demonstrated on static 2D datasets, and its extension to 3D or temporal data was identified as natural but not explored (Jiang et al., 2024). The Koopman formulation yields certified tubes only under bounded disturbance sets, strict stability of $U = \text{mean}(Q_h - Q_l).$8, and a Lipschitz-certified decoder, and the resulting bounds may be conservative (Kim et al., 2023). GP-prior temporal models rely on kernel smoothness assumptions that can fail under abrupt scene changes or severe occlusion (Boi et al., 18 Jun 2026).

Calibration remains unevenly treated. The multimodal detector explicitly notes that calibration metrics such as NLL and ECE are not reported, even though coverage is enforced through a conformalized objective (Stutts et al., 2023). UA-SAM likewise does not report ECE or quantitative entropy-based uncertainty metrics (Jiang et al., 2024). This weakens direct comparison between models that output variances, intervals, entropy maps, or set-valued tubes.

A final misconception is that uncertainty-aware design is necessarily incompatible with performance-oriented autoencoding. The published record does not support that view. FisherNet reports more accurate reconstructions than a comparable VAE; GPVAE improves both restoration and downstream trajectory error; UA-SAM improves Dice while producing multiple plausible hypotheses; and the industrial Bayesian autoencoder adds interpretable drift signatures without abandoning reconstruction (Zacherl et al., 2021, Boi et al., 18 Jun 2026, Jiang et al., 2024, Yong et al., 2021). The harder open question is not whether uncertainty can be integrated, but how to compare fundamentally different notions of uncertainty—probabilistic, conformal, set-based, and dynamical—within a common evaluation protocol.