29th Session of the IPHC Scientific Review Board SRB029, Day 3

Okay, so welcome back, everyone. We're going to go through the draft of the SRB report, and I think we'll start here in Section 3.4, the observer updates. The previous sections are the standard annotations that we've seen and noted the papers that were presented. So the observer updates, Ann Marie, we discussed this yesterday. It seemed like we'd adequately captured those here.

Is that right?

Yes, thank you. Maybe one thing is that the— instead of the AI aging approaches require verification, it was more about how that would be done.

Okay, great.

And we had a presentation on that yesterday from Basia about the verification approaches there.

So, I will try to change this one. It is, yeah.

Okay, um, moving on then, um, for, uh, research biology and ecology, noted the, uh, presented, um, and just a note here in 11, uh, that it's really, really helpful to have those tables for, um, different aging approaches, and it would be more helpful to have them combined in a single table so we can compare across the three.

Well, this is related to our discussion yesterday about trying to come up with a way of looking at trends in, well, depredation, whether it's geographic patterns or temporal trends. We talked quite a bit about the limitations of existing data, and one of the options that was discussed was sentiment survey, which is new, right, Malsha? Yes, we pretty much did the trial during the conference board meeting last year, a large group of stakeholders in one place, in a paper format, and then we set out to do the same survey in Ponoponika. So the survey in the electronic format is distributed with the brochure and the archives. Great.

So we talked about whether, whether we could use that sentiment survey, adding some additional questions about whale interactions on this recent trip, for example, to get more fishery-dependent data on our actions.

So the request here is just that the Secretariat think about it and at the next SRB meeting provide some options or decide whether this is feasible and be developed further. Yeah, we'll need to discuss a little bit whether— the goal there was to make it really easy and simple and kind of COVID that more by full year sentiment. So there's, there's already questions that, that it asks about the particular issue, that which issues you find as a particular, particular concern in this year. So that will give you like an indication of over time whether the whales are going to become more of a concern or less. That would be part of that kind of overall sentiment about what are the currently pressing issues.

It's on something more— it's like a little bit different type of question if you want to ask something specific to last trip, but this is part of the interview with the client and that the field staff contact when they collect the log. So they do ask So maybe, maybe we can look a little bit more about how these questions are asked as a part of the walk-through interview. Okay. More, but if you want something that you're interested in, something trip-specific versus the sentiment survey, that was really the goal is to get a—. Yeah, I mean, I don't think the SRB necessarily wants trip-specific information.

It's just whether, whether there's potential to survey to develop better understanding of real definition, whether at the trip level or some other level.

Yeah. Okay. Um, moving on to 13. Um, so we, we discussed this, uh, 13 and 14 are related to methods used in developing the epigenetic aging approach. Um, two, two suggestions there, uh, to make the models sex-specific for '13 and for '14 to use the Elastic Net model to essentially do dimension reduction and further developing the models.

So there's a note there about timelines. We don't have timelines in there on this, partly because we weren't sure what the appropriate timeline is relative stage that you're at now in the epigenetic, uh, aging development? Is there a timeline that makes sense? My hope is that by next June this will be— we'll be able to report on— so, um, So just add this to— okay, moving on here. And Anna and Mike, please feel free to jump in with clarifications and join the course, but I want to make sure that Anna and Mike can jump in if they need to.

For 15, this is— it was really useful to have in our discussion to hear what the next steps were in the, the fecundity analysis, how there's— how the Boff hypothesis will actually be tested. So just a just a request to fill in those final steps in the analysis. Sounds good.

Okay. 16, It should be recommended evaluating the feasibility of collecting data on energetic condition of Calibut through liver weights and liver lipid composition.

So, I think there's a request here to put the timeline on there. Joseph, we hear something back on this in June? Obviously, there can't be any new data collection. Yeah, in June, the data collection will not have started because this is— we've discussed the possibility of maybe doing a pilot testing of the feasibility in 2022. So, have a conversation with the FIS team and see how that's feasible.

So, okay, great. Hopefully, hopefully the plans can be delineated, uh, at, uh, SRP 30.

Okay, um, 17, uh, I, I just added this, um, so the, the SRB has not seen this yet, uh, so I added this based on our discussion just now, um, that we talked about, uh, potentially using growth increments for back-calculating individual growth, and then also using microchemistry to identify natal origin. It's a big hell of a— it sounds like there's been a lot done on these topics, and I think it would just be useful to go back in the notes, see what's been done, and if you guys have gotten to a point where you've concluded that this is not viable, just make sure that there's no new approaches that might make that more viable than it was in the past. That might include, for example, Colin having a discussion with— I'm blanking on his name now— Levi at UC Davis, and potentially looking at newer autoimmunochemistry approaches. A bit more literature review to see whether, um, molecular chemistry is being used on other species within this geographic area and what scales, the scales at which they can distinguish. Yeah.

I might add to A, for the young, like the focus on the early part of the age.

Joseph, the juveniles, what do you consider juvenile?

Juvenile is an immature fish, so a little bit before the troll samples, right? Uh, no, no, I would consider juveniles as, uh, the ones that we are catching on the troll. Okay. Yeah, is there a better word here for me? You understand?

Younger fish. Younger fish. Younger fish.

Okay, great. Moving on to stock assessment here. This will be noticed in the paper and presentation.

We really enjoyed the presentation on CL model development by Caleb, and I think that it'll be great to continue on in that path and then also to see whether WAM and SPOARC might have features that are still to be explored. Just one—. I think the last request had a very specific SRB. Is it SRB 32 or something? When they, when they said investigate SS, state space models?

Yeah, the goal being to have something prepared for the next, you know, whatever SRB that is. That's 32. Yeah, do we, we need a timeline for this request here?

I'll reference the existing request as well, which will stay as partially pending. Yeah, so that will, that will track it. Yeah, we're just updates on, yeah, on what you're, what you already put Yep. Okay, 20 was added here, a request to just continue to display uncertainty along with the point estimates in the documents.

And if possible, to look at their uncertainty assessment output changes with stock status. And I think the concern here is increasing uncertainty during periods of decline, whether that's a pattern that we shouldn't be looking out for.

We had some discussion about this, and we certainly acknowledge that most of the figures do include confidence bounds on both sides of the output. Yeah, I'm not actually aware that we produce any point estimates. Don't have uncertainty because we're integrating across the— for all of the decision tables is the uncertainty. So yeah, I guess, okay, you might add continue to display uncertainty instead of display it, lest somebody get confused that we aren't, but Okay, um, 21, moving on. Um, just a note that the sex composition information that's, uh, it's coming out now from fishery landings is really important to the assessment.

That's important.

MSC, uh, first noting the, the presentations and paper there. Um, so 23 here I think is, uh, is a tricky one. We are, I guess, just requesting that keep thinking about specific ways of defining depleted, and if that definition includes terms like recovery, that recovery also be defined specifically. So, you know, this just follows our discussion of how to operationalize the concept of the depleted reference point, and I think we've had a lot of good discussion about that. You want some ideas, Alan?

24. I'm sorry, go ahead. Did we, did we put the part about potentially taking recovery out and using continued decline as a basis rather than recovery in one of the other ones? Or did it just get lost somewhere? I think that got lost.

So let me try to add that here.

Does the sex ratio change dramatically with productivity in the areas?

Do you know?

Strong correlation between sex ratio and catch.

Age. So when we have a, when there's more fish and the age structure is more protracted, tend to catch more males because their sprees are longer.

Mike, does that capture it there? Yeah, I think so. It's just basically the idea that instead of thinking about it as trying to think about what levels promote some sort of increase, thinking about what levels might lead to long-term continued decline, and that that might be an easier way to define that depletion threshold. Yep.

Okay, great. Moving on to 24, talking about how the projections require some assumptions about assessment information implementation uncertainty. I think we have received presentations on this in the past, but those predate current SRP members, so it'd be, I think, useful to to go back and see how those are being included at this point. I think particularly the implementation uncertainty, we had a lot of discussion about that 3 years ago. Yeah, I think so.

And that's something that might be worth just reviewing the next time we start meeting. And I particularly like the potential alternative approaches.

Okay, 25. I put recalling here because I know we've talked about this before, but options for implementing an assessment model within the, the MSC, including potentially using some of the state space options that might be able to run faster than I looked up that previous request and now I can't remember what it was. Find out again. I said I should write this down, I'll remember it. Might have been 27.

That's— we'll find Okay, we talked about 26 already. Think about PDO projections based on time series analysis. So, the time series characteristics of the projections are similar to those of the historical record.

So, that's pretty big on time series analysis, but you trust me to do something. We trust you and look forward to seeing what you— We had enough discussion of that. Yeah, I understand what we're looking at. Yeah, it equilibrated at 50/50, which is not consistent with historical records.

So, I'll probably come back with a couple ideas on that.

27. Yeah, we— Annemarie, we spent a lot of time talking about whale depredation, I think, based on the the comments that you brought up at the beginning there, and one of the ideas was to think about whether the MSC could evaluate scenarios where whale depredation is increasing and determine whether the harvest strategy is robust and the assessment advice are robust to potential increases or current underestimates of, uh, welfare beneficiaries.

With this one, I like this one. I'm really interested in doing this type of analysis, but I'm curious why the SRB put this as recommended rather than at this point. Uh, we put it as recommended because we wanted the Commission to see this and see that we had heard and were being responsive to, um, to the feedback that we got from, uh, okay. Does that, does that create problems? I don't think it creates a problem.

The only issue that I was thinking is the program of work is pretty full right now, and recommending— it may have them think about the prioritization of all the other So, if you want to keep it recommended, I think that's fine, but maybe a timeline would be helpful just for me because, you know, we got new estimation model, reference line. There's a lot of things on the list. Yeah, I was going to suggest maybe indicating some sort of priority on this, given Alan's very full plate.

But yeah, I think that would be great. I think it indicates a way forward.

When we were talking about this option, that's why the first part's written the way it is, to evaluate the feasibility rather than to actually do it. And so the other wording we had talked about doing, so if you like this better, perhaps we switch it. Is to think about how you would further develop the MSE structure so that in the future you could include citation effects. And so basically just trying to in the future make it such that it, it's an easier option to consider if it continues to be high priority. Yeah, that's, that's a really good point, Mike.

And I guess I, did— that didn't really notice the feasibility. And I wonder when people do see this recommendation, they may think that, oh, this analysis will be done.

And I'm usually not one to argue against doing work. I'm just trying to protect myself at this point in time.

But conduct— instead of evaluating feasibility, conduct an additional feasibility analysis. Analysis. I guess we didn't want to make it sound too formal either. I mean, we're not, we're not, yeah, we're not looking for like a whole bunch of analysis. Just as Mike said, you know, think about whether the structure of, of the FSC allows for this analysis in the future or not.

I wonder if the new text that you wrote, perhaps instead of saying fully implementing, because a feasibility I could do pretty pretty quickly. Maybe that text you wrote would say something like, other items on the MSE work plan, and, and noted that this is not— or something along the lines like, this is not, this is not, we're not expecting conducting a full simulation analysis, but understanding that this can be part, can be done within the This MSE framework or your MSE framework, or something along that line. You get what I'm saying? I don't know exactly how to word it. Yeah, I mean, I've also just italicized feasibility and implementing.

Okay.

Okay. Yeah, that works. Feasibility studies and less feasible, you know. But feasible by itself.

Oh, that's helpful.

Uh, I, I don't think you need to write.

Okay, moving on then to, uh, best design.

So we have here in 29, just a note for presentations, but 30 here, 30 here, we continue to see 4B in the design and then through no fault of the Secretariat, it's not part of what actually gets done. So we're wondering whether there's whether we're at the point where it's time to really think more broadly about options for how this could get done. And perhaps something like interviewing fishermen in that area to see what it would take to make this appealing, what their constraints are for doing this. I mean, it sounds like you guys have an idea of constraints already. It's really far away and there's no processor, processor there anymore.

But trying to see whether, whether there's something the Secretariat could do to make this more viable, or whether there are alternative methods for gathering information from Florby, because it seems like right now it's just the approach is just put it in the design, see if you get bids, don't get bids. Wouldn't sign again next year. At some point, that's probably going to kind of cause trouble, especially with the concerns about bias creeping up in, in 4B. The SRB totally understood the fact that 4B has historically represented a pretty small proportion of the coast-wide stock, but the bias concerns that have been discussed at previous SRP meetings are still there. I would, I would just say if you're going to have 2019 in there, I would maybe include that the western end of IPHC regulatory or 4B has not been sampled since 2019.

Thank you. We have had sampling there in 2022 and 2025, but those were more like central and eastern parts of 4B.

Thanks. That's helpful.

Ray, any, any thoughts or feedback here?

Well, missed his chance.

Okay, moving to the space-time modeling then.

So we talked a little bit about this yesterday, that SDM-TMB, the newer package that Sean Anderson developed, could offer some advantages to the current, our in-law framework, and just, we're encouraging Ray to continue looking in that direction and seeing whether this is a potential time for a change there.

That's been an ongoing discussion.

Talk about that.

Okay, um, we in 32 requested that future displays of the space-time model outputs also include maps of uncertainty, essentially just help us better visualize the, um, the results of not having more recent samples in some of the outlying distribution. Moving on to the AI-based modeling that, that Basia presented, we in 33 noted the presentation there, and I think we're very excited about the performance of the 50/50 mix. I should evaluate it.

I think you want to break and bake.

Break.

So in 34 is just the note. 35 Was a request to think— continue thinking more about the trade-offs associated with using a mixed approach and think about what that optimal mix might be. And that might involve looking at how the uncertainty carries through to the assessment outputs. What is an acceptable level of uncertainty? It also involves operational considerations.

About how much time saved doing a mixture. Yeah. And as I said, the 50/50 was a kind of middle ground between like what would actually make a difference to manual aging versus what would seem not to be affecting the aging error substantially. Let me see.

No, I'm not sure how—. I'm sure we're ready to do assessment simulations. I mean, at a 50/50 mix, we've got effectively the same properties. That could be a huge amount of work to tell us that it's basically just like break and bake. Um, go ahead, Mike.

I guess I see you as at sort of a fork here, and if you go down this road I don't expect you're going to go back to all manual aging, and this seems like a really big step to me. And so, as well as I, I guess I have a fundamentally different view of the curves that were being shown yesterday about how similar they are. They look quite different to me.

Um, yeah, maybe we should look at that again because I think we're looking at different different things there. Okay. Maybe, maybe I'm— maybe what I'm looking at isn't the right thing. I think you made reference to some plots that were shown in the earlier meeting. That 30% and an 80% that I hadn't seen.

I don't think you showed them yesterday. Would you like me to quickly pop— except you have Oh, you have— can you go to our website and we can— to make sure we're talking about the same plot. Yeah. And so the other thing that we were talking about that I thought I recalled from last meeting was that I thought that with the break and bake, you're— or bake and break, you're getting, um, about an 80% concordance on the same age, whereas with the AI it looks like it's about 80% for plus or minus 1 age. And to me, that's, that's a pretty big difference.

So, but if I'm not remembering things right, that's entirely possible. My memory is not reliable these days. Dr. Meating?

I think that's all.

Just go to the current one.

The little pencil paper.

And the PowerPoint of the second one.

Okay, I open that. Let me switch the sharing.

Because they are derived very different way. The one that we presented with the mix versus just a simple comparison of the reference. Yeah, the comparison we need to look at is that aging and precision that's estimated properly, not the percent agreement. Because that's— doesn't include all of the information. You just need that first graphics that you had.

I'm not able to— slide you one.

Not the one before. This one. Yes. So we're talking about the difference between the black and the blue on this plot right here.

The black being the current—. I would think You're confusing it with the red. My apologies. Yeah, sorry. The red is the historical surface aging, which we actually do use those data, but that was— that's only prior to 2002.

And those are the 95% confidence intervals based on the full analysis of aging and precision, including all the probabilities of both being right, both being wrong. And this is done very different because this is like the first read against the second read for break-bake or first read against the second read or second read replaced with AI for other 50%. The—. I mean, there's definitely some consideration, as Varsha said, 50/50 is— was kind of an arbitrary choice, but the considerations are almost as much logistical as they are scientific at this point. I mean, we— with, with this level of precision, this is— these are far more precise even at a 50/50 mix than most people's aging that they use in their assessment, or most assessments that I've ever done.

So really, what, what we think with 50/50 we're looking at is an amount that's enough of a reduction to actually gain something in the lab, but also enough that we maintain plenty of manual aging. And we're— it's, to me, there's more to think about in terms of how we apply this across our different datasets and how much effort in manual aging we're going to have going forward and how we accommodate that. So we have fishery ages, survey ages, trawl ages, and recreational ages that we do every year. Year. And we can maintain, for example, we can maintain our survey ages being done manually and have those be the highest quality, and then some of these other datasets can start to be mixed in.

There's— we, we really don't know moving forward, I guess, what are exactly what our logistical capabilities are going to be. And the 50/50 method was explicitly designed also to maintain the kind of adaptability. That's why we're talking about If you have like a good portion of the outputs to continue the age in a manual, using manual methods, you can use that for training. So you have also always the model that's not even trained altogether. So it always includes the current sample as well as— and you have that portion of this 10% always for quality check to monitor the pair.

Yeah, well, you can if you're not doing what Ian suggested and taking some portions that are AI-aged and other portions that are a different set that are manually aged, right? That I think maybe worries me in terms of the cross-validation and make sure that—. Sorry, I was proposing that we would do the 50-50 mix method. We're not talking about doing AI solely for anything. That's what I thought you were saying.

No, no, no, just applying this new method to various datasets. I think about we, for the logistical reason, we did need to subsample. We are currently subsampling the commercial, uh, otoliths just because there's just too many and capacity. We did that analysis. We did a power analysis looking at the effective sample size and where we could reduce the number of ages and maintain the effective sample spitters.

We've not discontinued collecting those otoliths because we feel like we could always go back and supplement if we find that there's an issue. There's very little cost to collecting a few extra otoliths in the ports. But then we don't need to necessarily age them all real time here. And if we get a year where things don't look right, then we could go back and supplement. So we're maintaining the sampling rate in the field.

And the, I mean, the other outstanding issue here is that we still haven't done out-of-year test. So when we apply this method, we've got a model trained up and we apply it to the next year, then that's— we still have a question about how it performs one year ahead. Yeah, but I think Dr. Moser—. Yeah, because the, the current year would be always included. So you would only do— take the images first, then you would first do the 50%, then retrain the model, and then use it on the second half.

The question is whether you should perform. I think the only question is why we want to take images of those multi-year collections to make sure the model doesn't lose the predictive power when it's trained on multiple years and predicting multiple years at the same time, whether it performs substantially better if it's just like trained on one year and predicting the same year versus if it's including even more images, but it's It's designed to be kind of one model that predicts age, but by the way how I see this 50/50 design, there are never predictions of the competing year. There's always some portion of the outputs for that year, but it's just a completely different set of those outputs. But it's only trained on one year right now. So we need to know how it performs when we start adding years.

Yeah, that's what I'm saying. Differently as we start to apply the same model across years. And we'll— and we could obviously create a new model for each year, but that's a different— that's not the method that we have right now. The method would be to develop, train this model, and then apply it and add years. And it would add training data.

You're right, it's not a completely out-of-sample test, but it's a test of how it performs over time. And that's, that's where I think we need to explore a little more to know because we don't want to learn that it's breaking down while we're using it. We want to know that ahead of time. But I would just recall some of the earlier analysis that I was showing on fine-tuning, but it was not part of it here. But what we were showing is, yes, a completely out-of-sample prediction in New Year was performing worse.

It was still not particularly bad, but it was not performing as well. As with the example. Well, what we were doing, what we were, we're just testing fine-tuning. So what if you have, you know, the main model during one year and you added the portion of the current year outlets to fine-tune? And this was actually what we call it fine-tuning because it was very short training of adding that additional information.

And then we're getting substantially better results of that sample. With just a small number of new ontologies from the current year added. So if you think about adding at least 50% of the current database based on at least the fine-tuning experiments, it's— it indicates that this potentially could be the results. So yeah, I think that the analysis that we've done on at least on the 2 years that we tested, we have, we have 2019 and 2024, and by now we have a little bit of a 2025 as well. But we show that if you do completely just 2019 and try to predict 2024, that the results were not as good.

But even if you add, and we were testing with just like a 10% fine-tuning, so just the 10% of the current year, we tried adding 10% of the 2024 and that's substantially. So that indicates that if we have a method that is continuously adapted to include the current year, it should, in principle, attain better level of prediction. Maybe. I mean, also the current year will represent an increasingly small portion of the overall training data as the develops. So it might not keep up because it's fitting to the bulk of the data, which is the past.

So I think this is something to just keep an eye on. Yeah. And then, but that's why we also discussed that, that's why the part of the quality control. So you can also look under this method to still have a capacity to calculate and recalculate the error every year, so that you can— because the model would be then applied to the whole. So you don't only apply a new model, right?

You recalculate those AI ages for all the years that you, you are using the AI ages, and then you have to recalculate the errors. Oh, that's interesting. So you'd have a revision every year of the whole time series? Why not? If you're having better model.

Does mean they have to redevelop the precision matrix right here.

So I guess the question would be also if, if it's changing substantially, that's what concern, but if it's just an update, making sure they're staying on top of the—. Okay, let's get, um, thanks for that, uh, Masha. Let's, um, get back to the SRB report here and maybe I think that's a no.

Just paying attention. Yep. Mike, what are you thinking here? And I hope that was helpful to explain, just making sure we're on the same page in terms of which analysis of the error we are referring to. No, no, that was very helpful, and I was misinterpreting it before, so that's good.

Given that, I do have a particular fondness for simulation studies, but I'd be fine if we drop that because I honestly, what I'm thinking of is something that's sort of analog. The, the big ver— there's two potential versions of it. One is the big version that actually does it almost in an MSE context, which is similar to kind of what Basia did on assessment frequency for summer flounder, which could actually say, okay, what's the— how much value might you actually be losing by changing to this? And even though it might only be 1 or 2%, that 1 or 2% might be meaningful. And if you don't ever do it, you don't ever know what you might be giving up.

But at the same time, I agree that particularly for this case, trying to figure out what the future AI error for reading the zoodolus is going to be is going to be kind of the key point in doing it. And that's not a trivial thing to to try and figure out.

Yeah, I mean, it seems like it's changing so rapidly, just the last results that you presented, Basia, compared to these results. So maybe something I'm planning into— maybe that's a good request, something for presenting to the examiners to make it more clear that you would be interested in seeing that, you know, we've taken the author list from, years and present them to your— because I think that will be pretty much the, the last piece to— for the decision, like, to have a—.

This is the problem.

Is that good? Yeah, that's fine.

Okay, that brings us to the end here. So, SRB members, if there's anything that we're missing, speak now or we'll hold your hands.

Likewise, Secretariat, if there's anything that you'd like to discuss or clarify, this is the time to do that.

Okay, moving on then to other business. Dave, we've got some options for the June and September meetings next year. Yeah, so we've put in a couple of options there. As you know, the Commission has moved the work meeting from time to time, and that's why this meeting is a little bit later.

Yeah, so for Survey 30 and 31. You've got two options each. See if there's a preference on people's calendar. If not, we can always just leave it open as well and decide it via correspondence.

Let's get a— just looking at our calendars here.

Earlier in June is better for me and William.

I was going to say earlier in June is better for me as well.

September, really early dates in September.

Mike and Anna, earlier is better in June for me as well. Okay. Yeah, I could do either in June, um, okay.

So June 1st to 3rd, and then for September, Joanna expressed preference for the 14th to 16th. Does that work for Anna and Mike? Yeah, 14th to 16th works for me. It's better.

Olaf, we don't have an SSC meeting during that time, do we? I was just trying to go back. I hadn't put those on my calendar yet. Sorry, I'm just looking at that, and we do, Mike.

The, the Atlantic SSC that, um, Mike and I are both on is scheduled for the 14th, 15th of September.

Joanna, what does, um, the 7th to 9th look like for you?

No, it's right after That's not ideal for me. It's a bit tricky just because it's the first day of everything at the university.

School is a bit more complex next year because I don't have one. So I mean, it's not impossible. It's tricky because that Monday, it's a holiday. —Getting there. What about the 21st to 23rd?

It's not there as an option, but does that work for others? That works for me. It's currently open for me too.

I'm wondering if it overlaps ISIS meeting.

Put it down as tentative. 21St to 23rd. 21St to 23rd. Double check, we do have to wait for the commission as well to set their work meeting schedule. Always be at last minute change.

Okay, so we'll write down 21st to 23rd for the external meeting, 21st to 23rd, tentative September. Okay, great.

Yeah, and if you could check on the IC's meeting and confirm your availability for that late September, that would be great.

Okay.

I think from— all set then. Edited and adopted report.

Thanks for hanging in there despite the, uh, time zone differences, Anna. No worries.

Right, well, hopefully we see everyone in person in, uh, June then. Thank you very much. Okay, we'll close the meeting. Thank you, everybody online. Thank you, everyone.

Safe travels back.