29th Session of the IPHC Scientific Review Board SRB029, Day 2 Part 1

I will jump straight into it and pass to the Chair, and he's going to walk us through the first session, which is sort of a recap of day 1 and some of the recommendations that are requesting being developed. So over to you, Chair. Okay, great. So, yeah, we have a couple things just to check in with. First of all, Ann Marie, I— sorry, I got your email after we had started working on this, but I wanted to check in to see if we captured this.

The text here. Yep, that looks great. Sorry, I meant to send that earlier and things just happened. No worries. No, yeah, we had the notes.

Okay.

So then our— under the biology and ecology, just noting the paper, Joseph and Basia. So there are two different tables there and we're going to just think that it'd be useful to have those side by side and in consistent formats. We can do the comparison and see where each, uh, which falls. I think that hopefully shouldn't be too much, uh, work to do. The information's there and it's just, uh, just useful.

Um, and just also noted that the Commission thought this would be helpful to see that comparison.

Good. So Anna, I will turn it over to you for number 12 here about the sample sizes for genomic analysis. I think we just wanted to talk about what the limiting factors were for increasing those sample sizes and sort of what the thinking is behind that, whether there are logistical constraints or some power analysis that suggests that those are sufficient.

Anna, do you want to elaborate? Well, yeah, generally, when looking at that one graph where the genomic samples were presented in table, I just felt that they are really low. So, some subareas had only one or two samples. So, we were wondering that there must be some reason why. That few samples are collected, but would it be possible to kind of increase that?

Because it would like improve the accuracy of what— whatever is being found in the genome samples.

Do you have thoughts about what the limiting factors are there? Yeah, thanks for your comment, Anna. Yeah, the sample sizes were actually selected based on previous studies published in the literature, one of which evaluated the minimal sample size for the development of an accurate and precise epigenetic clock. Those, those numbers were in the order of, I if I recall correctly, between 100 and 150 samples. We're exceeding this in this case, 250 samples.

To my knowledge, this is the largest sample size for the construction of a clock using 250 samples of well-aged individuals. So we did not perform a specific analysis to determine this, that this is, that this was the appropriate sample size. But, uh, based on what I just said, the exceeding the number by a factor of 2, uh, of, uh, several published studies, uh, and the fact that we have, um, a fair coverage of the areas, even though it is true that there are some regulatory areas where the number of samples are very small. And then for D, we have 3 samples from the fist. But these are— I would like to reiterate that these are samples that have been double-aged, so the chronological age is fairly accurate, is as accurate as we can get.

So we could potentially improve, but I think it would be good just to see the results of the development of the clock first and see what level of accuracy and precision we get. And obviously, if needed, we could expand that with additional samples. We do have additional samples that are double age, and those are the ones that were selected for the out-of-year, for the future out-of-year validation. So we have another set of close to 50 samples that have been already sequenced that could be included in the training model if needed. But I would like to propose just seeing what we obtain with this set of samples initially and then make a decision as to whether to include in the training model additional samples or not.

I'm not sure, Anna, if this sort of satisfies your question.

Um, yeah, so in future you will have more samples, right?

Yeah, we do, we do have additional samples. Yeah. Okay. Yeah. In, in, in slide 4, you, you can see that, you know, the samples that we were drawing, uh, we had several thousand samples that were from the FIS between those 4 years, 2021 to 2024.

Uh, but we selected an equal number of males and females for each, uh, so we do have a larger set of samples that we could sequence. But that would imply, obviously, if you wanted to draw from those pool of samples, that would imply, you know, redoing, doing additional sequencing with additional costs, obviously. Yeah. We were thinking that maybe putting this note to the report might kind of something to promote that there would be more money for sequencing and sample collection. Do you think that would make any difference?

So, I guess, two comments.

Our limited experience with this is that That's the big limiting factor is the fact that you have, say, 150 samples, but then you have thousands of CPG sites. So I'm maybe not as concerned about the sample size, even if you have to have it to do the— develop a clock for males and females. I think the bigger challenge is like reducing the space like the number of CPG sites that are informative to something that then allows you to fit a model that's identifiable. Yeah. In that regard, we haven't completed the entire analysis, but just looking at the first— one of the first metrics, and that is the number of CPGs that are common for all samples without excluding any samples.

Without missing data, it's fairly high. It's in the order of 20,000. There's a paper that was just published this year on Atlantic cod, the development of Atlantic cod. And in fact, when they used all samples and they searched for CPGs that were common for all individual samples, and they had probably about half of the number of original samples, they got zero. CPGs.

So they had to go to the next step and actually use missing data. So in fact, they ended up selecting initially for the, for the initial matrix CPGs that were present in 80% of the samples. So in that regard, we're actually, we're way ahead of the, of the, at least the Atlantic Broad. So, you know, we haven't finished the preprocessing of the samples. We're going to lose a number of CPGs before we produce the final metrics.

But I think we're in a good spot. So, and we'll find out very soon. You could just look at the correlation within the CPG sites. Like, that's another useful dimension reduction. Right, exactly.

That's precisely at the step where we are now. Yeah. And it's basically filtering a number of filtering steps before we finally come up with the final number of CPGs that are going to go into the Yeah, I mean, going back to Anna's concern about limited sampling of some regulatory areas, I think maybe wait and see how the cross-validation looks for out-of-area prediction. You don't have any— you're not including areas of covariance, right? Just assuming that the age clock is representative across the— all the different regions.

That's the, that's the initial assumption. So if it is, I guess, but the out-of-area validation will prove whether that's—. So if it does turn out that there are some area differences, that, that would suggest that pumping up the sample size for particular areas. Yeah, but maybe we can wait, wait for their— until the analysis is done before Yeah, and anything else on this topic before we—. No, I think we can move on.

Great, let's jump to 14 then, um, because this is related to the, uh, to the clocks. Um, do you want to jump in here on this one? Well, I, I think it's just that was based on our experiences that it was, uh, We found the performance of the clock better. We did it separately, which, you know, like, there's this— I guess some people like to have— you could have a factor for sex in the model. But we did find that the CpG sites that were informative varied by sex.

So we tend— we prefer to have two separate epigenetic age clocks. I think it kind of fundamentally make sense here because the growth is so different. Um, but that will be something you can quite easily test. Yeah, yeah, that's, that's—. It's worth testing, not just throwing in the factor but actually fitting the cycle.

Yeah, that's a good point. Yeah, yeah, we're planning on doing that, looking at this index of sets.

So fitting separate clocks does cut the sample size in half for each one. See how that works.

Okay, great. I think we're okay there. The 13, we talked a bit about well depredation, and we'd had a little discussion about that in the very beginning of the meeting. And I know that, Ian, you talked about the— some of the the downsides of trying to look at well degradation trends with the existing data.

Over—. Over what—. Why we can't use existing data sources to look at trends in well degradation.

Yeah. Well, so we have 3 potential data sources. We have the survey information, which we know differs a little bit from the fishery in terms of our fishing behavior. So it's not— it's the analysis we did a few years ago, we made the assumption that the— when the whales were present, the fraction of the fish that they ate might be something that we could export and apply to the fishery. And that's the basis for the expansion we did on the fishery.

But that the frequency of whale depredation was probably potentially quite different to be experienced in the fishery just because the way we do the sets, the fact that we're actively trying to catch fish. In the, the other two datasets we have are the observer data, where the observers note when winds are present, and the logbook data, where since I think around 2017, some vessels have been reporting sometimes when whales are present. It's a challenging thing in the logs because they, probably rightly so, have some concern that extensive documentation of their interactions with marine mammals may not be a good thing for them at all. So the incentive to report— some people certainly report all of it and some certainly report none of it. And we don't really know what we're getting in aggregate.

And there's not really any good— we did look when a couple years ago, and we did make time series figures of the percentage of sets affected by sperm whales and orca whales. And we delineated that by halibut target versus mixed target. We actually found a higher rate of encounters when they were targeting halibut and sablefish. At the same time, which is surprising because it's the same fish. So we did actually kind of look at the frequency, but we didn't feel like any of the frequencies that we had access to— curiously, the observers were reporting fewer whale interactions, even though we thought it was going to be the other way around.

And we had lots of speculation about why that might be. It's not really the observer's job to do that. It's their task. Monitoring the cash flow. So that's—.

So the upshot was at the end of the day, we really didn't feel like we had a reliable metric. And we did, although we didn't— although none of them were sort of unambiguously the best estimate, we did go ahead and carry the calculation. See, okay, if this is the rate, what would it turn into? And it turned into something quite low. And in fact, we showed it number of people at various meetings, anecdotally, their response was, well, that's way too low.

It's 30% of our catch. But there's nothing in the data that happens. I mean, our peak counter rate, or our peak ineffective station rate from whales is just over 3% in survey. And it was, it bounced around quite a bit, but it was nowhere near going to add up to 30% of the catch. So I think we kind of laid out that avenue as far as we could get.

And the conclusion was at the time— and we can go back and I could refresh which documents that we put this into— but I think our collective decision at the time was there's really not much else we can do because The data that we have suggests it's a pretty small problem. It's tired, didn't really see an obvious way to proceed. For the whale depredation, when there is a whale depredation event, are they taking the whole fish or do they leave like the head behind still on the hook? So it depends. So for sperm whales, generally we get back bent, broken, or missing hooks, and there's no evidence of fish damage.

When they eat them, they're gone. And for killer whales, we generally get lips back on the hooks. So they'll pull the fish off and just the lips will stay on the circle hooks.

So are there recordings of those observations in any of the data sets? Yeah, we have, we have Lips Only back. That's what we use to determine when killer whales are actively feeding in the survey, unless it's just observed. So sometimes they're right next to the boat, you just see them eating fish. So that's one of the criteria that we use to exclude survey stations.

In the fishery, we've actually asked, we asked them to quantify damage to fish or gear in the logs, but The way in which they do that, it's difficult to delineate exactly what the damage was and where it came from because fish get damaged in normal fishing operations, gear gets damaged in normal fishing operations. And so it's not— they're not able to unambiguously identify damaged fish due to whales versus damaged fish due to other things.

Sharks just getting wound up in the gear and cut in half or whatever. Could the observer mandate be expanded to include whales, or is that a really difficult thing to do? That's outside our domain. And a lot of these boats, I mean, if they're fishing at night, asking an observer to leave a closed bait shack you know, get up on the roof to look for whales might be untenable. We don't, we don't really have the ability to task the observers.

We have done a number of special projects with the Observer Program. We can ask for things, can't really mandate anything for the observers. They're not our observers.

I thought you made a strong argument for not including whale depredation in the stock assessment model. And I was just thinking about, just in terms of this has been a concern of the community, if there's something sort of relatively easy that you could do to come up with a metric. And so the— I mean, and the other thing that came apparent to me yesterday was it sounded like the, these, the whale predation was fairly spatially structured. So just, it did cross my mind, maybe using one of those datasets you could, it doesn't seem to me that the survey would be representative of the predation, just in how you described sort of avoiding whales. But maybe the fisheries itself, you could do just a species distribution model where you have presence or absence of whales, or you could even use a multinomial response where you have like number of sets that are— have been in some way, shape, or form damaged by whale.

And that might just sort of appease the community in the sense it would at least show where historically the whales have been. Maybe that's useful information, I don't know, but it would provide an analysis. I find species distribution models quite useful for that sort of preliminary look at the sort of spatial impact, notwithstanding the fact it sounds like it's probably not an impact that people need to necessarily be worried about managing the stock. Perspective, but just as a way of sort of easily doing something.

That might be what we did. We looked at like, we used a multinomial model for it, we wanted to count just a number of sets. So we had like a, the fish was, the bait was gone, the bait was not gone, damaged hook, that sort of thing. And it was crates at the time, not sales, but since— yeah, so feel free to ignore it. It was just something I thought might—.

That was on commercial fishing logs or on—. That was survey. Survey is a bit more representative, I think. Yeah, and you know, like you could fit one of those in an afternoon. Side of the slide.

So hopefully that was useful.

Are there records of how often—. We have that information in the field notes. I think we have a field that we could quickly query, but it's there that they, they know if if they move, they know that.

It's—. Yeah, that's recoverable from the data that we have. It's honestly not a huge issue for our survey because of the way we fish. It's much different experience than we get from these certain people at certain places in the fishery. Certainly creating heat maps of where the fleet is encountering whales, we've actually just done similar things Please.

It's definitely something we could do. We don't have hook-by-hook information, fishery, but we have set-by-set information and at least where people are reporting whales. As a presence absence, we have presence data, we just don't have reliable— yeah, what's reducing over years. Yeah, we have seen some changes in the spatial footprint of whale depredation. It used to be that killer whales were pretty much only out west, and we've seen, I think in the last few years, we've seen a bit of an increase in Eastern Southern killer whales.

Everybody thinks about these in the Gulf.

Yeah. But besides the incentives that don't match and people misreporting intentionally, I think in terms of presence versus absence, the absence is much more reliable. For whale versus the pilot sample project we did. We had a lot of issues with plotting the sampling definition just because the way how it was set up, but the absence was compounded a little bit with just general damage. I think for whales, but again, as Ian said, the incentives and misreporting Based on—.

I could imagine kind of a bifurcated response from the fleet that some people would appreciate seeing it and be more likely to report it, and others might be less likely. Being concerned that identifying a hotspot might lead to a closure of that hotspot. So, I know on the West Coast we've had the crab fishery on the West Coast of the U.S. Currently in the middle of huge issues with gear entanglement. And they actually have depth and spatial closures to keep from setting fixed gear in areas where those are— in that case, it's back whales.

I guess the hope would be that the hotspots move around year to year. Yeah. So closures wouldn't be effective, but I guess you'll know that. And we— and to be clear, we don't have an entanglement issue along. There's really only 3 or 4 lights for each set, so it's a much different problem than commercial crab gear, but yet there is still some sensitivity, and there have been in many instances.

It never occurs, but yeah, I think there's some avenue to do some mapping there.

Do you think the reporting issues regarding whale encounters are similar across the US and Canada, or do you think that due to differences in laws and things, there might be differences kind of spatially because of that? Yeah, we do have an issue where the Canadian log program doesn't add fields when we ask them to. And so we, in 2017, we added fields to record marine mammal interaction, and they didn't add them into Canadian logs. And so our port samplers interview the skippers and collect that information and write it onto the logs. But the reporting, it can't happen unless we have a port sampler there.

So we may have a difference, an effective difference in reporting rate between the two countries, which could be an issue actually.

You can test for that too.

Yeah, and I mean, you might do some initial filtering to filter out those, those captains that are never reporting interactions and ones that stopped reporting them.

Kind of like what—. It could be subtle to pull that out because some people really don't encounter them the way they fish. They successfully avoid it, and other people don't even try. They just set the gear, and if whales are there, they just haul right through them and feed the whales. It's very, very diverse set of strategies.

Some people claim that whales is just a way to blame your bad fishing on something else. And other people, it's clearly not that. Other people get massive amounts of whale liberation. We have two sets of data. There's the observations of whales and the damage set.

So these could be also— how jointly these interact. There are also methods for interviewing people about activities that have a social stigma around. People interview techniques for like drugs.

Some of those methods might be useful for doing separate survey to get an understanding of where the hot books might be.

That might be something that could be added to the fisheries sentiment survey. So, we started to So we're interested in just kind of general feedback from the fishery. How do you think the fishery went? Just want that first, you know, sentiment about how the season goes. And what we're particularly interested in is changes over time.

So we have this feedback survey running for the first time this year. Those are novels? Yes. And this is like a very short Online survey, the field staff distributed workbooks. They don't necessarily see— it's not the paper form, so they can see the answers.

And it's a very short survey, it's like pretty much a couple of questions long. And there's also a question about the main problems of degradation, so we could at some point also quantify it. There's a percentage of responses that degradation has identify main problems. So a couple of like 1 to 10 question, 1 to 10 scale questions. They're just what you think what the biggest issue is.

Is there a geographic field? Yeah, so there, there's which regulatory areas you push, which sector you are, uh, which port you push from. There are a couple of these. And we did trial with the conference board. So that's where they're saying if people tend to complain a lot, you know, they come to the meetings like, oh, this is so bad.

But when you look at the results, it's like, oh no, my coaching was great. Like the scores were like 6, I think 7 out of 10. But I think that for this type of, um, scores, more interesting part is comparison year to year. January. Really interested then, like, we're going to run the second round for the conference, 4th of January, and then we'll have the first year this year for the fishing fleet.

That seems like a promising way to get—. If people doesn't come to report it, but it's associated with their name, of course. Yeah, and that was also because commissioners are also asking us to send them from the field, and we started like having some of the notes from the field, asked for by the field staff, but they were also including maybe when the fishermen were asked in person, what would you like us to convey to the commissioner from your perspective? It's just complaints, right? But that's a great excuse to work the concerns, you know, what we are worried about.

But that's why we came up with the sentiment survey to get the fun results, see how things change. And I think the interesting piece would be to compare how the score overall changes with the finance change.

Yeah, what, um, what kind of response rate are you getting? So we just pretty much started, so we wanted to be because we wanted to look something with fishing season. So we want to start at the beginning of the fishing season. So pretty much starts just here. Oh wow, okay.

So, so the— we did the test run for the conference board.

Conference board members filled it. Doesn't think about— there's trials that was maybe It's like 60 members, and I just start like first field survey coming in for this year. But I said the goal was to reflect 2026. This was the end.

I think this seems promising. Learning more about whale depredation. And did you— so it's voluntary, but it's distributed to everyone? So, well, we intend to distribute it to everyone. So it's distributed for the field staff who are present at the major ports, and a lot of vessels move out of those ports.

Of course, we will not reach out to all, but it's the most promising But hopefully it's in a simple enough— that's really like, they give them the flyer with a QR code, so just scan the QR code with your phone and you can fill out survey.

Is there a report associated? No funds for that. So we found that A $2 gift certificate sold at astronomically higher completion rates.

Maybe—. Wait, wait, wait, wait, wait.

Not even half a coffee, maybe a third. This is widely known.

So maybe we could put like something to fill it, like the $2 gift Gosh, Pacific halibut patch, right? Totally, just something like that. Well, we do have a pretty, pretty good response rate for the tagger, so we— our tagger reward is pretty much a hat or $10 gift card. We have a $10 gift card available for the taggers. We have a pretty good—.

Yeah, I mean, like a sticker or something.

A gesture to get people— I mean, if the port samplers are interacting, people are ready. Being able to say like, hey, you know, we're following this survey, I get a better sense of street experience, water, here's a sticker. So it's more of an incentive. So before, do you use it as an It's not an incentive, it creates a sense of obligation. Give me something, okay, I will.

Yeah, that's miserable. Maybe we can think about it for— so, so this year, so kind of trying it out, this was— we thought it would be like really interesting to see like a year-to-year comparison. Maybe a good idea.

Yeah, I would think about setting this up as like a real monitoring tool.

Think about the problems you get from it. Sentiment is certainly one of them. Potentially get a lot more, especially about things that are hard to understand on—. Yeah, so it was set up like a sentiment— there's also like this open question where we wanted to calculate the sentiment, but like It's pretty much like a liquor scale question.

Maybe we'll have some results a bit later. But as I said, just so far, it's just the conference board members, and pretty much it's in progress for the fishing.

Okay, let's keep moving And I don't think we need to talk about 15. That's— is that straightforward?

Uh, 16 is worth some discussion. So we, we talked about this a yesterday. Colin, you had sort of led your section with wanting to understand big old fat fecund females and whether they contribute an outsized proportion to spawning output. How—. I guess we were still looking for the link between the histology metrics and how Hypothesis for that?

Well, the histology data, so every fecundity sample we have is also sent out for histology. So every fish that we have, whether it's for maturity or for fecundity estimation, there is a histology sample. So one thing that people have looked at as far as histologically for fecundity estimation is what's called downregulation. So some fish species, as the oocytes progress towards spawning output, they actually reabsorb some of those oocytes prior to the spawning season. That's been found in people at like the Northeast Fisheries Science Center, some flatfish species in the you know, the, the Atlantic Ocean, um, as well as here on the West Coast.

Um, so that's something that we could use as a, a metric for the histology, is whether or not we see later developed females and whether or not they're showing high rates of atresia later developed. Now, currently, that's something that we haven't Now, it's something that is plausible. And it's something that people have incorporated into their fecundity estimations because they use what's called, like, potential fecundity versus realized fecundity, and they use follicular atresia rates to account for that realized fecundity estimation. So, so that's something that we can definitely use histology for. For in our estimations of fecundity.

I guess we're still trying to tie it back to the, the VOF question. Yes. So, what would you see that would say, like, ah, yes, the hypothesis is confirmed or rejected? Yeah. So, once we build, like, once we build out the autodiametric curve, it allows us to take oocyte diameter and extrapolate that out to numbers of oocytes per gram of tissue.

That's, that's the whole autodiametric curve. That allows— then if you have a fresh ovary weight, you can then extrapolate that information out to what's called potential annual fecundity based on number of oocytes per gram of tissue extrapolated to the whole ovary size. And then once we have that information, information, we can then figure out based on age, weight, length to see whether or not bigger females are exponentially producing more eggs, because right now the assumption in the assessment is there is no, it's basically just a, what, we're just estimated based on like the cubic function of length-weight, right? Basically, as females are getting bigger, it's just basically following— it would need to increase even— that's, I guess, that's something that I think share stakeholders struggle with a little bit of the concept is that they, they think that what Ian is currently assuming in the assessment is basically just like a, what, like a linear—. Yeah, so we assume that Dagg output output in proportion to the body weight, right?

Yeah. Go up and think about it in terms of egg output.

Right now, egg output is an exponential function. Yeah, cubic, basically a cube. But what we want to know is, is, is eggs per gram body weight constant, or is that actually— that, and that comes from the— if that's got a slope, yeah, then you're actually producing more eggs per gram body weight. Yeah. Which get big.

That's, that's the key. We haven't, I don't think we've thought yet about things like egg size, viability, and oil glob, other stuff. That sort of thing, in order for those things to be relevant, you have to know where the bottleneck, survival bottleneck is. But, but it very, going to be very attractive to document whether we're actually whether the egg output per gram body weight changes. I think that's the very first key.

If we see a relationship there, then automatically we know that bigger fish are more valuable to the species.

Yeah, that's super helpful to have that. This was kind of the missing link here between the BOP hypothesis and autodiametric methods. So I think in And in the future, we're going to be cool to see how, like, what are the next steps for building that out and all the way to the statistical test of linearity and the relationship. Yeah, that's definitely something that we're going to be internally working on over the next year as we actually start processing the samples and actually collecting data. Across to see where this is headed.

Yeah, yeah, yeah. It's one of the biggest, it might be the biggest question out there that a lot of people want to know about, stakeholders and everything. Stakeholders are—. Oh, absolutely. Yeah, when we have the research advisory board meeting in November, which is mostly like commercial fishermen and recreational fishermen, the last couple of years I've presented us collecting samples, and they all want to know, are bigger females actually producing more eggs?

They all want to know. So it's probably the both hypothesis. Maybe it would be better articulated if we indicate that that is really testing whether city hall follows a hyperallometric relationship. Hyperallometric.

And weight, basically. Or whether that is a directly, strictly proportional relationship between egg output and body weight, which is really what we're assuming right now.

So maybe we could deal with that.

Yeah, I wasn't thinking about it when— for the presentation yesterday, but have you seen much variability in egg size among individuals?

No, no. Yeah, egg size is pretty, it's pretty well determined. And, and we, we, um, in the past, um, several SRB meetings ago, we actually looked at potential egg size differences based on actual measured diameter, and there was, there was really no, no, no, no variation. Perfect. That's good.

That's, that's pretty well. Yeah, I mean, one of the, one of the, that is a side benefit of the autodiometric method because now every fecundity sample we have for every fish is going to have like 200 to 300 oocyte diameter measurements. So there's going to be a huge data set as far as egg size. And so that is definitely something that we can pay attention to as time, through time, to see whether or not there is any variation. But as of right now, we haven't seen that.

No, and there's the, we also looked at whether there's any difference in egg size in relation to age. There is not. So, you know, a 10-year-old female that is actually spawning, the eggs are about the same size as a 20-year-old.

Cool. As far as you know, they spawn until they die. They spawn, sorry? They continue growing and spawning. They're like—.

Well, we really don't know after which age, you know, the senescence will impact the reproductive output. So what would be the oldest female that we've actually assess maturity for?

Um, I think it's 41. The, the oldest, uh, aged female, uh, is, is a 55-year-old female that we have aged. That's the oldest one. So that's, that's actually pretty late. So a successful—.

A female that will most likely according to, you know, our data, a female that is mature and advanced enough, that female will— that doesn't show any signs of atresia, that female will most likely spawn between January and February.

And, you know, those ages are actually pretty old, so. Yeah, we've— we have yet to see senescence histologically in Pacific It's been shown in lots of other fish species that have done histological work, but we have yet to see it.

This makes me realize I don't fully understand how spawning biomass is defined. Is that the portion mature? It's based on the maturity Do you have that plug alone right now? Well, maturity, this, the spawning biomass is the maturity curve multiplied by the weighted age.

So it's effectively just the same as the biomass. Sure. And then if egg output is linear related to biomass, then that should be fine. Then there's no change. Well, there's not, then we just multiply.

Fly by the slope of that relationship. Exactly. But so, right now where we're talking about truncated weight distribution, put up a lot of larger individuals. We've seen, we've seen changes in mean weight through time. And so if there is a nonlinear relationship there, that, that's what environmentalists curve would not be proportional to reproductive number.

If there's a fecundity— really, a nonlinear fecundity relationship with body mass, then yes, we would get a different type of series. And that's why I've highlighted this, put people on notice. This is important. This occurs in the numerator and the denominator of And we've done some sensitivity to changes in the scaling of that just to show that if there is a relationship, it will change the trend. But we don't know how extreme it is.

So, sure, there's not really any way to— define the plausibility of those hypotheses until we get some data.

Uh, I'm trying to think about other indications. So the, the gonad mass, the, the navel somatic index, is that Seems to me, seems to be, we last, back in June, I presented at the SRB meeting a few figures of some preliminary GSI estimates with like weight and age for the fecundity samples that we have of those fish, and it didn't seem to show, because you would expect that if there was a hyperallometry going on, those gonads would be really big to produce more egg output. And we're— we've caught some over 200-centimeter Pacific halibut that have really big ovaries, but I think those ovaries would need to be even bigger. Their GSI though is similar to—. It's very— yeah, it's not like exponentially bigger.

Yeah, so then the only way that it would get hyperallometry as the egg size is smaller. That's right. Yeah, exactly. Exactly. Exactly.

So that's something that we—. If the GS size really dependent on when— yeah, much better to go through the autodiagnostic. But I did that. But I did that. I think I showed that even in June.

I think I had like every stage of vitillogenesis, like VTG1, VTG2, and VTG3, and it's still didn't show anything really. Yeah, yeah, well, I mean, certainly would be a nice answer the way that— relationship.

Yeah, yeah, great. Let's, um, on here, uh, I think the stock is Olaf, before we move on from that one, the other question I had with this part was, do you have any concerns about some of the individuals that you're indicating as mature histologically might resorb their eggs before spawning, such that you might end up— I'm just thinking about the difference in the younger ages between the maturity curves you were presenting yesterday?

I don't—. I mean, from everything that we've seen with Pacific Halibut, it wouldn't suggest that there's going to be an abrupt abort. That you— we just, we just haven't seen the levels of atresia that might be indicative of that.

And, and if that were the case, we're, we're dealing with very low levels of numbers that I don't think would have an impact on the maturity curve in itself.

That there's lots of different ways that people are calculating. We, I mentioned a little bit yesterday, there's, there's not a, there's no, there's no set criteria for how people are distinguishing a skip spawner or an abortive. It's, there's lots of different ways that people do it. Um, and in particular for us, because we have such a long time series with the visual data estimate, we're, we're using the histology to calibrate that visual estimate through time. And so that gives us, I think, a lot, um, sturdier, you know, better estimate because we have such a long time series.

Whereas some other assessments are using just a few years of data, we have such a tremendous data set here that I think it's given us a more robust maturity estimate because of that. And I think that's why we've done it that way.

Yeah, great. I mean, if I may just add, Yeah, I think from what Colin said and now our experience, and this particular species is actually fairly easy to identify whether that would be happening or not. Our conclusion is that it's really not happening. Pacific halibut female development is group synchronous, which means that there's only basically two cohorts of eggs present in a histological sample, and those are the ones, one One is those eggs that are actually advancing, and then there's a pool of very early oocytes that are going to be recruited for the next spawning event. So there's only really two populations.

It's not an asynchronous oocyte development where you see, you know, all stages at the same time in a histological slide. So if, from what Colin was saying, if there's really any sign of atresia, and we've seen it very, very occasionally, You would definitely be able to detect it, and we don't in sufficient numbers to think that really escape spawning, at least in our samples, is an issue.

Plus, like I think we mentioned yesterday, majority of the higher rates of atresia is happening at the transition zone. Between our classification of immature and mature females. So once they have progressed into vitellogenesis, we just don't see, we just don't see those rates of atresia. It pretty much is like a, it's like, yep, I'm going, this is happening. Yeah.

Cool. Because that's, that's the one thing that I— having that maturity curve shift younger might actually make a really big difference in the reference points and ultimate management advice. I think bigger than many of the other things that we've talked about. And so just reconsidering that, it looks like all of that is warranted to have higher maturity at younger ages. Is, I think, a really important part.

Great, thanks, Mike. Ready to move on to stock assessment?

So I think the first couple are pretty straightforward.

Well, actually, the first one's straightforward, and it seems like that was your plan already, to develop the— further explore the Seattle model and then look into Lance's work.

19, Anna, I think we had brought up a discussion of how the assessment uncertainty is expanding at low stocks. Size that could create additional risk. Do you want to elaborate a little bit, Hanna?

Hi. Yeah. So, the thought here was that it would be— like, we felt that it would be good to pay attention, similarly as on the average predictions, but also the uncertainty, and especially thinking about that uncertainty is not like same throughout the time, but that can show trends. Like usually a population gets smaller, there will be more fluctuations and more uncertainty about the average, and maybe this could be kind of better taken into account in the stock assessment, but also in biology research. Like, there were many figures, but that had only, like, parameter estimates, point estimates, but not the confidence intervals or something like that around it.

This might be a little challenging in the stock assessment, because the stock assessment, the ensemble is predicting since 1990. And it's been basically declining since 1992, especially stock status. But we're confounded with better data when stock status was small and more uncertainty when stock status was high because we didn't have the same quality and quantity of data that we have now. So I, I think this is a really interesting request, but I don't know that the stock assessment would necessarily have the ability to investigate this. The other thing is that we've got terminal year estimates, which are always more uncertain at the lowest stock size of seedlings.

But at least some sort of sensitivity analysis could probably be done, that what if we have some parameter a little bit off and how much does that affect affects the outcome.

Sort of a sensitivity, rather, a sensitivity to, say, stock status, or—. Yeah, sensitivity of the, like, the predictions or the whole, like, model outputs. And even if it wasn't like quantitative, or if it's difficult to address quantitatively, but at least like keep it in the discussion, keep it in mind.

That, for example, if we look at the different, like, probability of being at the depleted side, or where does that threshold go, we also think about like what is the probability of falling to the depleted side, even even though the average is about the depleted side.

Yeah, the other thing that I think intersects with this a little bit is that you also have fishing mortality rates declining or lower now than they were historically. And I would just expect with that that you'll get higher uncertainty in your estimates because fishing is a explaining less of the population dynamics.

Yeah, but especially the— like, we all know natural mortality is really difficult to estimate, and it has a lot of uncertainties in it. So thinking about how the system would look, how much it would look different if we thought about the uncertainty related to the natural mortality estimates.

I'm sorry, I don't think I'm following what the question is here. What are we trying to answer with this analysis? Is this to provide the decision makers with a better sense of the current uncertainty relative to historical uncertainty, or is this a— attempt to better understand the performance of the models and whether we're getting more sensitivity at certain stock sizes than others. Yeah, like when setting up quotas and that kind of, like, how much uncertainty in the stock assessment predictions can affect whether a set quota is sustainable or not. Like, and of course, that then comes with the question of what is the acceptable level of uncertainty, what is unacceptable.

Perhaps we could table this until we have our free-form discussion this afternoon. Sure. Yeah, I think we're, we're not fully, fully understanding how to put this into practice. So maybe having a bit more discussion about it later would help.

Great. I think we'll do that for those, those last 2 as well that are listed there, 21 and 22 discussion.

Yeah, we had some discussion last night about, so what do you do with these state space models? You alluded to this, you know, are they now multiplying out the number of models included in the ensemble because you've have your whole stock synthesis framework plus the new state space. The new state space model replaces stock synthesis entirely, and so you just have the same 4 scenarios evaluated in the ensemble that are coming from the state space model. There's lots of options there.

To look into those, those different options, or possibly going away from the ensemble.

But we'll, uh, we'll table that at this. Sure.

I think that is— oh, oh, MSC.

Can we take a 15-minute break and come back and see? So getting back together at 10:20. All right, back in 15 minutes. Back in here and talk a bit about management strategy evaluation. We'll start with— 7 here.

So I'm going to just pop in the definition of depleted from the current harvest strategy policy here.

The point was made yesterday by Joanna that the definition refers to recovery and recovery itself is not defined. So this might warrant some discussion here and might be helpful in getting us over the hump in— with a better definition or actually defining the difference for us.

Yeah, I think the— Yeah, I seem to remember this has come up in the past. And there's two things that have come up in the past. What does recovery mean? And if you are in a depleted state, what is recovered? What is the definition of recovered?

So right now for an overfished state, we have the definition 70%, 70% probability or above the overfished state, I think that. So we have a definition. If you're overfished, how do you come out of overfishing? But that might be a different question than the concept of recovery, but I think that does need to be added. Recovery though, I admit I haven't put a lot of— I guess I've been naively assuming that recovery is the ability to go back to an unfished state, unfished population size, get fishing stopped.

So that's the simple definition that I've been thinking of recovery. That seems simple. It's something that we would apparently never observe. Yes, right. Yeah.

So I'm not sure, I haven't thought beyond that. What are the consequences of going to a depleted? Is that where we're going here? Is like, is there a consequence that The population may increase, but not actually recover. I guess, but do you want to talk about it being recovered?

Like, is it an increase in population size for 3 sequential years? Like, I feel like it needs something that would signal that you were in sort of a like a good performing. Yeah. And it probably— or like, I think sometimes it's done with some measurable increase in the CSB, or right, some delta. But that might be ambitious here given where it looks like the stock has sort of bowed a bit.

Um, so maybe it's more important that it's not decreasing. So an increase or stable state for 3 years or something could be— it could be as easy as that. It need not be, but I was just thinking it's useful to have something that you could call covered. Yeah, and I see two cases. One is if we were to do like— well, when we were talking about how do we define depleted, there were three different methods that we put up on the screen there.

And let's just, for sake of argument, talk about the third method of simulation. The idea was that the population could recover if after 2 or 3 generations it approached the unfished condition after stopping fishing, and that's simulation thought. But the other 2 approaches are what does signal recovered when you're still fishing stock down to these low levels that we're currently at, and I think that's where coming up with a definition will be useful. Can you remind us What those two were? The first one was the minimum spawning biomass that we've observed in the time series, which is 2024.

So we don't know if we've come out of that, recovered from that, because we haven't had enough time. Yeah, and I think that's your point, like—. Yeah, well, does recovered mean it's higher? Yeah, yeah. And what it was last—.

Or does it mean that it can maintain going higher after a number years. So, and then the second method was spawning biomass that produced a piece of timber, which was 2012, and that was a bit higher, but that depleted what does depleted, how does depleted relate to that level. But I think this is really important because we need to define in the harvest strategy policy what recovered does if we are in a depleted state. And it—. I don't think it's as simple as you're above the depleted state, but, you know, we want it to be more sustained.

Yeah. And persistent, or yeah. And so I think that would be a really useful thing to think about. Yeah. And I think about it in comparison to the U.S.

Federal system where there's overfished and then before it comes into place. And it's not just getting back up above overfished, right, that triggers the end of the recovery plan. It's getting back up to the target. Right. Yeah, right, with a high probability above that target.

So we've defined that in terms of overfished defined as dynamic relative spawning biomass, a relative stock status, that there's a greater than 70% probability that we're above overfished. That's what the harvest strategy always says now. So it's still using that overfished threshold, but you're saying you've gone above it with a higher than 50/50. And so we can think about covered in a way that you mentioned, or in this other way of like you're above the depleted level with a high probability, or you're above some level above depleted, high probability. And so I'm thinking off the cuff here, so I don't know if I want to continue thinking.

Yeah, I don't have a good idea for you on this, but I would urge some caution with reinventing something that's equivalent to the US rebuilding standards just because of how problematic they've been.

Okay. And so I think in a general hand-wavy sense, we know what we mean with this, but trying to actually operationalize it is the real problem, the The problem we're trying to avoid is that as we push abundance or biomass down lower and lower, the probability of the stock not being able to increase again gets higher and higher. And so we don't want to get to unacceptably low levels of that. That being said, there's an enormous amount of uncertainty around that because it's something that you only get to see after you've crossed that sort of point of no return, and we don't want to do that. So, uh, it ends up being really tricky to basically phrase it as the ability to recover.

And but like I said, I'm not— I don't know that I have a good alternative at this point, but I'll just keep thinking on that. But It might be something that maybe, um, uh, phrasing it more as a level that has an unacceptably high level of risk of stock collapse or something like that, because I think collapse would be easier to define than recovery. True.

So of the alternatives that you mentioned, Alan, the simulation-based one, I think, is the one that's least attractive to me because the— at least in the way it's used in the U.S. system, Being overfished is something that shouldn't happen, the control rule. And so it's a way of saying, wait a second, our predictions are not going right, our understanding is wrong. Let's shift into this new paradigm of recovery rather than rebuilding rather than optimal yield. And if you're sort of still defining that in terms of your operating simulation, you never have the potential to bust out of that paradigm because it's not working. It's sort of related to exceptional circumstances, right, where, where there's a red flag here.

The operating says we shouldn't get into this situation based on the strategy policy, and yet here we are. So, defining it with simulation, I think you're You're always going to have a level of optimism. It's relying on that simulation model that, that's part of what got you into the trouble in the first place. Yeah. So you don't want to use the same model to basically understand those two ideas, which they're now interrelated.

Yeah. And the model doesn't have all the effects in it, so it's always going to say, Yeah. What about talking a bit more about this, um, this option, uh, lowest biomass that has produced, um, high returns? Yeah, and I mean, I like the concept of that, but, but I think I think of it, and I'll review the ICES system and how they do all of that. But with those concepts, I see that more as that's an okay biomass to be at, right?

Because it's produced a good recruitment. It's not a biomass that has a low chance of recovery. So how do you then define a depleted level or level of concern based on that biomass? Is it half of that, or, you know, is it some percentage of that? That's the challenging— that, you know, that's where it becomes ad hoc again, you know.

So that was the 250 even?

Yeah. So, you know, JC, it would be take some percentage of that and all that. I can't remember exactly how it works. It's actually the— yeah, the level itself. So it's like a hockey stick.

Up, right, to stock recruitment. It's the biomass at which start to go into the Pacific. Yeah, and I remember because I, I was a reviewer for herring stocks when we got to that point in the review, a lot of discussion. It was very challenging. I don't even think we solved it there, and it got solved later, but I remember it's like, oh boy, I see the challenges with this, especially with noisy stock.

Yeah, yeah, yeah. I think if we want to pursue this one, it would be helpful to look at stock recruit.

Is that something that you could—. A stock recruit, um, Well, say the estimated stock recruit relationship.

Yeah. Yeah, yeah, for sure. And we did an analysis of compensation a few years ago as well. It was very uncertain, but—.

That might be useful to look at this afternoon. Okay, see if I can dig that out. It's— yeah, I might even be able to just bring up the analysis I did for deepensation. They might have some plots in there that are useful. Anything more to say, Anna or Mike, about this one at this point?

I don't think I have anything else. Okay, we can also revisit this, uh, afternoon.

Yeah, I think, I think we're not quite at the point where we have a specific request or recommendation here, but I get the sense that we're heading in that, that direction. So it'll be useful to talk more this afternoon about what's, what's feasible here. And maybe some kind of investigation of what other management bodies do, really understanding helium and the IC system and some of the other options for the reference point. Yeah, I think that's useful and that's something we've done for other concepts of a strategy policy. So it's good for me to learn more about all those.

Yes.

29 Here, we noted the stability of the projections. Projections depend on the assumed levels of assessment information, information error. Can you talk a little bit about what those assumptions are? Yeah, so we've talked about, and there's an outstanding request from the SRB and the commission actually in the program of work to improve the assessment error module in the framework. And basically right now, because it was really challenging to mimic the ensemble stock assessment in a simulation, and we did a lot of work in 2019 on trying to put together an estimation model with stocks in this this, and it just— we had something running at first, but it just never really gelled, had to get progress made.

So we ended up looking at uncertainty on stock status and uncertainty in the mortality limit coming from stock assessment in a bivariate correlated multinormal distribution between those two, because that's really what we are trying to get out of the TCY limit and then stock status in case we need to invoke the control rule.

And, and so it's just basically a multinormal with assigned CVs from an analysis I did a long time ago. I think they're like 15% CVs on each of those axes and then a correlation between them. We did—. It's been a while since we've done that work, but that's what's been implemented. So it's a very simple way to just inject uncertainty, assessment uncertainty into the simulations.

And I completely agree. I think I would like to have an estimation model in there, and that's something that I've been working in spare time trying to put together, especially with TMB models. Model, trying to get something more simple. So that's assessment error. With implementation error, which we call decision-making variability— I don't like to use the error when talking about it.

Error is statistical sense. Yeah, yeah, yeah, exactly. Managers don't see it like that. And the decision-making variability, I took a look at decisions the commission had been making and, um, the amount— how they were departing in the past from like 13 or 15 years of data, how they departed from the assessment advice. And, and I thought, oh, when it's higher, you know, maybe they're following it.

When it's lower, they're not. But it was just all over the place. So I'm not— yeah, this is bringing back flashbacks of me doing this analysis of trying to figure out a relationship. And I think in the end, Ian and I sat there discussing it, just said, we don't know what's going to happen in the future. So let's just put a normal distribution around the assessment advice.

And they could go lower and they could go higher, but there's some level of certainty. And I believe I parameterized it as a CV. So it's a smaller absolute range at smaller values, but then it can grow at higher values or something like that. I can't remember the specifics of it right now. It's something you just code in, and then it's working there.

But it's, again, it's a pretty simplistic approach. Just it doesn't have— I think it might have some bias at the lower end, but a very little bit. But it's just basically adding. Adding error onto the assessment advice, which is coming from the uncertainty in the assessment, and then additional uncertainty on top of that, the commission may choose to— or the decision may be lesser or greater than the advice. And so what I also run are then cases.

I like to run cases without any of those errors. So there's 3 sources of errors in this. There's the assessment uncertainty, the observation uncertainty from the FIS, because that can, that can come in and— does that come in nowadays? It's not coming in in distribution anymore. Good, because we're randomly sampling from past distribution percentages.

Um, where's it coming? I can't remember, but there's, there's that sort of observation uncertainty from the FIS if it's used elsewhere. But that will come in if we have an empirical rule set instead of the assessment. Um, and then there's, um, the decision variability. So there's 3 sources of variability and, um, management variability.

And then I do runs with those turned off. So you can see the difference. And what you'll see is the AAB on the TCUIs are like 30-some percent with all this variability. And then you take away all that variability and the AABs, I think, are more like 5 to 10%. So it's a major source of interannual variability.

I was trying to get at— trying to understand— people can see the effect of those, understanding of it, but I think both of them can't be proved.

Is there, uh, I know this is already on your, on your to-do list, is there something specific there that's— I think also, how does this interact with the shift to a state space? Right, right. And, and the shift to those models may— I mean, they, they can run pretty fast. It might be, you know, because one thing I was thinking and I started on was writing my own estimation, TMB, and seeing how I can really optimize it for speed in a simulation context. But seeing some of these new models, how quickly they can run, like, Stock synthesis can take a lot of time because it's trying to do so much.

These are pretty streamlined. I might be able to develop an estimation model. The difficulty is mimicking the ensemble. So what we looked at 2019 was, could we say, use 1 or 2 models that would mimic the uncertainty in the assessment because we're not trying to do the tactical assessment in the simulations. But it would be good enough, right?

And we found that we actually— we ended up using two models and they would run parallel in the simulation. It would spawn both of them, do this quick assessment, truncate data, do all sorts of things to speed it up and We got pretty good behavior, but there were some— I think there were some concerning patterns where it would go off the rails. And, you know, what do you do about non-convergence? That happened a bit, things like that. So it was definitely could use more time or more investigation.

And as you said, maybe moving to these new models will help. But I think recalling the SRB, request from last time to, to do the estimation model with Swole, because I thought that request was really good. I just haven't had the time to meet that request yet. So I think we're calling that, um, and then, um, I mean, if the SRB would like, um, you know, maybe a presentation, the assumptions on these three sources of variability, management variability, and I guess you call it management procedure variability. They're all part of the management procedure presentation on these 3 sources for further discussion in the future.

Might be one way to start this, and we'd all be on the same page again. Obviously, I code this stuff years ago, and then you can't remember exactly how you implemented it.

Am I wrong that, um, remembering that that Sean Cox was really, um, not to spend a lot of time on the assessment. I think that was certainly early on because we did spend some time on it and it was like, yeah, it's going to be a challenge and it looks like this assumption of this bivariate uncertainty is performing well enough to inject some uncertainty from it. But then I think the concerns that we talked about— that might have been 2 SRBs ago— was how, like, it's not capturing the biases now. We had— when we did that, we had a more consistent survey and all these other things. Now, you know, there's these differences in sample sizes being collected that could have an effect on the assessment model, the estimation model that would be good to include in the simulations.

And we're more interested in investigating things like what happens when we do reduce the footprint of the survey. Um, you know, and we did some investigations in the past, and so I had to make a lot of assumptions how that would affect the assessment error. And I just—. Yeah, but, but this would be a more direct tie.

Yeah, I— okay, that's helpful to check my memory of these questions. Yep, and we had talked quite a bit at the time about the value of information, the analysis of value of information with this. That would require that the MSC have the assessment. Yes, yeah, I think that's a big part of it as well.

How long does the MSC take to run now? Um, 500 simulations, if I, if I run it on one processor home computer, um, to do 500 simulations, it's a few days. It's probably 3 to 4 days. Really? I— then there's obviously parts of it that I still don't understand.

So, okay. What was your expectation? Shorter than that? I was expecting maybe a couple hours. Yeah, I wish.

Um, yeah, because without the, without the assessment in it, I would, I would expect it to be really fast. Yeah, it's even though it's spatial, it's still not that It doesn't seem to be that complicated of a model to me. It's the application of the management procedure that takes a lot of time. In fact, yeah, I'm remembering. So the model itself, if I had constant fixed catch in there, it would be less than a minute to run, you know, 100-year simulation into the future.

And that's what I do with no fishing. But that's one case. So it takes a few hours to do it with no fishing. What takes the time is the calculation of the SPR. And I'm not doing that through any shortcuts.

I'm actually running the model for every loop of the— to calculate the SPR. It's doing an iterative approach to get right at it because I found that I couldn't figure out a method to really do it accurately on an analytical sense. So it puts in a TCY, and it's— and that was actually an improvement I made this year, the starting point for this. I forget what you call that, type iteration, where you're just going left or right until you tone in on it. And yeah, so that is really the time-consuming part there.

But as I said earlier, the time-consuming part that it's calling an R script every year. Well, actually, every single one of those iterations to do this. And the calling of an R script from an executable took a lot longer than expected. Yeah, yeah, yeah, yeah. I can see that now.

I can see that. The way I've done this in the past is I make my own really tight function where I just code my own Newton algorithm for it.

I, it usually only needs like 7 steps. It's usually at the answer in like 3 and I just push it to 7 just to really make sure. But that one ends up being almost the same as just applying a constant f. OK, it adds almost no overhead to it at all, and so there might be ways to significantly speed up that part.

Yeah, yeah, yeah. It's something that I certainly would like to think about. And yeah, this does beg the question of, in the back of my mind, I mean, we wrote this 2018, not '19, so it is aging. And so, you know, now with these platforms like Spork that does really good spatial, has integrated MSE, I'm almost thinking of, would it be worth doing what we're doing with the assessment and investigating platforms to help us do the MSE? Or at least, how can these platforms be used to benefit the MSE?

And should we rethink our paradigm on this code that we have? Yeah, yeah. Even thinking of a version of Spork that might be the operating model would be interesting because you could feed it the similar to what you're currently doing. You could feed it the biomass estimates from the stock and give them catchabilities of 1 and then fit to it, or at least I assume you could.

Yeah. Yeah. And so I think as part of the learning process, learning Spork as part of the assessment process. I actually approached Ian about a month ago and said I'd like to learn it as well as a potential for an operating. So I think we're going to try to do that together.

Yeah, I think that's a good idea. Does Seattle not have MSC? It, it, it does. It has some internal MSC stuff as well. It doesn't have the spatial, right?

Yeah, we talked a bit yesterday about the PDO projections and alternative ways of— so I think from my understanding of it, you have an increasing probability of regime shift as you spend more time in that one machine.

And I think Mike brought up the idea of doing, you know, full-time series analysis so that you are sure that the behavior was consistent with historical. So you fit the time series model and we'll be objective. Does that seem—. I was thinking the exact same thing last night when before dinner last night, and I was sitting in there writing down some ideas on how to do this. So, yeah, I like this idea.

Great. So this is an easy request.

Um, and then we talked a bit about whether the MSC might be a place to think about whale depredation. We spent a lot of time about well depredation, not wanting to— not wanting to ignore the clear concern from the industry about well depredation, but also recognizing that the empirical approaches are pretty limited for the reasons that you talked about before, asking whether we can at least bound the problem in the MSC to either indicate that this is something we really need to be paying attention to and designing research projects around monitoring it, or convincing ourselves that this is not yet something that— that level. But we don't have any specific ideas. I mean, it would be pretty easy to even just add a lead that it's whale depredation. Turn that to 0 or turn that on, assign some selectivity to it.

I imagine I'm thinking off the cuff a little bit. So, I see it as a possibility and, you know, in doing— using the MSE for these sorts of things is why, you know, one thing I was really excited about developing framework, I just wanted to develop a framework that could be used for lots of different things and we found a lot of uses out of it already. Size limit analyses, things like that. My hesitation, as you can tell, is that my program of work is pretty large right now. There's a lot of work, and we're talking about new estimation models, potentially new operating models, you know, and so on.

It might be useful to think about the program of work and the current program of work and those first two slides in the MSE presentation and what might be rising to the top as sort of now and what might be maybe a little bit later. And I think we did that with the assessment, you know, identifying— investigate state space models for 2028. That's helpful, those sorts of thoughts. Would it be— well, then just putting a note here, the SRB notes that this is a lower priority.

If you feel this is a lower priority, but it might not be, and maybe it's like developing a new operating model, the lower priority. Yeah, it would, you know, it would be helpful. I found it because I just noticed this list, this program of work can expand real quick, and having the prioritization helps me focus on certain things.

The way I might think about it is instead of making it a specific request to yes, please do this, more have it be a request that future, um, work on the framework would allow this kind of scenario to be evaluated. And so that way it kind of lowers the hurdle model of what it would take to do it in the future.

Especially if you're going to be making, like, if you're thinking about making wholesale changes in the operating model in the relatively near future, just trying to think through all the things you might want it to do and then make sure you have kind of pathways forward with that would be a useful exercise.

Yeah. Yeah. So maybe something like a request that your operating model development include the potential to use the MS-8 for investigating well degradation experience. Would it be useful to maybe even during our freeform session to maybe brainstorm arm on a few other concepts that would want to be included in there, such as reference point analysis, size limit analysis.

Okay, so I think we're, we're now back on track and ready to do the presentation.

Ray, if you're listening, if you can come on down, so we will jump into Paper 08. So this is under General Item 4.2. There's nothing under fishery-dependent data. 4.2.2, We have the fishery-independent data section, and GSA is going to present—. Did you want to use a snap?

Do we need to maximize this? Yeah, you can.

Thank you.

Okay, so this presentation will be focused on the upcoming survey for 2027 design.

For that survey, you reviewed initial take on that design in May. We'll also look at designs for upcoming years so we can see how the survey is planned to progress. And then some modeling updates. And at the end of that, that's the shorter part of the presentation. To give us some background, especially since we have newer members of BSRB, our survey is the most important source of data on Pacific halibut.

It's the data we use for getting our indices of density and abundance, the weight per unit effort, numbers per unit effort. Numbers feed directly into the stock assessment. Weighted mean method is used for estimating stock distribution.

And they're, um, one of the most reliable methods we have for assessing the trends in the stock over time. We also get biological data for the assessment from the survey, and we've had an annual survey since 1993. Over the first years of the The survey design was expanded to encompass all IPHC regulatory areas, and we undertook further expansion over the last— over the previous decade into unsampled waters. We noted that commercial fishery was catching fish in waters deeper than the survey was sampling, so we expanded the depth range of the survey. We also filled in some gaps in coverage that had not previously been included.

So, the full FIS grid is on a 10 by 10 nautical mile grid. It's shown here, represented by the orange dots, 1,890 stations from Northern California out to the Western Aleutians and the Russian boundary in the Bering Sea.

As well as our own survey, we have— we make use of data from is troll survey in the Bering Sea. It's a very large region with very low density of Pacific halibut. We have fished it a couple of times, at least southern two-thirds of the Bering Sea, in order to produce a calibration with the troll survey data, which just allows us to make use of the troll survey data in years in which we can't survey ourselves. We also have additional stations most years in Northern Sound, which is this far northeastern part of the Bering Sea in Alaska, from the Alaska Department of Fish and Game troll survey there, which is a similarly run survey to the Nooks Trolls survey. And some years when Nooks haven't run their surveys up there, we've— that's been the only source of data we've had in the northern Bering Sea.

So, recent history, this shows the more patchy nature of our survey over the last few years. The lighter colors represent the stations we fished most recently. The darker colors are stations we haven't fished for 5 or more years. And so we can see that the effort has been concentrated in recent years in what we refer to as the core of the stock, where the densities have historically and are currently highest. From 1B, 2B, 2C, 3A, and 3B, with less effort expended in the fringes of the stock, 2A, 4A, 4B, and 4CD.

These are areas that not only generally have lower densities, but they're also much more expensive to sample. And we've had constraint— financial constraints on the survey in the last few years, and those areas have been the ones that have been harder to include in the survey. And not only we have financial constraints, but we also have less interest from— the survey runs is undertaken by charter vessels. They need to post bids on the work, and if it's very expensive, like it will be with 4A, we're less likely to attract bids. And that has been the case in recent years as well.

We are trying to rectify that.

The objectives driving the way the survey is designed, starting with the primary objective, which is scientific motivation, sampling the stock for stock assessment and stock distribution estimation. And that generally determines the distribution of stations, at least what we want to achieve with with the station distribution, the number of stations, and how much— how many skates per station.

But we also have to bear in mind cost-effectiveness. This is a survey that's been— until very recently, was meant to be self-funding. We fund the survey by the sales of the fish that we catch.

And so we have to consider the effect— cost-effectiveness of the design, so we're balancing operational feasibility, logistics, and so on with the scientific needs. And in recent years, we've set an aspirational target reserve of $4 million, so that when the survey is not expected to generate the income in the year that we need to cover the costs, we have reserve there to enable us to fish the stations that we want for our scientific needs. And then we also can use the survey as a platform for other purposes. We've done work with other agencies on a cost recovery basis. Sometimes commission priorities extend beyond what we might need scientifically, and so they may want to add additional stations to the survey footprint for other needs, and so on.

The review process, essentially, for reviewing the designs of the survey essentially extends throughout the entire year. Our annual meeting occurs generally in late January, and following that, the Commission has essentially a final design for implementation that summer. And so following that, we begin the process of preparing designs for the coming year, and we're installing that in the spring. Scientific Review Board gets to review these for the first time in usually in June. This year it was in May.

And, um, then they might suggest further design work, and it's been sort of, um, difficult to do that given our cost constraints. There's the opportunity to do that, um, update the designs and look at the— one of the designs, um, over the summer and over the some of the current design is being implemented, and we start to get updated information on costs, prices, how much to sell, fuel costs, vessel costs, and so on that we can use as inputs to predicting the cost of the designs for coming years. Commissioners saw revised designs last week. And you are getting another look at the designs. Following that, the data from the current year survey will be finalized in early October, including quality control and finalized costs, which will then be used to revise the cost projections from the ones you'll be seeing today.

So what, what you'll see today in terms of cost projections still remains preliminary.

And the Commission will make a decision in November, recognizing that we can modify that somewhat out to the annual meeting.

There we are. Okay. Some recent history. Following the expansion the previous decade, we developed what we call rationalized FIS designs, recognizing we can't sample 1,890 stations every year, nor will we necessarily need to for scientific purposes. So we propose that we would sample randomly from the stations in the core to maintain consistent spatial coverage in the most important part of the stock.

Randomly sampling from the dense grid is less efficient in general because of increased running time between sessions. Which makes it harder to fish a compact block of 3 or 4 stations in a single fishing day. So that cost is kind of amplified once you get away from the core of the Stauff, and the costs are greater anyway. So we propose sampling blocks of stations, groups of stations, elsewhere prioritized to maintain precise estimates with low potential for bias. And, but in 2020, of course, a lot of surveys were impacted by COVID-19.

We did run a survey in 2020, but it was only in the core of the stock. 2021 To 2022, our proposed designs were largely implemented, the random sampling in the core of the stock, walks at stations elsewhere, but we were impacted by a lack of viable bids in the western part of the stock in 4B, and we did not sample stations that we had planned. 2023 To '25, we had reduced designs implemented due to financial constraints, as I mentioned earlier. We had little sampling outside the core of the stock in 2023, with no sampling in 4A, 4B, or 4CDE. Further reductions in 3A and 3B in 2024, but some sampling in 4CDE, commission priority that year.

Sampling in 3A and 3B was, um, targeted to complement the sampling of the previous year. So at least we had a picture of the stock over, um, for at least 3 or 4, 2 or 3 recent years, a more complete picture of the stock. And we sampled higher density parts of 2A and 4B, 4A, 4B for the first time in 2022 last year, and saw that play, how that, um, graphically played out the reduced designs in the figure previously. Gaps in recent coverage. So this is the design that was adopted for 2026.

We have again very high density of sampling in 2B and 2C, which is— which are the areas that are revenue positive. Essentially, no policy at the cost of sampling elsewhere.

We had, again, complementary areas for sampling in 3A. I should have mentioned perhaps that we'd move— that's coming up in slides. I mentioned we've been sampling at a random sampling from the core of stock. We changed that a couple of years ago to more complex sampling as costs increased, more operational efficiency in the core of the stock. So we don't have comprehensive spatial coverage.

In a single year, we have patchy coverage in a single year, but we're trying to build up comprehensive coverage over a relatively short period of time. But we didn't have the resources to sample every area that we targeted. We had done 3 of the 4 charter regions in 3A, 1 in 3B out of 5, and this charter region in 4A, which we had not sampled since 2019. We did not get— again, we had difficulty getting a viable bid for 4B or any kind of bid for 4B, and we did not sample it in 2026. 4B is also considered a biological region, which Ian's probably talked about in his presentations.

And so we are trying to get data from every biological region for the assessment each year. Not able to do that before.

So the base block design, sort of what I was alluding to, um, considered our optimal long-term FIS design in terms of balancing our scientific needs and cost-effectiveness. We prioritize some annual sampling in each biological region for assessment purposes. We ensure that all charter— the design ensures that all charter regions core of the stock, while not sampled annually, are sampled over a 3-year period. So we build up a picture of the stock over a relatively short period of time. And the coverage in other areas prioritized to minimize bias potential and maintain relatively precise estimates.

So while having gaps in the coverage on a single-year basis does bring in the potential of bias, we seek to minimize that by by having at least comprehensive coverage in the core of the stock over a short time. And the sample blocks rotated over time, and the Commission has noted that the use of this design will be the focus of future planning and annual fiscal proposals from us, the Secretariat. So, we're having gaps in coverage. It's important just to not just look at the current year or the next year's design. We need to look at where we're going in subsequent years as well to see how addressing those gaps.

And over a 3-year period, as I mentioned, the core of the stock, 2B, 2C, 3A, and 3B would have comprehensive coverage. But not— that is not the case over a shorter period of time. So in 2027, the design that we propose, base clock design, features sampling half the charter regions in 2B, 2 or 3 charter regions in 2C, half of the charter regions in 3A, and 205 in 3B, with additional sampling outside of the core of the stock in 2A, 4CD, and 4B. And again, this area in the far west of 4B is the region we tried to sample in 2022 and got no bids on, and we'll see how it goes.

It has not been sampled since 2019.

2028, Basically trying as much as possible to fish complementary charter regions in the core of the stock. This almost completes the design except for one charter region in 3B, which will be picked up in 2029. And again, rotating the stations in the fringes of the stock as well. And We also have designs for the following 2 years, which I think I'm going to come up with in a bit. So we do an evaluation of these designs based on the output of the space-time model that we did each year to estimate trends.

And so essentially, we're using simulated data from that model, data generated from that model for subsequent years. We don't project any trend in those years. It's just essentially the output for the terminal year, samples from the Monte Carlo samples from the final year of the model, and those are plugged into the designs for the subsequent years. So this predicts the coefficients of variation for each regulatory area and biological region for the base block designs that you just saw. We have a target CV of less than or equal to 25% by regulatory area.

And this achieves that comfortably in the core of the stock. The only place it— and close to achieves it elsewhere. The only place it missed in our assessment was 2A, and we address that by adding an additional 20 stations to the design we originally looking at in this assessment to address that. So, we expect that to be less than 25% or less in 2027 as well. But otherwise, precision is generally not a problem with, with these designs.

Our greatest concern is bias.

These designs have a lot of stations. So I mentioned them, we also have potential designs for 2030 and 2031. Again, I call them potential designs each year, at least in recent years, and probably, and likely in 2027 as well. The design is not the same as the base block design. Either it's been reduced to some degree, or it's possibly in 2027, it's going to be larger than the base block design, and we will modify the designs in subsequent years to account for the errors that we've, we've actually— but this again shows the rotational nature.

It shows the design. It shows where we're picking up parts of the design that weren't included in previous years, in the Bering Sea shelf edge, and for both 4CD and 4A.

Okay, so the cost evaluation of these designs, that secondary objective of our design. So these are estimated with several assumptions, one of which we already know is incorrect. The designs are optimized for the number of skates for sets of 4, 6, or 8 100-hook skates used. Sometimes we make more money with a longer skate, generally in the core of the stock, you set more skates, you catch more fish, make more money to offset the costs. And that's not always the case on the outside of the stock.

And depending on projected catch rates and bait costs, and projected Pacific halibut average sales price of $27 is unchanged from our best estimate of $26, and that has evolved since we made these projections now that the survey is complete. Projected halibut landings per station remain unchanged in each charter region from the 2026 values, which were based on partial data from 2026 and data from last year. The price of the chum salmon bait that we use remained at the average price for 2026 of $2.40 per pound. We already know, assumption is not true. It's close to $4 a pound.

We know because we were boarding— for next year, we needed to secure it early. And so that's going to be a factor that increases the costs that you see projected today. Other factors might reduce that, like we had higher catch rates in some areas than we expected. We had other factors, other costs that might be different as well. Prices were also high.

This year. So this is the zoom in on the base block design for 2027. Projected net revenue was a loss of $131,000, which if you're new here, you might not recognize how good that is. That is, when I first saw that, oh wow. And the Commissioner basically said the same thing last week, that this is as close to revenue neutral we've been for quite some time on the design that we're doing together.

Is that mostly because of the rice? Rice, exactly. How much has the price of palta changed?

Uh, so the year before it went up by like 30% from 24 to 25. This year it's gone up—. I don't have the number—. It could be another 20% or something. It's 5 26 years.

Yeah, adjusting for inflation. Yeah. Um, it's the highest in 26 years, but we're still, still, uh, right, because we're, we're not only are we—. That's, that's true. And part of this is because at the same time catch rates have gone down.

Um, it's a long-term trend in the stock is, is downward in terms of weight PR output. And so areas like 4B, which used to have high catch rates, are now very very expensive to fish. 2A, it was never that great, but it's, it's, it remains expensive to fish. But these, while a small component of the stock overall, are of local importance to the people who fish there. And so we still want information from, from those.

4CD in particular has, has a lot of interest from, from the Pontus, halibut fishers, those who fish other gear there because the high Pacific callopod bycatch affecting those other stocks. So yeah, it's still an expensive undertaking, but this still represents a much better than expected outcome. If only that stays like that. If the costs are revised. Price this year is, uh, 19.5% up compared to last year in Alaska.

And on this itself, we achieved 30.8% higher price, right? So that's 30% increases over each year over the last 2 years. Um, and we're optimistically— we've been told to optimistically assume that that's going to be sustainable, not the increase, but that the price we achieved in 2026 will be maintained next year. We've usually tried to be a little more pessimistic on that, to have a bit of a buffer, but commissioners, I guess, I feel— So this is a breakdown. This is mainly for the commissioners, but again, I note that the cost estimates are preliminary and do not fully account for 2026 realized costs.

So there's considerable uncertainty in these projections.

And that uncertainty gets larger the further ahead in time you get. We do have projections for costs, which we presented to commissioners last year for subsequent years. I don't think I included it in this presentation because it's not your primary focus. That's something that's important to keep in mind.

And we'll update these projections very soon. Fortunately, you don't get to see that. So, um, with things looking kind of good this year, um, we were asked to prepare additional options for— sorry, options for additional sampling.

For the FIS. And we prepared 10 modular add-ons that propose design that you just saw. And they have a number of goals. One is reduce this loss potentially further by adding in a couple of charter regions that are projected to be revenue positive.

And that we've done this in past years to when things were more more expensive. We can also catch up on the regions that were not— that were part of the 2026 base block design but were not fished.

So that would— she effectively reduces our potential for bias because of unsampled regions from last year. We could reduce the heterogeneity in net revenue among future years by sampling some of the more costly regions ahead of time. We have a kind of quirk of the design as it stands, the base block design is that it bounces around between close to revenue neutral and million-dollar deficits. And it just happens that— I mean, there are practical reasons for that. We're trying to fish regions, charter regions that are adjacent to each other to make it more appealing for— to fit on the survey.

But it turns out that adjacent charter regions either do well or they do extremely So that— bounces around, this would mitigate that some. We had one modular add-on or change that would fix the charter rate at a high value rather than allow the charter vessel to bid on a particular rate in order to increase the chance of receiving bids. People might not be put off by thinking, well, I might do that, but they would never take what I want for it. So we basically set it at how much higher than we'd ever got it for? I don't remember.

It's a substantial amount higher than we'd ever got for this charter region in order to attract bids. And also fully clear the backlog of uncertainty created by reduced designs over the last 4 years to provide the most informative data. For assessment management groups, full catch-up. And I'll show you what these— these areas— where these areas are in the next slide. But the first two were the revenue-positive areas that we could add back in.

Commissioner Smith, I think, we're not in favor of doing this, given that the relative increase in revenue is quite modest.

And fishing for the sake of making money is, even though essentially that's how the survey is funded, it shouldn't be a primary goal in the Commission's eyes. These other areas are, we have our catch-up area, Shalagok, we wanted to fish last year. That's 26, 22, we wanted to fish this year, but we didn't. So that's done within 2024. so this would catch that up. Um, Haydack is a 4B charter region that we did not fish in 2026.

Um, part of it was last fished in 2021. This would allow us to catch that up. On Alaska, um, is supposed to— western part of 4A is planned for 2028, bringing it forward, but it's high cost. Bringing it forward would reduce the costs in 2028. Likewise, the Bering Sea Shelf Edge for CE North and Central are planned for 2029 and 2030.

They're high cost, bring them forward, reduce costs or potential reduce costs in subsequent years.

And then, so, these have been already had— I've already had funding agreed to— agreed by the United States with supplementary funding of $749,000 to fund options 4 to 9. So we have add-ons to the survey in addition to the base block design already budgeted, at least at these projected costs. And then we had a couple of other options: expand sampling to 2-way. We have some sampling plan for next year in 2A, that's California, Washington, Oregon, and that fixed contract rate in ADAC, which would be an alternative to Options 6 or 7 at very high cost.

So that's where these options are. The options that the Commission has seemed reluctant to go ahead with are Options 2 and 3. That's here in 2C and here in 3A. And then the options that we are going forward with, 4 through 9. So, Shalakov here in 3A, Unalaska, ADEC, and the Durgan and the Northern Bering Sea shelf edge.

And the base block design included the southern shelf edge, so this will be full sampling of 4 CD stations, as well as the islands, where density tends to be relatively high. So this would represent a substantial improvement over the surveys we've had in recent years.

So, in summary, the base block design has a preliminary projected net loss cost of $131,000, which bear in mind, it's very preliminary, will be updated very soon, which can be implemented on its own using FIS reserves without the need for supplementary funding. Two extra areas would add potentially $62,000 more revenue. With supplementary funding, the design is able to add charter regions in regions 3A and 4B that would part of the approved 2026 design but not fished this year, and bring forward one or more revenue-negative regions to reduce projected costs in 2028 and potentially beyond that.

And we are asking you to note that this— the paper, which includes base block designs for the next 5 years, including preliminary projection of revenue and expenses for 2027 and potential modular add-ons.

So there is a much shorter second part on modeling updates. I don't want to proceed on that or discussion of this, which is quite a distinct topic. Let's see if there's any questions.

Skate. Skate. What is a skate? And so the— fair enough. The, um, The sets that are long strings of gear, each skate is a piece of that string with 100 hooks.

It's just a unit of gear and it's basically our primary unit of gear. We don't fish 50 hook bits of gear, we fish 100 hooks and a minimum number of skates is 4, a maximum of 8. Beyond 8 becomes hard to fish much in a day. At beyond smaller than 4, and we start to get what we call an end skate effect, where ends of the skate tend to have higher catch rates because they're less— competing less for fish than the hooks on the interior part of the skate. So you have slightly higher average catch rates.

Fewer skates you get, the more impact that has on your average catch rate. So we have them orders from 4 to 8. Really doesn't make much of a difference. So our range is 4 to 8, tend to fish 8 when we get a lot of money out of that, tend to fish 4 when it makes more sense. Fish that, sometimes it's a compromise, sometimes the— this year we reduced from 8 to 6 in one area because the vessel just couldn't manage fishing such large scales with a relatively high density of crayfish in UK soils.

Is that enough? Yes, it is. You can picture it.

Great. We'll talk more about potential ticket bids for the—. Yeah, not sure what I can say. We're not at the part of the process yet where we're soliciting bids, but It's, it's going to be, it's going to be difficult for, for B, um, as well as being an area generally of relatively low density, is very expensive to get to. The tides can be, um, very strong, and so the window for sampling can be quite narrow.

Sometimes you have to wait it out, and they will go in a very 2-week period where you try to get as much done as you can. And so, and there's no processor in 4B to which to land. There used to be a processor in Adak, which hasn't been for a few years, I think. And so you have to long run to go back to Dutch Harbor, I think generally, to just to land the fish. So it does make it logistically challenging.

We did have sample ferry stations in 4b last year successfully. So, and there's also very few vessels that are equipped that have the experience to go out there as well. I don't think the vessel that went out there was super happy with the work, with the conditions that they came across.

But for one vessel that we did, this 4b expansion in 2017, they fished all 200-something stations. And that vessel is the vessel that has fished foray the last couple of years. And they also seem— we discussed that there was potential for them to fish some Portby stations this year as well, but in the end they decided not to. So it really just depends on our ability to negotiate a small number of vessels and ensure make sure that they're aware that we have the potential to give them more money than they think we're able to give them in order to get this work done. And as I said, some of these stations haven't been fished since 2019.

That far— the Artut region in the very far west next to Russia— used to be an area with quite good catch rates, and we really have no idea what's going on. There's not a lot of commercial fishing, so there's no real No real data from any source. I don't know what's going on there. So, you know, we, the modeling will come up with some numbers for there and we'll form the trend, but really, we really have lost our understanding what's happening in that part of the stock. So I don't know.

I don't think Kayla, our survey manager, is super optimistic about our chances of getting vessels out there. So we'll see. So there's not really a It's not really a plan. Well, at least we— well, we have a plan. It's just can't happen.

Okay, a plan with a realistic chance. It sounds like you're not very optimistic. I don't think it's impossible. I don't think we think it can't happen, but a lot of things would have to go right, I think.

I'm just wondering how accurate past revenue predictions have been, and I'm curious if that's been part of sort of the spatial evaluation of the success of the model, because doing these sorts of spatial predictions is something that You have a really unique system here where you're kind of doing that for every survey design, trying to predict what you will catch, and it's not something I've seen elsewhere.

Yes, so they've varied a lot.

We have experienced losses much greater than we anticipated. Experienced revenue greater than anticipated in the last couple of years. So I think that's the goal of having a reserve of $4 million. We're never going to get it quite right. We can't anticipate how the price is going to change from one year to the next or how fuel— fuel went up this year.

We didn't see that happening last year. So that wasn't built into our projections. So there's a lot of variability. But the commissioners, we do present tables showing We present a sensitivity table, shows that— which I haven't done here, but we, while we're projecting $131,000 loss, we could, with just a 10% change in price or 10% change in catch rate, we could be looking at a million-dollar loss or a million-dollar positive revenue. So there's a lot of uncertainty in the even if once we've got the final data.

So while we can have a reasonable projection of what I think the stock catches per station might be, the variability of the cost in the prices in particular, there's a lot of uncertainty there. So I don't know if that answers your question, but having a target reserve, having a reserve that will allow us to weather those that uncertainty and not have the reductions in the design that we've had the last 3 or 4 years. Has there always been a reserve? It's just grown in size? It's come and gone.

We had a reserve and it was like, was it a couple of million dollars or something? No, I think back in— When you got here? Before the expansion study, we had like $6 million. $6 Million reserve. Expansion study used it all up a year in advance.

And then we've slowly been trying to figure out a way to build them. We were essentially asked to spend that reserve down. Yeah. So now we understand, now that times are tough, we understand that the value, Commission understands the value of having a reserve.

I guess my part of my question I was most interested in was the how accurate the predictions of catch rates have been.

Those are also widely variable because even though we have a general sense of where the coast-wide stock is going, the relative abundance in the areas can change pretty dramatically year to year. So our ability to predict what we're going to catch in an individual charter region is quite limited. And so we've got those two axes, the catch rate and the price together, as Ray indicated, less than a 10% change in one or both of those gives us a million dollar or up as much as a $2 million difference across the two axes in just relatively small changes. So it's really, we have this tool that allows us to do planning, but it's Nobody, nobody calls this a prediction.

Have, have you gone through and like calculated the currency and the cost from past years? Yeah, I mean, every year you have a projected cost and every year you have an actual cost. No, we haven't done that.

Costs vary from Well, no, I mean, revenue change. Yeah. How good is that projection essentially?

But then that would also bring in the cost that we have for a lot of these things, which is also a preseason guess. So, I think the challenge with doing that is that we've produced these at different points in the process. So, some years we've done them as late as January. And in other years we've done them as early as late summer, so the information that we have is wildly divergent in those periods. It's really hard to compare anything.

But we can look at the kind of base reference, for example, the current year, where we are this year, that varies quite substantially between regulatory areas. So I'm looking in terms of at least the revenue, and it's a slightly different way how it's projected. But if you look at just a kind of general reference, what you would kind of think from going from status quo, things not changing, to where we are this year, and it was like between something like between Charter Regions, from 116, 154% in some of the regions to down to like 70%. It varies depending on the region, but it's substantial.

Ray, do you have a sense for where we are on the increasing potential for bias from the examples?

Is that just continuing to increase or is there a bias aspect to it? I think we don't know what's going on.

We've gone to— so we looked at how, I think a couple of— it was last year or the year before, produced a figure showing how large the average index can change and how much the average index can change in 1 year or 3 years.

We can see changes of up to 30% in 1 year, or more likely over a 3-year period, either up or down in the, in the Explorer and IPHC regulatory area. Probably worth bringing that figure back up again because the commissioner— commissioners are asking last week, do we even need these stations that sort of, sort of building here? Um, um, but so, you know, if we're— if it's been 7 years since we've done a part of a charter region, most of a charter region, or in terms of OB, it's like half biological region. Yeah, if we've got a consistent trend that we're completely missing, bias very large. I don't know what number, it could be over 100%, it could be in either direction.

So, and that's half the Charter Region. We did sample for B Summer last year and it at least revised things a little bit. Yeah, maybe to put that in perspective for the Coastwide stock, we've had 5 to 7% of the of biomass in 4B. Right. So, if we're off by 100%, it's not a huge bias, coastal trend, but in terms of making management decisions about what's going on in 4B, it's very—.

And it's a biological region that you need to bring those in assessment as well. Whereas 2A is— also has a couple hundred stations like 4B, but it's much less. It's a very small component of the biological region too. How much anxiety is there on the part of the Commission about 4B, for example? Is this something that's really worrying them, or are they thinking more small coastline, small contribution to the coastline?

Good question. I don't hear as much concern about 4B as I do about 4CD and 2A, but there are very strong local stakeholder interests in what's what's going on in 2A and industry and external interest in what's going on in 4CDE. So we hear more noise about those areas, which is why they were added on the Commissioner's direction a couple of years ago. So, maybe because there's not so much fishing going on. Yeah, in the last couple of years, the US has provided additional funding.

Specifically sampling for bees. There's—. There must be some sort of level of pressure, I guess, to make sure it happens, but we don't hear, you know, from stakeholders specifically that needs to be done. Uh, and then that comes out in the lack of bees as well. It's just part of that is that the sole processor out there closed a few years ago, and so there's a bunch of quota that's effectively stranded because it's assigned to small vessels that they can't fish very far away, can't get to the processing facility.

But should that processor reopen, I think it's much more of an issue. The quota take on Area 4B is just like— 41% now, and the quota is taken substantially less than any other areas. And it's been a trend over the last couple years that walleye is underutilized in the expansion. It's not a big chunk overall, it's, it's been a couple of years.

One of the pressures to survey for A, B, and C, D, and E is the bycatch limits ABM is determined from EBS trawl survey and the FIS index from all of those areas combined. And I think some pressure does come from the trawl industry to, or I don't know if it's coming from them, but there is pressure to, there's comments that IPHC has not surveyed and thus they can't trust the index to decide on the right level.

This is probably not as far out as we want to go in talking about this, but why did the— sort of close that there.

There were a lot of reasons.

Corporate management. Issues. But I understand that the owner walked away and the equipment wasn't decommissioned, so the existing structure out there, it's— after a few years of sitting idle, is actually just like— somebody buy it.

It's always been a challenging area. There's not— it's a difficult place to— for a processors. Again, you'll see that just—. Probably more dependent on other species than Pacific halibut. It's the processors.

Halibut, for most processors, especially out there, halibut's a bigger problem. It's being driven by other catchable species.

I was just trying to understand the sort of A number summary, this $110,000, it's— that's the cost to, to run the survey for— that's the anticipated cost of doing the survey for next year, right? Yes. Um, right at the top of this one, this one here. Yeah, that's the difference between our projected expenses and our projected—. So that's— so if that number is positive it's not costing the commission to actually do the survey, right?

We have some money we can set aside for next year's survey. So looking back in time, it sounded— you said at the outset that's a pretty good number. So for, for the recent, the last 4 or 5 years. So is the, is the, the goal, I guess, of the commission to get that always in the positive, obviously, like it's not costing them anything to— I'm just trying to understand it in the context of how big a deal is it. Yeah, and I, I don't think it's, it's necessarily a goal to, to always get it positive.

It's just— and that's why Ray was saying, that's why we have this, uh, $4 million reserve aspirational target, um, it's to make it self-funding.

And in whatever form that might take, whether it be each year it fluctuates and we just absorb that loss or full gain accordingly, stay at or just below that $4 million reserve. That's probably sufficient at this point. There is discussion around as well whether we should be funding the survey with the sale of fish at all, should be government funding because there are no funds that are likely to come in in sufficient and consistent volumes to fund this.

A lot of this dead anyway, so we might as well sell them. Yeah.

Okay, anything else, or should we—. Okay, the last one's going to be quite short. Um, it is just, um, it's sort of an extension of what you heard from Colin yesterday.

Um, let's see. Okay, the information's there. So at present, the models that we fit, the spacetime models we fit to the data, um, more than this that you've just seen. Um, I fitted at the IPHE regulatory area, and then so we fit a model for each regulatory area. Um, when I started doing this, it was all very slow.

Um, trying to fit anything larger would have, would have, um, been, um, unfeasible, infeasible. Um, and then the output of those areas is aggregated to produce estimates for larger geographical unit, biological regions, coast-wide waters. But fitting models to coast-wide datasets has become more feasible due to substantial reductions in the runtime. Software has improved, hardware has improved. And so I'm looking at sort of a longer goal.

Well, one of the goals is to see if I can put everything into one model, one coast-wide model for the, for the estimation of the indices of density and abundance with parameters that vary across space. So we incorporate the heterogeneity that arise naturally from fitting models area by area, but it also gives us potential to model other aspects of data that we collect, including biological data on individual fish, which may have shorter time series or larger spatial gaps. Than the fish catch rate data, and maybe possibly modeling the environmental data that we collect on water column profiles. So almost every year since 2009, we have been collecting environmental data on every station, assuming things go well, typically, mechanically, using water column profilers. So we have dissolved oxygen, salinity, pH, chlorophyll.

Thank you. Those are the main ones we have good quality data for. And bottom temperature, which can be used as environmental covariance in some of these modeling, or just modeled on their own in order to understand the environmental conditions that are occurring at the same time as the survey. And we have used those environmental data for some modeling. In 2-way, we have dissolved oxygen concerns.

We have hypoxic events that occur off the coast of Washington, Oregon on a reasonably regular basis that impacts gullibet catch rates, gullibet density. So, have had the— what are they called— the blob, the heat event in the Gulf of Alaska around 10 years ago. That impacted a number of fish stocks. And so we've looked at these environmental data and done some modeling to help relate that to what we're seeing with Pacific coho. But anyway, this—.

Oops—. This, though, is this sort of test application I've been working on, is applying these coast-wide models to the histological maturity data that Colin— and I presented this a year ago. Some preliminary models, and this still— I haven't done as much as I'd hoped to do over the last few months, and so this is still preliminary modeling, but it's an extension of what I presented last year. So, some recap of the data that I'm talking about here. We've collected histological data on Pacific halibut maturity since 2022.

Maturity assessed on individual female Pacific halibut and each fish in approximate capture location given by the midpoint of the location of the set, which we have, which was captured. So we have specially these data not incorporated to Colin's analysis that we saw yesterday, they're spatially indexed. So, you know, do we have an area, but we have a specific location, ocean, at which these data were collected. So the recent modeling between maturity probability and age has used generalized additive models to estimate the relationship ships, but that's based on the assumption these essentially are independent observations. And, and, but we have more information on that.

So last year we presented results, preliminary results of a coast-wide spatial modeling maturity data, and there's more information in the appendix of the accompanying paper, which is what I presented last year. In fact, I'm not going to review that. Such a picturesque talk about, um, woody palm species. So the data, histological maturity data, I guess you saw this yesterday, I wasn't able to be present for most of Colin's presentation, um, has been sampled over the last 2 years, we have 3 years, we have, um, most of the data comes from regions 2 and 3, smaller amounts from region 4, and a very small amount in 4B. Additional data from 2025 that will be added to that soon.

But this is the data we're working with here. So it's not spatially comprehensive at this point. And from a year— from year to year, there's a lot of variation, and even within Regions 2 and 3, and where we're actually doing the sampling.

So last year we presented results of some preliminary modeling. The modeling was based on only 3 years of data, as I mentioned, with partially inconsistent spatial coverage across 3 years. The models used— the models that we presented last year, the results combined the data across 3 years without accounting for changes over time or temporal dependence. Now, As discussed, we not only have the 3 years of histological data, but we have 20-something years of visual assessments of maturity. And so Colin talked about the calibration that we did on their GAN models, and essentially we fitted all this to the visual data for 2022 to '24, histological data, and to the ratio of the, of the, curves, maturity curves, and they gave us our calibration, which we applied over time.

So we can build that into the modeling as well, spatial modeling, instead of having something that's done essentially outside of the model. So, including the visual data with the histological data allows us to estimate a calibration between the maturity curves. From the 2020 to 2022 to 2024 overlap, and also then model the entire time series of maturity curves, including modeling of both spatial and temporal correlation in the model parameters. And then we can predict the histology-based maturity curves, essentially correcting for applying the calibration to past years across space and time. That's the goal.

Um, And so far, so we have more background. The models are based on logistic model with probability of maturity related to age, log age as in Collins modeling because that's a better covariate, with spatially varying intercept and slope. So while Collins was fitting GAM models, we looked at logistic models as well, which is easier to work with because you have an intercept parameter and a slope parameter. Parameter on the logit scale, and we make those parameters very— both those parameters very spatial. And we add some GAN-like flexibility into the model through a random walk term in the age relationships, which allows some variation from a strict logistic model, which is clearly important based on epidemiology.

And so this year we extended those models to include temporal independence, the logistic parameters. So the slope and intercept not only vary spatially but vary temporally. We essentially— dynamic regression model where we use an AR model for the temporal dependence. And we have slope and also have slope and intercept fixed effects. And we allow those to differ between the visual and histological maturity data, and that's our calibration.

And so while the full dataset is available, I'm building this up one piece at a time. The more data we add, the slower it becomes, and so becomes cumbersome to troubleshoot the model. So I haven't been fitting the model so far to the full '24, whatever year— here's the data that we have.

But I've so far been using 2018 to '24, so including a portion of the historical other visual data together with the 3 years of maturity data.

So, this is just a very quick summary. The spatial dependence is strong, which is something we noted last year, but the temporal dependence is moderate. So, in the, in the FIS data, the catch rate data, temporal dependence is very high. It's in the order of 0.9 to 0.95. So, patches of high density There are patches of high density in Pacific caliper, and they persist.

This is less true of the maturity data, where we have early maturing fish. We might have a patch of that, but it's less likely to persist over, over time. And as you'll see, that makes it harder to use the model to predict in areas where you're not sampling, and then use those predictions to to build up a coast-wide picture, coast-wide maturity curve. And yeah, the calibration fixed effects produced exactly what we expected. Curves based on visual assessments, the result of lower maturity probabilities at age than those derived from histological assessments.

So, curves have essentially shifted, as you saw with Colin's— I saw part of his presentation yesterday, where he showed the A50s over time with visual and the histological, you know, there's a shift up in the histological curve, I say, and values relative to the, to the initial curves. That's essentially consistent with what we've got from the spacetime model. So as everyone likes maps, I like, like maps, these are the estimated A50s from the model we've fitted so far from the spacetime model. We have, this is mentioned, quite strong spatial correlation. And we have patches, really large patches where maturity, the A50 is higher, the lighter the color.

And we have variation within each regulatory area, variation between regulatory areas or biological regions.

Yeah, areas where maturity is higher. Maturity, where the A50 is lower, areas where the A50 is higher. And we have variation temporally in where those patches are and how strong those patches are in terms of how high the A50s are over this 6-year period. I dropped off 2018, even though it was included in the model, because it looks nicer having an even number of panels. But we also have big shifts in areas where we're not sampling model is predicting big changes in Bering Sea from year to year, but it's based on very little information.

And that is, I think, a consequence of the, one, the weaker temporal independent, temporal dependence of the model is estimating that the data are showing, and the fact that this is a very poorly sampled area with little data anyway. So, I— next slide, I've overlaid the station locations where we have not only visual, but also not only histological, but also visual data. And you can see that we don't have a lot of data in the Bering Sea at all. We have nothing outside the shore edge because we're not getting assessments from the Bering Sea trawl survey. We have consistent information otherwise in terms of weight development and age data.

Great. Sorry if I'm going to interrupt. Please do. What is the functional form of spatial structure you're assuming? It's a matter of—.

Okay, so that correlates to the core lesion parameters we showed on the previous slide. For the spatial structure, yeah, that's what I'm talking about. This slide? Earlier, I guess. Yeah, where I talked about spatial dependence being strong.

That's it. The range is— so we have a variance parameter, range parameter. Range parameter is very large. Range is large. Range is large, yes.

Is that a function of some of the— so you create a match? So how are you doing that? So what do you mean by create a match? So you've got these— I guess once you're using the survey You're using the survey locations directly, are you? That's correct, I'm using the survey locations as the location which we— yes, it's the observed survey locations.

Okay. So I'm not fitting or lining anything up. I'm not— it's not like all these squares where things fall into, it's the actual latitude and longitude converted to a spherical coordinate system.

The range is— it comes out as radians, but you can convert it to— I don't have the number off the top of my head, except that it's large and almost suspiciously large. But I want to look at that more closely. One thing, adding the visual data into the model is it seems to have stabilize some of these parameters a bit, just being able to go beyond the 3 years and have more comprehensive spatial coverage even within the areas that we're sampling.

But yeah, next time I present this, I'll have more detail on what these parameters are. Does that help? Yeah, I think, and I'm— did I understand correctly that you're doing this spatial, like you're doing, you've got these areas and then you've got regions, right? Which— And they're not part of the modeling. I ignore these arbitrary boundaries.

I'm just— to geographical. Okay. So the idea is that once we've done this without the artificial boundaries, we can then bring those in and aggregate predictions at the stations to form an value for the management units that we're interested in. And the way we did that, I think Colin might have mentioned that, the way we did that was we used the output of the modeling on this catch rate data, the numbers per unit effort. We get a prediction of numbers at each station and use that as a weight and then get basically a weighted average coverage maturity curve for a region based on essentially population size.

And so that allows us to fill in any gaps. It allows us to properly weight the stations when there are gaps in coverage. So one thing we get here with just the— what's been done so far is that your curve will vary depending on where you sample. So we may see trends in time curves, but part of it might be an artifact that we— well, we sample this area in 3A, and that's where the maturity curve is. So this is supposed to deal with that.

But one of the issues I have, which is the main question out of this brief preliminary results— whoops, sorry— is how do we deal with these large persistent coverage gaps when we're trying to build up aggregate picture when the result is, is we don't have strong temporal correlation. Um, filling in that gap is, is a lot more uncertainty with that. And, um, yet you want to build up a consistent curve for a region or for coast-wide. Um, is that something we can reliably do when, um, when the results look like this. So I think, you know, where we see high density of stations, we tend to get a fairly stable picture.

It varies with time, but it's not changing drastically. Where we have not much of anything for B, for C, D, E, we're getting great big shifts, which mean probably not. And so Do you ever plot the CV beside the—.

Certainly can do that. Oh no, yeah, I do. I can map so many plots. No, I can map that as well. Yeah, exactly.

And I've done that. I think last year I did that for— last year I didn't— I showed an A50, but I also showed the spatial maps of this, the slope parameter and the parameter, and also of the, the, um, standard deviation. Yes. So you will see that. You'll see big CVs out there that tells you exactly what you expect.

Um, and what software are you using to fit these? I'm using R—. Pardon? R—. Rnlux.

Okay. Have you tried Sean Anderson's? I haven't yet. It's just inertia. I—.

It's, it's— I started using Rnlux in 2016. Took me a while to get it all set up, and I know there are I think, I know at the time, well, not long after that, VAST became available, but it didn't have all the features I wanted. Like, and I think, but I think things improved. For example, I'm, because of the great distances that are involved in this, using a 2-dimensional coordinate system is not really gonna cut it because the distance vary, the lats and longs vary, but in terms of what distance means. So I think being able to use spherical coordinate system wasn't something that you could do initially in VAST.

I don't know if you can do it in SHORM. So I, it's interesting, about the same time I invited the people from who wrote our in-law and then the TMB folks who had the Laplace approximation and the first covariance matrix stuff coded up. To Toronto to a workshop just to try and understand which one to use. Right. And the Basque people.

So the Basque was incomprehensible at the time. We ruled that out basically because we couldn't understand it. Then the RN law folks and I mean Anders came from TMV so he could answer every question on the spot. So we ended up using TMB, which in— so, so, but we always used InLab to create the mesh, which would have been more to, to consolidate like the survey points, right? Yeah.

So, so I at that point sort of stopped, started using anything TMB related, and then Sean Anderson, I think, was down here at UW and did, uh, started this spatial-temporal modeling work using TMB. And so for anything survey-related on the East Coast, we now use Sean's stuff. I think it's the go-to. It's— yeah, but it's hard to— yeah, like, but I think it calls our in-law. I think you're right.

Yeah. Yeah. And I'm not up with— I know they've been making additions and improvements constantly. I'm not up with the most recent things. So it's— it comes up almost every year I present these things.

Yeah. Olaf talks about it and it's kind of on my list of things to do. And I think he's thought recently about exactly what you're asking, which is like, what do you do when you have this— these areas where there's no survey effort? Yeah, right. This, like he does like a cross-validation or something.

Um, anyway, it might be worth looking at, not to recode it all, but I have found his stuff. Yeah, no, it sounds great. Um, it's just finding the block of time here to, to try it out and compare it. Um, the other Yeah, for the meantime, what I'm thinking that, that we essentially is what's happened anyway with the curves. We just cut out that part of the area.

So, don't use any predictions outside the range of the survey and recognize that when interpreting this coast-wide curve or whatever regional curves, it doesn't apply everywhere. And that's true of the curve that we have now. Was also true of historical curve, which came from a very small subset of the stock. And the other thing is, can we— and, and Colin's touched on this as well— can we bring in other covariates that might inform the areas in which we're not able to sample? Bottom temperature, for example, I don't know if this is important, but we do have that from the normal stroll survey.

And so if there's some kind of relationship, it might help. We do that to some degree with the other survey data. We use depth as a covariate, and we— a recommendation of the SRB, and we use distance from the shelf edge as a covariate because density decreases from the shelf edge. So we don't have sampling here. It gives us something to— and so there may be things we can bring into the model and to help fill in those gaps.

This is really basic. There are no covariates at all in this, except for age, of course. And you don't have any— so the depth, I guess, you get from an external layer. You're not— We can't. Yeah.

So, depth— so, depth we use— the depth is recorded at the station. So, I use actual observed depth, the average depth of the set. But when doing projections, I use a bathymetry depth, or well, a little complicated. We, I actually use the average long-term measured depth at a station because there's some variation where the station fishes each year. So I take the average as the objectively good measure of depth, but we would never fish dissation, they come from bathymetry.

That's the best estimate. This is not a—. Doesn't matter too much. It's pretty consistent.

Yeah. Wait, can you talk more about the approximation to the GAN? How—. Right. So this is—.

You're wondering about the potential overfitting, over-flexibility in this, and you see it maybe in some of these small hotspots? Right. So this is where I I'd love to hear what you know about these things. I've tried to find people fitting GAM, really flexible models, spatiotemporal models, and I haven't found much of anything. People talk about it, but nothing.

Did you look at—. So I guess Simon Wood's MGCV? Right. No, I used that, but can we incorporate that into this?

So I'm fitting a logistic model because it's easy to fit a parametric model, even parameters vary in time and space, but that level of flexibility I hear, I read, at least the folks talking about GAM-like flexibility. We try to mimic the flexibility of the GAM, and here I've done it through a random walk.

Sorry. Yes. So, I'm not sure I fully understand. So, your model's logistic because of your response Right, it's logistic, but we're fitting essentially a logistic regression with a slope, whereas we're fitting a logistic GAM. It doesn't have a strict parametric form for that.

So I'd like to be able to do that in the spatio-temporal modeling, and I don't know if there's a plan. Now that would get me quickly converting to the TMB world. World if they offer that functionality. I don't know the answer to that. I mean, if it's—.

I would again have a look at John, what John Anderson—. I've tried, and I mean, I did, it was a year ago. Yeah, I haven't seen him in about a year. Yeah, so I mean, I'll do that, but when I looked, I didn't find what you're looking Yeah. So I have— so to get back to Olaf's question, the GAM flexibility is the coastline.

It doesn't vary spatially. It's basically— did I say GAM flexibility? I mean the random walk. So it's one random walk relationship between maturity and age. It essentially gives you deviations about the average maturity curve, deviations about the fixed fix, I suppose.

Is that there's a fixed slope and intercept term and then the random effects and the slope, especially this, and I showed it last year, a new little figure showing what the deviations look like and they make complete sense for the coast-wide curve. But it's pretty, I'm not worried about overparameterization because there's really only one parameter associated with that, that's variance parameter. So I think it's, and the results make complete sense. I have not worked out within the framework that I'm using how to make that more flexible. Tried different things to get the variance parameter vary, say, by region or by area, and I—.

Nonsense. I guess I don't understand how to parameterize that within the code.

So, my next step, so I'm working on, yeah, go ahead and slide. So, summary slide, we're getting to lunch soon. So, I'm testing alternative prior models that the, I don't know what, I haven't looked enough at Sean's stuff to know what priors they're using, but they love penalized complexity priors here. So, I mean, I don't see At least in its standard form, it's not a Bayesian implementation. It's a non-Bayesian, so there's no prior specified.

I mean, you can sort of do it with a penalty term if you want, but—. Okay. Yeah, that's the other big distinction, I guess, between MLaw and TMB. Right. You can run standard TMB to include priors, but—.

It's not required. Okay.

There are pros and cons to that, right? Yes. Yeah, I talked about the— yeah, and I'm also just checking the FedLM crops validation. So, yep, and extending the time series back to 2002. So we have that full picture of how maturity curves have changed over time.

Can I ask a question? Can you show the previous slide? Yeah. Because you just mentioned the temporal changes in maturity. What we saw in histology was a really strong temporal effect between 2022 and 2023, and then to 2024.

That's, that's not Not 100% clear. Clearly seen here. What would you think that is? Let's see. So let's probably should pick this apart regionally, especially in biological regions 2 and 3.

That's where we have 3 years of data. So region 2, we saw fish.

How did it go? They were maturing.

So the A50 value was, was basically decreasing over time.

So in 2023 was lower than '82, and 2024 was lower. That's right. And you're correct, that's not what we're seeing here. You—. I mean, if, if that was reflected here, you should see a lightening up of the color, especially in, you know, we see a darkening.

They're like, you're right. Yes. And this is the opposite of what we see. So that's, I think I did notice that. Don't forget, it's going to have— you can't just look at the change in color here because different areas are going to have a different amount of biomass.

And so the coast-wide curve is a biomass-weighted curve, right? Right. But I think they're talking about the regional curves that we So, in the regional curves, region 2, it was a clear shift of the curve to the left, which means the fish are maturing earlier. And that's not clear here. So, it'd be really helpful to have the same summaries for the two approaches.

And I did, I did do that last year. Right. But I haven't got to that point. Yeah, you're right. You're absolutely right.

You know, part of it is that those curves were— it may not be worth digging into this right now. This is without those summary curves, but we are—. I couldn't read the scale on the—. No, it's horrible. Sorry about that.

It goes up to 15 and down to, um, but the scale stays the same across these. Same scale stays the same. But I think it goes up to 15. I capped it at 15. There were some weird spots that went higher, but, um, outside the core of the stock, then down to 6, I think.

6, 8, 10, 12, 14, 16. Yeah, 16. Um, so we're generally getting in this around the 10 area for Region 2, and that's basically what we see, but it does look like it's lightening. But we have areas like from the Charlotte that are darker, um, from 2024 to '23. So it might be that it's hard to see the whole picture, uh, summary by taking in the No, you're right.

I think last year when I produced this, it looked more consistent. Have to look more closely at that. The other aspect is we're combining visual data here as well. And there's a potential that they're not as consistent in their patterns as the— I don't know if Colin remembers that.

But yeah, it's something to— that's something to investigate. What's the effect of including the visual data? On the trends here. Is there something else going on that we're not modeling? Have you tried to do the same with the histological data?

Just the histological data? Just the histological data. Right. That is what I presented last year because that didn't have any of the 19, 20, 21 data in there. Because I think there, then you—.

I think we did. Yeah. So it may be there's something going on with the visual data, we need some— Yeah, there's a lot of checking to— that's why I keep calling this preliminary. There's a lot of— and having these discussions helps highlight some of these questions.

There's some pretty big gradients over small spatial scales, and if you're looking at, like, the 2024 off the north end of Vancouver Island to Paraguay, right, basically spanning the whole range of And that is—. I'm sorry, wrong direction. That is a place where we didn't have any sampling. Um, so that's—. So, yeah, so where is that gradient coming from?

You have this over the standard deviation side. Yeah, if I had the standard deviation side, it would show a lot of uncertainty. I mean, I guess I would have thought that that would flatten out, that it would be close to where the sampling took place. So it would be because of the strong spatial—. Oh, I see there is sampling in 2023 there, but not in 2024.

Something in 2023, which is weak, relatively weak.

Yeah, I see. Okay, right now you can see it. It's right where the sampling stops, 2024 there. Yeah, you're basically using 2023 samples, I guess. It's drifting.

Yeah. Okay. Yeah, I don't want to get too far in the weeds, but it just helps to understand the model, what the model is doing. Yeah. And there's a lot more I should— I'll work up next time to help illuminate what's going on, not just for you, but for me as well.

As I said, it's very early look and I haven't sifted through everything. And then go, is it something that worked?

Yeah, and I know, you know, we've— the SRB has kind of pushed to go in this direction, but seeing some of the model behavior here, given the sparse sampling, so I think it's maybe not ready yet. Well, it's definitely not ready, but yeah, um, what would it take to get it? Ready for production application.

Yeah. If you—. I was just going to say, one of the things that surprises me with it is how high that correlation is for those spatial effects and how sort of blocky the regions end up being. I think because of that high correlation, I'm just— it makes me really want to try to visualize the data some way and not a model version to try and see what the model's picking up on that's really driving some of those really stark patterns. Is it like just a couple of fish or is it really a lot of fish informing some of those regions where you're getting a break or something?

And so— That's a good question. I tried to do that. A year or two ago, I actually worked up the code in R Shiny to do an interactive where you could look at what the individual maturity values were.

Trying to remember what it exactly did, but it was like that. You could interactively click on and off mature, immature fish and look at ages or something like that. There were different scales in there, but I can work that up again. Because I did find that helpful to ground truth what the original GAN models were showing. You're right, some of these stations, while you're building up a coast-wide picture, going into some of these stations have very few fish informing the station result, which is why when we coast-wide value, we are weighting it by the numbers per unit effort.

For that station. But yeah, that, that, that's a good suggestion to, to help understand as well. Go revisit what I did, um, for that because I, I did find it helpful.

You were gonna—. Yeah, I was just curious if you fit the whole time series all the way back to 2000 to what sort of impact does having more years of data impact the current years that you're showing here? Right. So, some of—. Because we are— the data that you're showing here, a lot of those years are reduced FIS design years.

So, if you start going back in time to 2000 earlier and where we actually have more consistent sampling across coast-wide, what sort of impact would that have on this? These maps that we're seeing here? It's going to have an impact because of the way the model is structured. We have parameters that are fixed through— some parameters are fixed through time. The range parameter and the variance parameter of the covariance temporal correlation parameter, those parameters are kind of global parameters that don't vary.

And so they'll be informed while the model is dynamic. It allows things to evolve with time. Those parameters will be informed by the large amount of data. And so we have a better indication, better estimates of those parameters, and we may stabilize some of what we're seeing here.

That's one of the benefits of including a larger than 3-year data time series.

But the more I added, the longer this thing took to run. 2018 Did a few minutes for just 3 years data. Couple hours for 6 or 7 or whatever it is. 7. So it does blow up.

Take it one second.

Yeah. Thank you. All right. Thanks, Red. Um, see what's here.

We have one more presentation based on information paper, so maybe break for lunch, come back in an hour, 1:30, um, present that and sort of be able to that kind of discussion. So that's good. We're not covering AI. I see, I see. Okay, that's an information paper.

Yeah, it's just we walk through information just because to run information paper about the 30-day deadline, and I got the chance to put some additional around. Okay, so for those listening online, we will adjourn now. We'll come back at 1:30, um, probably for about a half-hour session. Thank you.