SteelheadTracker

Counts through October 4 are in, and the chart from September 29 is redrawn below through October 5.

One change in how the Snake pages read. Until now they counted at Bonneville until October 1 and then switched to the Lower Granite scale, which put a 30 percent step in every number with no fish behind it. They now count at Bonneville all season, where their fish are counted first and the season is 95 percent complete by October, and show beside the finish the share expected to reach Lower Granite, measured from each river’s own tags over the last three seasons (about 0.78 for the Clearwater and Grande Ronde, 0.81 for the Salmon), with the Granite count so far as the running check. The grey line on the chart is that expectation.

Today: Clearwater 12,921 at Bonneville and 10,027 expected at Granite with 6,254 counted there so far; Salmon 13,053 and 10,534 (7,663 counted); Grande Ronde 7,680 and 5,960 (4,413 counted). The Clearwater keeps coming down and the Salmon has passed it. That is fish, not method: the Clearwater added 33 fish at Bonneville in the three days after October 1 while its own timing expects about a sixth of its run after that date, the late B-run, so every day those fish do not arrive lowers its finish. The Salmon is 91 percent in by October 1 and holds. In tagged fish the Clearwater has 155 past Bonneville where the last three seasons had 433, 899 and 399 by this date.

Also this week: the Granite expectation uses each river’s own measured share in place of one Snake-wide constant; checked against Idaho Fish and Game’s published per-basin numbers it halves the error. The method is on the accuracy page.

Every river page except the Deschutes now has a card called Fish on the way. It counts the tagged fish bound for that river that are past Bonneville and past the next dam up this season, how many have reached each sub-basin’s antennas, how many to expect in the next two and four weeks and the rest of the season, and when they usually arrive. These are tag counts, not fish counts, so read them for pace and timing. The flow column carries the NOAA river forecast for the gauge nearest the antennas and names the days a rise is forecast; the counts are not flow-adjusted, because I tested that against 14 seasons and it did not help the two-week number, while the rise days do say which days fish move.

The measured part surprised me. South Fork Clearwater fish reach the South Fork around the first week of March whether they crossed Bonneville in July or October; they hold in the main river through the winter. Flow moves them: on days after a three-day rise of 25 percent or more at Stites, arrivals on the South Fork run two to nine times the other days’ rate, in every month from October to March. Last fall was the second-highest fall flow on record there, and 60 percent of last season’s South Fork fish were in before January 1, against 0 to 31 percent in every other season.

Before putting the forecast up I ran it on each of the last 14 seasons as if that season were under way, and scored it against the simplest alternative, assuming the same arrivals as the season before. It beats that in most seasons for every river except the Dworshak trap, which the card marks low confidence. Researchers say the Columbia has been through several steelhead eras in this record; I tested training the forecast within those eras and it did not help, so it uses the last three seasons. The method and every score are on a page of their own.

Also in this update: counts through October 1. Bonneville added 155 fish in two days. The Snake river pages switched to the Lower Granite scale on October 1, as they do every year when the run reaches Granite, so their season totals and finishes now read at Granite and are not comparable with the September figures.

Four more days of counts, September 25 to 28, and every forecast on the site came down. Bonneville added 652 steelhead in those four days; the three past seasons this one is being compared to added 1,300 to 2,800 in the same four days. The forecast assumes the season keeps pace with those years, so when it falls behind, the fish still to come are scaled down for every river.

The Salmon and the Clearwater took the biggest hits, 15,000 to 14,000 and 16,000 to 14,400. The Salmon's own tagged fish simply did not arrive: none crossed Bonneville in the four days, where the comparison years averaged five. The Clearwater's did arrive, eight of them, but a Clearwater forecast leans on the whole run's pace because most of its fish come late, so it fell with the run. The Grande Ronde and Imnaha, mostly in already, barely moved.

Nothing in the model changed between these two builds. This is the first revision this month that came only from fish.

This is the first season this site has run, and the honest question is how much the forecasts are worth. Between September 12 and 29 the fish barely moved and the forecasts moved a lot, and most of that movement was me changing the model, not fish.

I spent the month working with Idaho Fish and Game on how their numbers are built. Every change was an improvement against their run reconstruction: the level error on the Idaho hatchery counts went from 22 percent to 11, the Clearwater-to-Salmon split from 6 points to 3.5. But each change also moved the forecast for the season under way, and more than I wanted to see.

The Salmon and Grande Ronde forecasts held together through it, inside 13,900 to 17,500 and 7,000 to 8,700 fish all month, and they are converging. The Clearwater's did not: 14,800, 11,600, 20,400, 14,500 and 16,000 across five builds on nearly the same fish. It is the latest-running river and the one every method change lands on hardest, and its band of plus or minus 28 percent is the model saying exactly that. I am not yet sure the Clearwater lines are going to meet.

From here the method holds still and the forecasts move only with fish. The first such step is in: four more days of counts, September 24 to 28, moved the Clearwater finish from 16,000 to 14,400, the Salmon from 15,000 to 14,000 and the Grande Ronde from 7,960 to 7,780, all downward because the fish are arriving slower than the analog years did. That is what a few days of data is worth; the earlier swings were not that. The chart shows each build's projected finish with its band, one panel per river, and it will be extended as the season runs.

I wrote to Idaho Fish and Game asking why my PIT tag count for Dworshak came out at about 60 percent of their run reconstruction every year. Their hatchery evaluation program answered. A PIT tag estimate under-counts the run because some tags are lost and tagged fish survive at a slightly lower rate than untagged ones. They call the shortfall PIT representation, measure it every year against their genetic count at Lower Granite, and it is not a constant: about 70 percent for the Clearwater and 90 percent for the Salmon and Snake over 2013 to 2019, but 82 to 92 percent and 89 to 101 percent over 2021 to 2025.

What changed on the site:

• The Dworshak correction this site fitted to the 2013 to 2019 reports, and applied to every year, is gone. It assumed 62 percent representation forever and was too large by about a third for recent years. Completed years now use IDFG's own representation figures by basin and year. Against the site as last published, the Clearwater hatchery count at Bonneville for 2021 to 2026 is up 5 to 14 percent (MY2025: about 37,600 to about 42,100). Most of that is a change made on September 18, when unclipped hatchery fish that the dam count files as wild were moved to the hatchery layer of the basins that release them, using IDFG's own proportions. The representation change on its own moved the Clearwater down 2 to 6 points of share against the intermediate build it replaced, which was never published.
• The current season is a PIT tag index with no correction, on their advice that a historic figure should not be applied to a run still under way. It will be revised when the year closes.
• The accuracy page now scores the Idaho hatchery numbers against the run reconstruction, which IDFG describe as their most accurate estimate, and shows their two numbers and ours on one chart. Until today it scored against their fall PIT charts and called that deliberate. It was not the right reference: those charts are the same method as this site, and IDFG say they under-count.

They also confirmed that 2019-2020 is the newest run reconstruction, with later years in progress, and pointed me to where unclipped hatchery releases are published by program.

Until today every river's run timing was Bonneville's. The model scaled the dam's curve up or down for each river, which changes how big a season is but not when its fish move, so all twelve forecast rivers came out between 73% and 74% through their season on the same date. The Umatilla is largely finished by mid-September and the Clearwater has barely started; they cannot both be three quarters done.

Each river now carries its own timing, measured from 10,291 adult PIT crossings at Bonneville since 2012. The spread is 58 points, not 1:

• Wenatchee, median 3 August, 92% past by 11 September
• Salmon, median 31 August, 65% past
• Clearwater, median 18 September, 34% past

Tested by hiding one year at a time and forecasting it from the others, this is right within 13.8% against 29.4% for the old shared curve. The Clearwater's projected finish roughly doubles as a result.

The forecast band is now the error we actually measure. Each river's band comes from how far that river's forecast has missed in the past, so the Wenatchee reads plus or minus 3% and the Clearwater plus or minus 28%. That difference is real and you were not being shown it.

Four rivers lost their projection. The Umatilla, Walla Walla, John Day and Asotin do not have enough tagged adults to measure their timing, and they were only ever showing Bonneville's curve. They now show no forecast rather than a borrowed one.

Also: the Asotin's 2015 hatchery number was zero because the reference table we anchor to has no Asotin row that year and the code read the gap as a count of nothing. And the hot and cold share chips had been frozen since 29 August because the script that writes them had been failing silently; run again, it says no river is running hot.

Every hatchery tag stands for some number of untagged fish, and that number was being built badly. For a hatchery that releases into several rivers the model used one pooled rate, so Irrigon's Imnaha fish and its Wallowa fish were both counted at 31 fish per tag when the real rates are 15 and 48. The Imnaha came out more than twice too heavy and the Grande Ronde a third too light.

Worse, the pooled rate itself was a running average that gave the last release in the file half the say. Dworshak's 2024 rate was effectively set by one small, lightly tagged release, reading 87 fish per tag against a true 64.

Both are fixed, and the Grande Ronde's current season rises about 78% as a result. A separate check now refuses any weight below one fish per tag, which caught 41,617 tags standing for less than the single fish they were in.

The error bars were also mislabelled. They were captioned as a 95% confidence interval and are not one: the low end is the fish tags have already placed in a river, the high end assumes the unplaced tags go the same way. For most river-years nothing more is expected to resolve, so the bar sits on the estimate. It now says that.

• The sub-basin chart used to split a drainage across its forks even when a fork's detectors recorded almost nothing, which pushed those fish onto neighbouring forks. Such a year now shows a single hatched block saying the forks could not be resolved.
• The calendar-year views on the destinations chart published 41 river-years where a river's Lower Granite number exceeded its Bonneville number, which is impossible. They are withdrawn.
• The projected season finish now appears on the front page and the phone, not only on the river pages.
• The forecast band and the average miss are two different quantities and were being quoted as if they were one. They are now labelled separately.

August 24, 2026: the site learns to talk, and every river gets a page.

New today: dedicated river pages (browse all 16) with each river's run timing, forecast wedge, season totals, and real PIT-measured travel times ("Clearwater-bound adults take a median 20 days Bonneville→Granite"); a daily "What the data says today" plain-words briefing on the Run Timing tab; the timing wedge charts promoted to the landing view; and a River-shaped alternate layout on the Migration Flows diagram.

3 September 2026. Two audits acted on, and the physics enforced.

Two more independent reviews of the site, twelve readers in all, and the numbers they found were fixed the same day. Upper Columbia rivers no longer sum to fewer fish than the dams that counted them; the Grande Ronde no longer reads more fish at Granite than at Bonneville; the chart caps and the river pages read one file of projected finishes; Cottonwood Creek moved to the Grande Ronde, where it is; three transposed and duplicated rows left the tagging-rate table; and which rivers are held to a typical share is now decided on the model's raw output, not on a chart the decision had already smoothed. The phone page is built with the site from the same data. One rate that stocked five rivers from one hatchery, the Klickitat's, is being priced by the river its tags were released into rather than by the hatchery's total.

2 September 2026. The season in progress, the Washington tag rates, and an audit.

The current season is now split from this season's own PIT tags at Bonneville. Each release group is sent to the rivers its fish actually reached in past years, which matters because Lyons Ferry stock does not go home and barged smolts are filed under a Lower Columbia release site. A river with fewer than twenty tags so far holds its usual share for the date. Upstream dam counts are no longer used as caps on the current season, because they lag Bonneville by weeks and a count to date is a floor, not a ceiling. Before this the Snake read 86 percent of the run so far against tags saying 47 to 50.

• Tag rates for Washington and Oregon. The Fish Passage Center's release table records zero PIT tags for Skamania, Lyons Ferry, Irrigon and Wells in most years; PTAGIS holds ten to forty thousand a year for each. The rate is now the PTAGIS count over the FPC release total, and where both sources counted tags they agree to within three percent. Fish tagged at the Lower Granite bypass are expanded by FPC's daily collection count there. The Fish Passage Center is now a named source.
• The Methow's history comes down. The Methow and Okanogan had been allowed up to two times the Wells Dam count, which is the tolerance a check uses, not a measurement. Seventy percent of Methow and Okanogan fish tagged at Bonneville are later seen at Wells, so the cap is now that conversion.
• An audit. Five independent reviews of this build, one per page, found about seventy items, from a Yakima count of zero caused by a missing map entry to sentences describing a method the site no longer uses. The numbers are fixed in this build, the copy is being corrected, and a second audit follows before anything is published.

1 September 2026. The level, and two rivers that were four times too big.

August fixed which river the run goes to. This update fixes how many fish are in it, which turned out to be a separate problem hiding behind the first one.

• Ten thousand fish a year were being deleted. The pipeline checked that the rivers never added up to more than the dam counted, and never checked that they added up to it at all. Every year came out at 71 to 95 percent of the real run, with the remainder simply gone rather than marked unknown. The hatchery level now sits about 2 percent above Idaho's published figures instead of 17 percent low every single year.
• The Methow and the Tucannon were roughly four times too large. Both for the same reason: the step that decides which river a fish is bound for was counting raw tags rather than expanded fish, so a river that tags heavily looked like it received more fish. The Methow read 13 percent of the Columbia hatchery run in 2012 and now reads under 3. Two mislabelled detector codes also sent Tucannon fish to the Tucannon that had never been there, including one that is a trawl in the Columbia estuary, 200 miles downstream.
• Every river now fits through its own dam. A river's fish must pass the dams between the ocean and home, so its number cannot exceed those counts. That check now passes for every completed year; the season in progress is not capped, because upstream counts lag Bonneville by weeks. It did not pass before.
• A new check the site cannot fake. Every PIT tag seen at Bonneville, multiplied by how many fish its release group stands for, reconstructs the season total to within about 10 percent of the physical dam count, without being told the answer. That is run by hand after every model rebuild. If the expansion ever breaks, this catches it, and nothing else could.
• Three missing years of Idaho wild data recovered. Spawn years 2017 to 2019 had been marked unavailable for months. They were published all along, as a picture of a chart with no readable text behind it.
• Forecasting Lower Granite from Bonneville. Snake fish pass Bonneville two to four weeks before Granite, so Bonneville's count is a real head start. By mid September it forecasts the Granite season to about 12 percent, against 34 percent for assuming last year. This site previously told you to assume last year. That was wrong, and the forecast page now says so.
• Said plainly what the wild numbers are. Roughly a third of the run is wild, and our wild figures are weaker than our hatchery ones. Outside the Snake there is no river-specific wild estimate at all. The accuracy page now spells out exactly which rivers and years that affects.

August 2026: a big update, and an honest one.

This site ran without updates from late February to late August 2026. During that gap I found several real problems in my own numbers, and rather than patch quietly I rebuilt the validation from the ground up. Everything below is now checked against published sources before it ships: Idaho Fish and Game's own figures for the Idaho rivers, the Yakama Nation's counts for the Klickitat, and dam-by-dam censuses everywhere. Rivers with no published figure are checked only against the dam counts their fish must pass, and a hard gate blocks the site from deploying if the numbers fail.

Fixed in this update:

• Eight charts restored: the live build had silently switched to an older script and dropped them (including the calendar/migration-year toggles).
• Ocean age corrected: 1-salt fish were dramatically undercounted by a day-counting bug; ages now derive from migration years and are validated against length distributions.
• One number per quantity: the cumulative run-timing chart and the destinations chart could disagree about the same river by 5x. Both now read from a single source. The Dam Passage tab's daily split by destination is a different quantity, the share of each day's tags seen later at each river's entry arrays, and it does not match these totals; it is being rescaled to them.
• Hatchery and wild estimated separately: the PIT-tag sample is 93–99% hatchery while the actual run is 52–72% hatchery. Pooling them made every destination estimate quietly a hatchery estimate. Wild allocation now uses wild sources and the dam fin-clip census.
• New: sub-basin destinations: Lochsa, Selway, Lemhi, Pahsimeroi and 18 more tributaries, from the separate hatchery/wild allocation.
• Confidence intervals repaired: some intervals didn't contain their own estimate after a late rescaling step; whiskers now wrap every bar.
• Impossible destinations removed: basins that enter the Columbia below Bonneville can't be destinations for fish counted at Bonneville.
• 2011 shown as unresolved: only 63 usable tags exist for that migration year, which cannot support per-river estimates. The bar now says so instead of pretending.
• Two data clocks: dam counts update daily (automated); PIT detection data runs through the date in the header: it refreshes with each model rebuild, not nightly, and this line states the real date rather than pretending otherwise.