Sep. 202610 min read

Your Canadian truck already drove that lane empty. Northbound finds the freight it could legally have carried

10,479 completed legs, 1,596,093 miles, two months of Roadstar Trucking's real dispatch data. Coming home empty happened six times. The real empty running is inside the United States, between loads, and whether it could have carried freight turns out to be a legal question, not a routing one. 19 CFR 123.14(c)(1) permits US point-to-point carriage when it is part of the return of the vehicle to its base country, so the verdict turns on which way the load runs. 217 legs, 66,702 miles, $155,815 at ATRI's 2025 industry cost, already travelling toward the border and legally able to carry. GLM 5.2 on SPUR reads the offer, a deterministic TypeScript engine cites the statute. Ablation on 100 real offers in five broker formats: 100 percent correct verdict with the model, 40 percent without.

RoadStarHackathonGLM 5.2SPUR ComputeCabotageCross-border trucking19 CFR 123.14TypeScriptRegulatory AgentFreightHackathonShowDev

Watch the demo Source code

I wrote this post for the RoadStar Hackathon 2026, built Northbound on Roadstar Trucking's own dispatch data. Source at github.com/JonathanSolvesProblems/northbound, open with no key and no account. #roadstarhackathon

The claim in one sentence: on 10,479 real legs from Roadstar's own two-month history, 217 empty trips (66,702 miles) were already travelling toward the border and legally allowed to carry US point-to-point freight under 19 CFR 123.14(c)(1). That is $155,815 at ATRI's published 2025 industry cost of $2.336 per mile. None of the inputs are mine: the legs are theirs, the rate is ATRI's, the statute is the CFR's.

Northbound load board: a Canadian truck's next possible haul, coloured by whether US cabotage law permits it under 19 CFR 123.14(c)(1). Green international, amber heading home, red moving away. — A load board that knows which US freight a Canadian truck is legally allowed to haul. Built on Roadstar's own dispatch history, not on a synthetic market.

Their data corrected my thesis

I went in assuming the problem was coming home empty from the Midwest. Roadstar's data said no: across 10,479 completed legs and 1,596,093 miles, six empty legs home in two months. 222,635 miles ran empty overall (13.9 percent), and the biggest bucket is repositioning inside the United States: 778 legs, 126,874 miles. Whether any of those legs could have carried freight is a legal question rather than a routing one, because a Canadian-domiciled truck may not haul US point-to-point freight unless a specific carve-out applies. That is the whole engine.

Where the empty miles really are in Roadstar's two-month history: 6 legs coming home empty, 778 legs repositioning inside the US, 2,847 inside Canada, 92 crossing south empty. — Where the empty miles really are. In 10,479 real legs, coming home empty happened six times. Repositioning inside the US: 778. That is the pool the legal engine works on.

Cabotage is directional

Almost every distance-based optimiser treats US cabotage as a flat prohibition. It is not. 19 CFR 123.14(c)(1) permits carriage between US points where it is incidental to international traffic, and explicitly says a US point-to-point move counts as incidental when it is 'as part of the return of the vehicle to its base country.' So the customs test turns on which way the load is going. Columbus to Brampton is green (international). Columbus to Atlanta is red (moving away from home base, and Northbound names the exposure: section 592 of the Tariff Act and the driver's B-1 admission). Columbus to Detroit is amber (US domestic, but headed home, so it clears with both citations attached).

The regulation cited verbatim: 19 CFR 123.14(c)(1) allows US point-to-point freight as part of the return of the vehicle to its base country. — The rule the whole engine turns on, verified against two independent primary sources (law.cornell.edu and govinfo) and tabulated in LEGAL.md. The verdict turns on direction, not distance.

The headline number

217 empty legs, 66,702 miles, already travelling toward the border and legally allowed to carry US point-to-point freight. At ATRI's published 2025 industry cost of $2.336 per mile, that is $155,815. It repeats, which means it is fixable with standing agreements rather than luck: Fairburn GA to Walton KY 10 times, Morris IL to Richmond IN 13 times, Scottsville KY to Walton KY 11 times. And 120 of those empty legs ran within 150 km and 48 hours of one of Roadstar's own loads going to Ontario. Fairburn is twenty miles from Atlanta. Same company, two trucks, one of them empty and legally able to carry, one of them loaded to Etobicoke.

Northbound's headline card: 217 empty legs, 66,702 miles, already heading for the border and legally able to carry freight. $155,815 at ATRI's 2025 industry cost. — This is not a projection. Every leg happened, the miles are theirs, and the cost rate is ATRI's public 2025 number. Northbound counted what was already legal and already paid for in fuel.

How a verdict is decided

A dispatcher pastes an offer the way it actually arrived, as a broker's email or a one-line chat message. GLM 5.2 on SPUR Compute reads it into structured fields (origin, destination, commodity, weight, pickup and delivery windows). That is the surface the dispatcher touches, and it is the part regex is bad at and a language model is good at. Then the law decides in TypeScript, against the cited regulation, because a model should never be the thing asserting what a statute permits.

An amber verdict from Fairburn GA to Walton KY: US point-to-point but heading home, cleared with 19 CFR 123.14(c)(1) attached. — Paste the offer as it arrived. GLM 5.2 on SPUR reads it, the regulation decides. Fairburn GA to Walton KY renders amber because it is heading home.

A red refusal from Columbus OH to Atlanta GA: US point-to-point moving away from home base, refused with section 592 of the Tariff Act and the driver's B-1 admission cited. — The refusal. Columbus OH to Atlanta GA moves away from base, so it is red. Northbound names the exposure by regulation, not just by feel.

Plans the clock across the border

The same driver has different hours on each side of the border. Driver 84 in Roadstar's roster shows 53 hours on the US cycle and 11 the moment the truck crosses at Bluewater. Border wait is priced live from CBP's public API, commercial and FAST lanes separately, and a wait over 30 minutes is credited as the break the driver already owed. The board renders 34 real trucks against two rulebooks and live crossing waits.

The dispatcher board: 34 real trucks, two rulebooks (US 49 CFR 395.3 and Canadian SOR/2005-313), live CBP border waits, hours-remaining per side. — 34 real trucks, two rulebooks, live CBP waits. Driver 84 has 53 hours on the US cycle and 11 the moment he crosses at Bluewater.

Measured what the sponsor's model contributes

Every entrant at a sponsored hackathon asserts the sponsor's model is essential. Northbound measured it. Same 100 real orders, rendered in five broker formats (email, chat message, structured web form, PDF quote, phone-transcript style), read twice. With GLM 5.2 on SPUR in the code path, 100 percent reach the correct legal verdict. With the model switched off and a fair rules parser standing in, 40 percent. That gap is not a claim about model quality in general, it is a number about what happens on messy freight offers a dispatcher would actually receive. The ablation is recorded to docs/ablation-n100.txt and locked by scripts/check_claims.py so it cannot drift silently.

Ablation card: same 100 real offers in five broker formats, read twice. GLM 5.2 on: 100 percent correct legal verdict. Model off: 40 percent. — Ablation, not assertion. The gap between 100 percent and 40 percent is what GLM 5.2 on SPUR contributes to the verdict on messy freight offers, measured by a script the sponsor can run.

What is not claimed

$155,815 is a floor on what was already legal, not a projection of what will be captured. It measures what Roadstar's own two months of empty running could have earned if the dispatcher had had the verdict at offer time. It does not model any change in behaviour, any lost margin from repositioning contracts, or any operational cost of chasing amber legs. Every number in the README is recomputed from the raw sheets by scripts/check_claims.py, which fails the build if the prose and the data disagree. It caught four mistakes in my own writeup the first time it ran. Of the 31 tests, one exists only to lock the headline, so a change to the geocoder cannot quietly move the money figure.

Related project

Northbound: your Canadian truck already drove that lane empty, this finds the freight it could legally have carried

View the project

A camera assistant wears smart glasses through a shoot and walks off set with the continuity paperwork already written

The last job on a film shoot still done entirely on paper is continuity. Miss a line, find out at the edit, pay for a pickup day at $1,440 to $3,020 (Giggster 2026). Dailies watches a take through Ray-Ban Meta Gen 2 glasses for 1.6 cents at Google's published Gemini rates, 63 cents a forty-take day, 0.04% of the day it is guarding. Gemini reads every take, ClickHouse Cloud holds the observations, and a plain-English question box lets an agent write its own SQL through the official mcp-clickhouse server. The agent cannot destroy data: three independent layers, and the one that matters is a readonly cluster credential refusing DROP, INSERT, TRUNCATE, CREATE TABLE and ALTER with ClickHouse code 497. Median rolling verdict is 4.4 seconds, spoken back into the wearer's ear. The catch in the 2 min 31 s demo lands at 0:49.

Sep. 202610 min read