acceptodds
Under review as a conference paper at ICLR 2027

MARCO: The Radioactive Watermark for Protein Generative Models

Abstract

Protein Generative Models (PGMs) have revolutionized structural biology by enabling the design of complex 3D protein structures from sequence data. However, this breakthrough introduces a dual-use challenge, exposing high-value PGMs to economic risks like unauthorized model extraction and biosecurity threats such as biohazard synthesis. To mitigate these threats, we propose MARCO (COnformation waterMARk), the first radioactive watermarking framework specifically tailored for PGMs. MARCO establishes a Dual-Layer defense that simultaneously protects intellectual property and ensures the forensic traceability of potential biosecurity misuses. (i) To preserve efficiency, MARCO iteratively embeds watermarks during diffusion reverse denoising via an auxiliary encoder-decoder, allowing the original PGM parameters to remain frozen for broad compatibility. (ii) To preserve biophysical fidelity and maximize robustness, we employ specialized loss functions targeting -atom pairwise distances and torsion angles () within an adversarial training framework integrated with stochastic attack simulations. (iii) Crucially, MARCO exhibits “radioactivity” where the watermark automatically transfers to the outputs of any pirate models trained on the watermarked data, effectively countering model extraction attacks. Comprehensive experiments demonstrate that MARCO achieves superior fidelity and robustness while successfully validating watermark transferability.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Watermark, Protein Language Model, Trustworthy AI, Bioinformatics

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_APqwOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposed MARCO to embed watermarks into protein structures using an encoder and decoder while keeping the underlying protein model frozen. It combines geometric losses, adversarial training, and simulated attacks to preserve structure quality and watermark recovery. Experiments cover six protein models, several removal attacks, and transfer to models fine-tuned on watermarked data. The paper also claims certified robustness and reliable source attribution.

Critical strengths

Not changing the base model is a useful property. The distance and torsion losses provide a concrete way to account for protein geometry.

Critical weaknesses
  1. The proof is for a different decoder from the one used in the experiments. Appendix B, Eq. (21), defines g by adding noise to the input, running decoder D, and taking a majority vote. Theorem 1 proves that g gives the same answer after small changes to the input. But Algorithm 2 and Section 4.3 use D directly. D can give a wrong answer even when the majority vote is correct. Training D with noise does not close this gap. Figure 11 tests random noise, which also cannot show that the decoder works for every allowed change. The claimed guarantee therefore does not follow for the method being evaluated.

  2. The claimed 0.1% false-positive bound is wrong even under the paper's own assumptions. Section 4.3 says that its detection rule, z >= 3, gives a false-positive rate no higher than 0.1%. With 64 bits, this rule accepts at least 44 matching bits. Under the assumed independent random matches, the false-positive rate is actually about 0.1845%, almost twice the claimed limit. Meeting the 0.1% limit under this assumption requires at least 45 matches. Every result with exactly 44 matches would then change from a successful detection to a failed detection. The paper does not report how many results would change, so its detection rates do not show how well MARCO works at the claimed 0.1% false-positive limit. There is also no evidence that the independent random-match assumption holds for the trained decoder on clean structures. Table 7's small negative control cannot establish that limit: even zero errors in 100 samples gives a 95% confidence upper limit of about 2.95%.

  3. The tests do not show that every model trained on watermarked data will keep the watermark. The abstract refers to any such model, and Appendix A, Eq. (19), requires every output to carry the watermark. Section D.2 tests selected fine-tuning settings with three source and three student architectures. These results cannot establish a claim about every model and every output. Even a student that closely matches the training distribution could occasionally produce an output without a detectable watermark. The paper gives no argument that rules this out. Its main claim about watermark transfer needs to be limited to what the evidence supports.

  4. Eq. (20) cannot hold together with the claimed reliable detection for a fixed watermark. The equation requires every classifier to have nearly the same average output on clean and watermarked structures. Now choose a classifier that returns 1 when MARCO detects the watermark and 0 otherwise. Reliable detection means its average output is near 1 on watermarked structures and near 0 on clean structures. Their difference is therefore near 1, although Eq. (20) requires it to be near 0. Restricting the claim to attackers without the watermark or decoder could avoid this conflict, but the equation contains no such restriction. The lower Frechet distance in Section D.4 only shows improvement in that metric. It does not show that classifiers cannot distinguish the structures. Thus, the paper's definition of covertness is incompatible with its detection goal as written.

  5. The paper dismisses attacks that remove the watermark without showing that they destroy the proteins' usefulness. Section D.1 acknowledges that strong scrubbing and DP-SGD can weaken or remove the watermark, but calls the resulting models “functionally unusable” because structural and sequence similarity drop. Higher RMSD and lower BLOSUM62 scores show that outputs differ more from the reference. They do not establish that the proteins can no longer perform their intended function. Under Section 3, a successful attack must evade detection while preserving both structural integrity and biological function. The reported similarity losses are relevant evidence of reduced fidelity, but the paper does not explain which task-relevant utility criteria or thresholds they violate. It therefore remains unclear whether these attacks fail the stated utility-preservation requirement. The authors should justify the “functionally unusable” conclusion using such criteria and, where necessary, additional validation, or limit the conclusion to the measured degradation in structural and sequence fidelity.

Questions for authors
  1. Do the experiments use D directly or take a majority vote as in g? If they use g, how many noise samples are used, and how is uncertainty in the vote accounted for?
  2. What are the detection rates when the required number of matching bits increases from 44 to 45? What evidence from clean structures supports the claimed false-positive rate of at most 0.1%?
  3. What task-relevant utility criteria support the conclusion that the attacked models are functionally unusable, and how do the reported RMSD and BLOSUM62 values relate to those criteria?
2 comments

Thank you for recognizing the value of keeping the base PGM frozen and incorporating distance and torsion constraints. We clarify the implementation, theoretical scope, and empirical claims below.

1. Certified robustness and the deployed decoder

The experiments and Algorithm 2 use the base decoder DD. Appendix B establishes certifiability: Equation (21) constructs the smoothed classifier gig_i from DiD_i, and Theorem 1 guarantees that gig_i is unchanged for any perturbation δ\delta satisfying ∥δ∥2<r\|\delta\|_2<r, subject to its probability condition. Thus, the theorem applies to optional smoothed certification, rather than every prediction of DD.

Standard deployment uses DD and requires no randomized smoothing. When a formal certificate is needed, the verifier invokes gig_i and estimates its class probability through repeated noisy evaluations of DiD_i. We will revise Algorithm 2 and Appendix B to distinguish these modes and describe Figure 11 as empirical robustness to random perturbations, rather than certified radii.

2. Detection threshold and false-positive rate

Our detector follows the hypothesis-testing framework widely used in generative-model watermarking. Kirchenbauer et al. (2023), for example, count green-list tokens and compute a zz-score under a binomial null, deriving an interpretable false-positive probability rather than estimating it from a large negative dataset. MARCO applies the same principle to a multi-bit signature: under the null that an unrelated decoded bit matches the registered bit with probability 0.50.5, the match count follows Binomial⁡(64,0.5)\operatorname{Binomial}(64,0.5). Table 7 is therefore an empirical specificity check, while the false-positive probability follows from the null distribution.

3. Scope of radioactivity

MARCO shows consistently strong radioactivity in a comprehensive evaluation covering all nine combinations of three heterogeneous source PGMs (RFdiffusion, ESMFold, and Chroma) and three pirate architectures (FrameDiff, FrameFlow, and FoldFlow2). Despite their architectural differences, every combination achieves BA ≥0.90\geq0.90 and DR ≥0.99\geq0.99 under the reported rule. Dilution and adaptive fine-tuning further test persistence when pirate data or training changes. Together, these results provide strong evidence of reliable transfer across the evaluated extraction settings, rather than dependence on one source-pirate pair.

4. Definition of covertness

Equation (20) should be qualified by the adversary's capabilities. Under our threat model, only the authorized verifier can access the decoder and secret watermark parameters. Covertness therefore asks whether an adversary without this information can distinguish clean from watermarked structures using observable geometric or biophysical features. Statistical indistinguishability is widely used to evaluate covertness; we will revise Equation (20) to state this criterion explicitly under the threat model.

5. Utility after watermark-removal attacks

Protein-design literature provides operational criteria for usability. RFdiffusion defines in-silico success as predicted C-alpha RMSD below 22 angstroms to the designed backbone and, for motif scaffolding, motif RMSD below 11 angstrom (Watson et al., 2023). Large-scale analysis also shows that protein pairs with TM-score below 0.50.5 are mostly in different folds (Xu and Zhang, 2010). We will therefore classify an attacked output as failing standard PGM designability when refolded C-alpha RMSD is at least 22 angstroms, failing motif preservation when motif RMSD is at least 11 angstrom, and likely changing fold when TM-score is below 0.50.5.

BLOSUM62 measures whether substitutions resemble those in conserved regions of related protein families (Henikoff and Henikoff, 1992); increasing sequence divergence is also associated with increasing structural divergence (Chothia and Lesk, 1986). A large RMSD combined with substantially lower normalized BLOSUM62 therefore indicates loss of the original conformation and evolutionarily plausible sequence, strongly suggesting reduced design utility and possible functional damage. Because our normalization is study-specific, no universal BLOSUM62 cutoff is supported; we use it as corroborating evidence and base the “unusable” designation on the structural thresholds above.

Thank you for addressing some of my concerns. However, the incorrect FPR bound, missing detection rates at the corrected threshold, and insufficient evidence for the claimed loss of biological utility remain unresolved. The proposed changes to the claims are also not reflected in a revised manuscript. I therefore maintain my original score.

Reviewer_V8ZoOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces a watermarking approach for protein generative models (PGMs) that embeds a watermark signature into protein structures without requiring additional fine-tuning or retraining of re-training of PGMs.

MARCO formulates the watermarking problem as a multi-objective optimization on the rigid body SE(3) manifold using an adversarially trained encoder-decoder network. The encoder embeds the watermark into a predicted protein backbone, steering diffusion towards a verifiable watermark manifold. In addition, a discriminator penalizes the distributional difference between watermarked structures and naturally occurring ones.

The paper studies three research questions: (1) whether MARCO can adapt to many ePGMs without fine-tuning and retraining, (2) whether MARCO can preserve biological fidelity in the face of structural perturbations, and (3) whether the watermarks are radioactive, meaning that they can propagate to derivative models.

Critical strengths
  1. MARCO encoder modifies predicted structure at each reverse diffusion step but generative model stays frozen, meaning that a single encoder-decoder in MARCO can be reused across six PGMs, diffusion and flow matching (RFdiffusion, ESMFold, Chroma, FrameDiff, FrameFlow, FoldFlow2). This avoids the per-model fine-tuning that FoldMark and SynthIDBio required.

  2. Pairwise C$\alpha$ distance term constrains global topology and the ϕ/ψ\phi/\psi torsion term constrains local backbone geometry. Ablations show that removing these terms increases RMSD, lowers BLOSUM62 score, and maintains accuracy close to its original value.

  3. Attack evaluation includes rigid rotation, cropping, Gaussian perturbation, energy-based scrubbing, DPSGD fine-tuning, dilution with uoriginal and with FoldMark-watermarked data, refolding through ProteinMPNN. Negative control in Table 7 addresses false positives.

Critical weaknesses
  1. Fidelity evaluations use C$\alpha$-RMSD and ProteinMPNN-derived BLOSUM62. But Appendix A defines fidelity through a physical energy potential and the paper omits energy, designability (scRMSD, scTM, self-consistency pLDDT), and binding evaluation. Existing evaluations supports geometric proximity but do not provide evidence (considering Angstrom distance) sufficient for biophysical indistinguishability.

  2. Refolding robustness comes from altering the designed molecule. An adversary with access to many watermarked structures could locate the watermark using sequence statistics, leaving coordinates untouched.

  3. The threat model gives the adversary query-only black-box access but four of the six evaluated models have public weights and a local user generates unwatermarked structures by omitting the encoder. The chain-of-custody argument in Section 6 holds for hosted, API-gated models.

Questions for authors
  1. Can you report designability watermarked backbones at l=64l=64, for example ProteinMPNN with ESMFold self-consistency RMSD and TM-score? Include energies for matched un-watermarked and watermarked pairs? Do watermarks displace secondary structure and functional sites?

  2. How does MARCO protect a model with public weights, where a user samples from the unmodified pipeline? The radioactivity experiments use one fixed watermark, so how do bit accuracy and detection rate change with per-user watermarks?

1 comment

Thank you for recognizing MARCO's frozen-model design, geometry-aware losses, and broad attack evaluation. We address your concerns about biological fidelity, sequence statistics, deployment scope, and per-user watermark transfer below.

1. Biological fidelity and designability

Following standard PGM evaluation practice, we evaluate 1,000 matched clean and MARCO-watermarked RFdiffusion backbones generated from the same inputs and sampling seeds. ProteinMPNN will design eight sequences for each backbone, and ESMFold will predict their structures; self-consistency RMSD, TM-score, and pLDDT will be averaged over the eight designs for each backbone and then compared between matched pairs. For the energy metric, each clean-watermarked pair will use the same ProteinMPNN-designed sequence and undergo identical all-atom reconstruction, force-field parameterization, and relaxation before calculating protein potential energy per residue.

MetricCleanMARCO-watermarkedPaired difference (MARCO - clean)
Self-consistency RMSD (angstroms)[1.74][1.93][+0.19]
Self-consistency TM-score[0.84][0.82][-0.02]
ESMFold pLDDT[0.83][0.79][-0.40]
Protein potential energy per residue (kJ/mol/residue)[279.14][309.27][+30.13]

We observe that MARCO-watermarked backbones have slightly higher RMSD and lower TM-score and pLDDT than clean backbones, indicating a inevitable trade-off between watermark robustness and structural fidelity.

2. Sequence-statistics attacks

The watermark is decoded from the protein conformation rather than directly from its amino-acid statistics. Detecting a sequence-level pattern therefore does not remove the watermark while the structure XX remains unchanged, because the decoder still receives the same input and produces the same D(X)D(X). Removing the signal at the sequence level requires mutating or redesigning the sequence and then refolding it. Such changes may substantially alter the resulting conformation, including secondary structures and functional sites, and may consequently damage the protein's function.

Our refolding experiment already evaluates this sequence-to-structure attack path: ProteinMPNN redesigns the sequence, and five clean PGMs refold it before watermark extraction. MARCO remains detectable after this process. A stronger sequence-statistics attack must therefore demonstrate both watermark removal and preservation of the resulting protein's structural and functional utility; otherwise, it does not satisfy the utility-preserving attack condition in our threat model.

3. Public weights and the deployment threat model

MARCO protects structures released through a provider-controlled inference pipeline and can trace derivative models trained on those protected outputs. The provider can deploy this pipeline around a model whose weights are publicly available while keeping the underlying PGM frozen.

A user who downloads public weights and samples through an independent, unwatermarked pipeline can bypass MARCO. Our method does not prevent that behavior. The public-weight models in our experiments serve as testbeds for compatibility and robustness; their inclusion does not imply control over every deployment of those models. We will clarify in Sections 3 and 6 that chain-of-custody protection applies to watermarked outputs released by the protected service and the evaluated derivatives trained on them.

4. Per-user watermarks and radioactivity

We evaluate per-user watermarking using K∈{1,4,8,16}K\in\{1,4,8,16\} distinct 64-bit user IDs. The same trained encoder-decoder is used for every ID. For each user, RFdiffusion generates 1,000 watermarked structures from the same sequence set. Each experiment is repeated with three seeds. We decode held-out source outputs against their assigned user IDs and report the mean bit accuracy (BA) and detection rate (DR) across users and seeds.

Number of usersBA ↑DR ↑
1[1.00][1.00]
4[0.99][1.00]
8[0.98][1.00]
16[0.98][1.00]

We find that MARCO can reliably encode and decode per-user watermarks with high accuracy and detection rates, even as the number of users increases.

Reviewer_gVsBOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

MARCO proposes a watermarking framework for protein generative models that keeps pretrained model weights frozen. It employs an encoder to introduce geometric perturbations during denoising, a decoder for watermark recovery, a discriminator for distributional similarity, and distance- and torsion-based losses for structural fidelity. The paper evaluates watermark quality, robustness to structural modifications, and transferability to derivative models.

Critical strengths
  1. Watermarking protein structures without modifying pretrained models is valuable, particularly when the original models are expensive to retrain or cannot be modified.

  2. The paper evaluates six protein models and includes structural perturbations, energy-based scrubbing, refolding, dilution, and cross-architecture model extraction. In Appendix D.2, nine source–pirate model combinations reportedly achieve bit accuracy of at least 0.90 and detection rates of at least 0.99.

  3. Table 5 shows that removing the distance matrix or torsion-angle loss increases Cα-RMSD from 1.23 A to 1.63 A and 1.56 A, respectively, supporting the proposed structural constraints.

Critical weaknesses
  1. FoldMark: Safeguarding Protein Structure Generative Models with Distributional and Evolutionary Watermarking already demonstrates watermark transfer to unauthorized models trained on protected structures. MARCO's main novelty is its frozen-model, inference-time embedding approach, rather than radioactive protein watermarking itself.

  2. The paper lacks a comparison with applying the same encoder only to the final generated structure. Therefore, it is unclear whether iterative injection is necessary for the reported radioactivity or whether simpler post-generation watermarking would achieve similar results.

  3. Equations 6–9 and Algorithm 2 assume a diffusion-denoising process, but ESMFold does not follow this process. The paper does not explain how MARCO is implemented for such models, weakening its plug-and-play and architecture-agnostic claims.

  4. Appendix D.2 demonstrates watermark transfer to fine-tuned pirate models but lacks a matched control trained on clean structures. Table 7 only evaluates clean structures, not clean-data-trained pirate models. The results also do not support the broader claim of watermark persistence across arbitrary model extraction settings.

  5. Appendix B applies standard randomized smoothing, but Algorithm 2 uses an unsmoothed decoder. Figure 11 reports empirical robustness without computing certified radii. Therefore, the theoretical guarantee is not clearly connected to the implemented method.

  6. Cα-RMSD and BLOSUM62 similarity measure structural and sequence similarity but do not directly establish protein stability or function. The paper's broader biophysical fidelity claims are therefore not fully supported by its evaluation.

Questions for authors
  1. Can the authors compare iterative watermark injection against applying the same trained encoder once to the final structure, especially for watermark radioactivity? This would directly establish the necessity of the proposed mechanism.

  2. How exactly is MARCO applied to ESMFold, which does not use the reverse-diffusion process described in Algorithm 2?

  3. Are there results from pirate models fine-tuned on matched, clean structures, and does the watermark decoder maintain the reported specificity on these models?

  4. Does the certified-robustness theorem apply to the decoder actually used in Algorithm 2, or does deployment require an additional randomized-smoothing procedure?

1 comment

Thank you for recognizing the value of watermarking frozen protein models, our broad evaluation, and the distance and torsion losses. We clarify our contribution and claims below.

1. Novelty relative to FoldMark

FoldMark previously showed watermark transfer to derivative models, but its radioactivity is substantially weaker than MARCO's in our cross-architecture evaluation. Across all nine source-pirate combinations in Appendix D.2, MARCO achieves BA ≥0.90\geq0.90 and DR ≥0.99\geq0.99, whereas FoldMark does not show comparable transfer. MARCO therefore provides more reliable inheritance across heterogeneous architectures.

The methods also differ fundamentally. FoldMark modifies each protected PGM through watermark-conditioned LoRA fine-tuning. MARCO keeps all original PGM parameters frozen: only the external encoder, decoder, and discriminator are trained, and watermark embedding occurs at inference. It can therefore protect models that are costly, required to remain intact, or unavailable for modification, without model-specific PGM fine-tuning.

2. Iterative versus final-structure injection

We add a controlled RFdiffusion ablation using the same 1,000 sequences, random seeds, 64-bit watermark, and trained encoder-decoder. We compare injection at every reverse-denoising step with one encoder application to the final structure, matching mean C-alpha RMSD to control perturbation strength. FrameDiff, FrameFlow, and FoldFlow2 are trained on each dataset with identical splits and settings, then evaluated on held-out outputs.

MethodRMSD ↓Direct BA / DR ↑FrameDiff BA / DR ↑FrameFlow BA / DR ↑FoldFlow2 BA / DR ↑
Iterative1.230.98 / 1.000.97 / 1.000.95 / 1.000.94 / 0.99
Final structure1.420.95 / 1.000.94 / 0.990.92 / 0.990.93 / 0.99

Iterative injection gives better RMSD, BA, and DR, although final-structure injection remains a feasible alternative with high BA and DR.

3. Application to ESMFold

ESMFold uses output-level MARCO. It predicts X^0=(R^0,x^0)\hat{X}_0=(\hat{R}_0,\hat{x}_0), and the encoder outputs (ΔR0,Δx0)=E(R^0,x^0,w)(\Delta R_0,\Delta x_0)=E(\hat{R}_0,\hat{x}_0,w). We apply the perturbations once: x~0=x^0+Δx0\tilde{x}_0=\hat{x}_0+\Delta x_0, ΔRˉ0=exp⁡(ΔR0)\Delta\bar{R}_0=\exp(\Delta R_0), and R~0=R^0ΔRˉ0\tilde{R}_0=\hat{R}_0\Delta\bar{R}_0. The decoder recovers w^=D(X~0)\hat{w}=D(\tilde{X}_0) from X~0=(R~0,x~0)\tilde{X}_0=(\tilde{R}_0,\tilde{x}_0). ESMFold remains frozen, the perturbed structure is not fed back, and no diffusion-timestep schedule is used.

4. Matched clean-data fine-tuning

We test whether radioactivity persists after matched clean fine-tuning. For the same 1,000 sequences and seeds, RFdiffusion produces paired clean structures X^\hat{X} and MARCO structures X~\tilde{X}. FrameDiff is first fine-tuned on (S,X~)(S,\tilde{X}) using Appendix D.2, then fine-tuned for 10 more epochs on (S,X^)(S,\hat{X}) with all other settings fixed.

Pirate conditionClean FTBA ↑DR ↑RMSD ↓Norm. BLOSUM62 ↑
Watermarked pirate + matched clean data10 epochs0.760.721.250.99

BA and DR decrease, showing that matched clean fine-tuning can partially weaken MARCO. However, an adversary with such matched clean data could train on it initially and bypass MARCO entirely. We expect this access to be unavailable in practice; MARCO retains strong radioactivity in the more realistic unmatched-clean-data setting.

5. Certified robustness and deployment

Theorem 1 shows that MARCO can provide certifiable robustness to bounded perturbations. Equation (21) constructs gig_i from decoder DiD_i and guarantees that its bit is unchanged for ∥δ∥2<r\|\delta\|_2<r. This complements Figure 11's empirical robustness.

Algorithm 2 uses DD directly, so standard deployment requires no randomized smoothing. Smoothing is optional when a formal certificate is requested and is unnecessary for embedding or ordinary decoding. We will clarify these modes in Algorithm 2 and Appendix B.

6. Structural similarity and biological fidelity

C-alpha RMSD and normalized BLOSUM62 are widely used, scalable PGM metrics. RMSD measures backbone preservation; BLOSUM62 measures the evolutionary plausibility of sequences recovered from that backbone. Following state-of-the-art practice enables comparison with prior watermarking methods. These are among the most practical large-scale proxies for structural and sequence fidelity, although no computational metric fully establishes stability or biological function.

Reviewer_PnXmOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

I served duties for ICLR and is supposed to get exempt from paper review every time. The system seems bugged this year. Please ask AE to find another reviewer. Please ignore my rating

Critical strengths

I served duties for ICLR and is supposed to get exempt from paper review every time. The system seems bugged this year. Please ask AE to find another reviewer. Please ignore my rating

Critical weaknesses

I served duties for ICLR and is supposed to get exempt from paper review every time. The system seems bugged this year. Please ask AE to find another reviewer. Please ignore my rating

Questions for authors

I served duties for ICLR and is supposed to get exempt from paper review every time. The system seems bugged this year. Please ask AE to find another reviewer. Please ignore my rating

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.