Churchill Corridor Atlas

Drag along the route — the seabed under the keel

Every 2 km from the Port of Churchill to the open Atlantic: measured depth in blue, model-inferred depth in amber with its calibrated σ band, and the water nobody can answer for yet in red. Hover for the provenance of each point: the CHS survey year and CATZOC grade under the keel, or the distance to the nearest sounding. Of the 198 points with a published sounding, 103 (52%) come from surveys before 1980. Trust grade along the route: A 17% · B 27% · C 34% · D 22%.

Two independent tests, all the numbers, including the ones that hurt

A completion model is worth nothing until it is scored on depths it never trained on. We ran two such tests and publish both in full.

Baselines a hydrographer would actually use. Copying the nearest sounding is a floor, not a method. The classical gap fill is a gravity trend plus interpolated residuals; we report it two ways (Delaunay natural-neighbour and inverse-distance-squared on the 12 nearest residuals). On Test 1 the model beats both (4.9 vs 11.2 / 8.7 m); on Test 2, 13.3 vs 16.8 / 16.2 m. *GEBCO (NCEI global mosaic, 15 arc-second) is shown as the chart-world reference, not a baseline: it ingests NONNA-100 through IBCAO v5 (the Arctic Seabed 2030 compilation) and the same NCEI cruise archive, so it has seen both test sets; on the 100 m grid it scores 3.0 m (Test 1) and 6.2 m (Test 2) — lower than the model, and that is the point: those are the residuals of resampling data GEBCO already contains, not predictions. Where GEBCO has no source soundings it is SRTM15+, the gravity row above. Gravity leakage, measured. SRTM15+ V2.7 (April 2025) inherits the cumulative NCEI multibeam archive of its point releases, which is where the Test 1 cruises live. Its error on the Test 1 cells is 5.5 m against 15.2 m on 282,193 ordinary NONNA-sounded shelf cells in the same blocks: the prior has almost certainly seen the cruise data. That leakage favours the baseline, which the model still edges on Test 1; the model itself takes gravity as an input and inherits some of it. Test 2 is the clean comparison: NONNA shelf soundings SRTM15+ does not resolve, where gravity scores 18.1 m against the model’s 13.3 m.

What the tests do not support. Under 50 m of water the pre-2016 model reads 6–9 m too deep — the dangerous direction, and the reason the mean depth is the wrong target for navigation (see the hazard exhibit). Within 500 m of an old sounding, copying that sounding still beats the model (5.6 vs 7.1 m); the model earns its keep beyond that. Test 2 is the full-size (34.8M-parameter) model, trained 17 minutes on a rented RTX 5090; the 6.8M model scored 13.5 m on the same test, so size is not what limits this. Coastal cells under 50 m in Test 1 disagree with every source including CHS’s own nearest sounding and are excluded pending a look at the GMRT coastline product.

Ships do not ground on the mean depth. This predicts the shallowest point.

Every bathymetry model, including the one above, predicts the average depth of a 100 m cell. A keel meets the shallowest rock in it. Where CHS holds both 10 m and 100 m data, the shallowest point within 500 m sits a median 6 m above the 100 m mean on this route. So we trained a second, 34.8M-parameter model on 475 NONNA-10/100 tile pairs to predict that shallowest point and its uncertainty, and turned it into one number per cell: the probability that a 10.5 m draft touches.

P(shallowest point within 500 m < 10.5 m), every cell. Black is safe, yellow is a coin toss. Blue: the route. Red rings: the 26 route-km the mean map would have called safe.

The groundings that already happened, hindcast

For each TSB-documented Arctic grounding we gave the hazard model only the soundings that existed before the ship hit (dated by CHS’s Survey Index) and asked what percentile of danger the strike site got among the water that looked safe on the mean map within 25 km. On 4 of 6 scored groundings the hazard field placed the strike site in the top 10% of danger among water that looked safe on the mean map (Thamesborg 96th, Akademik Ioffe 93rd, Hanseatic 99th, Nanny 2012 95th). It missed Clipper Adventurer (84th), Nanny 2014 (79th). Baselines: the mean model’s own σ (“where it is blind”) and the shallowness of the nearest old sounding.

The same hindcast, blind, and how much the denominator matters

The table above still lets the model see every published sounding around a site at inference time, and NONNA carries no dates after 2016, so a post-grounding survey could be in that input. Here every sounding within 10 km of each site is removed from the input as well (the hazard model was trained with every tile group within 0.4° of the sites excluded from the start), and a 25 km hole is shown too: at 25 km the site falls outside the model’s inference envelope and it returns no estimate, which is the right answer, not a miss. Then the two choices that set the percentile — the comparison window and the definition of “looked safe” — are varied. Thamesborg with a 10 km hole in its input: 93rd percentile, nearest remaining sounding 10.0 km away, absolute P(shoal < draft) 1.4%.

Skill, not recall. By construction 10% of apparently-safe cells sit above the 90th percentile, so the base rate is known exactly. With the 10 km hole, 4 of 6 sites land above it (binomial p = 0.0013); on the four uncharted-shoal groundings alone, 3 of 4 (p = 0.0037). With all soundings visible: 4 of 6 (p = 0.0013). The precision side of the ledger is stated too: corridor-wide, 18% of cells exceed P = 5%, and on the route 26 km of 1762 apparently-safe km are flagged. Six incidents cannot bound a false-alarm rate; what the flags buy is a survey order, and the survey is what tests them.

Thamesborg (table above) uses all published soundings because NONNA carries no dates after 2016; its position is from AIS and press, not TSB. Akademik Ioffe and Clipper Adventurer sit 9–12 km from any pre-incident sounding, at the edge of what any completion should be trusted for; their scores are reported anyway. The two Nanny groundings were navigation errors in known narrows and are included as controls, not claims. Full per-site JSON in the validation repo.

The channel no one drew — computed from depth and uncertainty

Given only the completed seabed and its σ field, the model traced the least-risk channel from Churchill to the Atlantic itself: 2,782 km, never shallower than 22 m, mean σ of 2.7 m — and not one kilometre over unanswered water. Where the straight route crosses red, the found channel goes around. Read it as a map of where the risk is, not as a sailing direction: the channel is 455 km longer than the straight route, about 20 hours at 12 knots, roughly 21 t of fuel and one day of hire per voyage (a Handysize at ~25 t/day; ~US$25k). Nobody should pay that on every voyage; the point is that the water the straight route crosses has never been answered for, and that is what the survey plan below is priced to fix.

Cyan: the computed channel. Dashed grey: the straight-line route. Cost per km = distance × (shallow-water penalty + σ/8 + no-data ×13); water under 20 m and land impassable.

The corridor, completed — and honest about which pixel is which

82,121 km² with published soundings → +351,305 km² completed, every inferred pixel within 6 km of a real sounding. Blue is measured; amber is model. No pixel pretends to be the other.

Where the corridor is still uncertain

The insurers’ map. Bright veins are the highest-σ water — and the route threads straight through them in the Hudson Strait approaches. Uncertainty, not ice, is what limits the shipping season; this is the first map of exactly where it lives.

Tasmania Islands, Franklin Strait — September 6, 2025

The MV Thamesborg, an ice-class freighter bound from China to Baie-Comeau, struck what the Transportation Safety Board calls an “uncharted shoal” (case M25C0241) at the western tip of the Tasmania Islands. Salvage took a month and an icebreaker. In July 2026 the Dutch Maritime Disciplinary Court suspended the Master’s licence for deliberately routing through a zone whose survey reliability was rated CATZOC C — poorly surveyed water — when a better-charted lane lay further west. In the published archive, 55% of this water has no sounding. The ship struck 1.7 km past the edge of the surveyed swath — and the model, which had never seen that seabed, ranks the strike site in the 98th percentile of uncertainty for the whole strait (calibrated σ ≥ 14 m at 68%, ≥ 20 m at 95%). CHS’s own Survey Index shows the nearest dated survey to the site is a 1960 single-beam line graded CATZOC C; the modern multibeam swath ends 1.7 km short of the strike. The right panel is the same water, completed.

The model cannot see a shoal nobody measured — but it knows where it is blind. In the left panel the surveyed swath passes west of the islands, the route the court said the ship should have taken; the strike is at its margin. Right: the white haze of highest uncertainty sits exactly there. Grounding position from AIS/press reports (TSB has not published coordinates).

Where a survey ship earns its day rate first

The σ map, converted to procurement: the ten highest-uncertainty boxes within 35 km of the route, priced in ship-days. Coverage assumption: 40 km²/day, from a multibeam swath of ~3× water depth (300–400 m at 100–130 m), 8 knots for 20 hours (~300 line-km/day, ~100 km² raw), 20% line overlap and a 50% weather-and-ice downtime factor across an Arctic season. Two public contracts bracket the cost. Lower bound: CHS Arctic hydrographic contract to Sedna ROV Services (Iqaluit), C$6.5M over two seasons, 9,000+ line-km in the first (DFO, March 2024): ~C$720 per line-km; at a 350 m swath with 20% overlap, ~C$2,600 per km², which prices the plan’s 8,753 km² at about C$23M. Upper bound: the Coast Guard’s July 2026 polar-icebreaker charter at C$183,000/day (C$22M for ~120 days), which the table uses. For scale, CCGS Amundsen 2021: ~38,700 km² mapped over an ~80-day season (Copernicus ESSD 2024), i.e. several hundred km²/day in transit mapping of deep water; 40 km²/day is conservative for 100–130 m shelf water.

Total: ~219 ship-days for 8,753 km²: C$23M at the CHS contract rate, C$40.1M at the icebreaker charter rate to de-risk the worst water on the route — a fundable season, against “more than a decade” for full conventional coverage.

1,314 predicted depths, sealed on 2026-09-02 — with a clock we do not control

Every cell in this file has no published sounding today. 314 sit on the found channel; the rest are drawn at random across the corridor. Each carries a predicted depth and calibrated 68% / 95% bands. The next CHS, Coast Guard or research survey through this water scores it, and we publish the score, whichever way it goes.

Not just the corridor. Canada.

The same model, run over the entire national archive: all 437 NONNA blocks — three oceans and the Great Lakes. 429,682 km² of surveyed seabed extended by +1,088,182 km² of completed seafloor, every pixel carrying its own σ. This is the asset the corridor exhibits are cut from.

Every charted Canadian water, completed. Blue: the CHS archive. Amber: SeabedNet v5, within 6 km of soundings.

And the national uncertainty field — the survey-priority map for the whole country, the same way Exhibit E prices the corridor.

What we claim, and what we don’t