How TollChat estimates commute costs

Where do the toll prices come from?

Prices for dynamic I-66 Inside the Beltway and 95/395/495 Express Lanes come from VDOT's public feeds. For the Dulles Toll Road and Dulles Greenway, TollChat uses reviewed 2-axle E-ZPass schedules reconstructed from rates published by the operators. TollChat is independent and is not affiliated with VDOT or any toll operator.

How are current prices different from annual ballparks?

A current-price request uses the latest supported observation or a published fixed rate for one route. An annual ballpark uses recent, complete round trips from the same date to calculate daily P25, P50, and P90 scenarios. It then multiplies each daily value by the number of commute days you choose. These are historical scenarios. They do not predict next year's tolls.

How should I use the take-home-pay estimate?

TollChat's take-home-pay figures are best-effort estimates, not calculations of your actual pay. They assume one-third of gross income goes to taxes and apply $0.685 per straight-line mile on priced toll segments as a fixed vehicle-cost allowance. They exclude untolled parts of the commute and do not account for your actual taxes, deductions, vehicle, or other expenses.

Use the figures as a starting point, not for financial planning, salary or job negotiations, tax filing, or legal decisions. TollChat does not provide financial, tax, or legal advice.

What do the four map pins mean?

The main number on each pin is the P50 annual toll for a 2-axle E-ZPass commute to Washington. It assumes travel Monday through Friday, leaving at 8:30 AM and returning at 5:30 PM for 240 commute days. The detail panel shows P25 and P90, how much evidence was available, and when the snapshot was generated. The figures exclude costs from untolled roads. They are not forecasts or quotes from toll operators.

How does TollChat keep the model on supported roads?

TollChat matches ordinary place names to a limited set of route points. A read-only route oracle checks the direction of travel and returns fixed pricing keys. The pricing tools retrieve and add the supported toll components; the model does not. If a route is unsupported, TollChat says the price is unavailable instead of making one up.

Can a supported trip still be incomplete?

Yes. TollChat shows missing observations during active tolling, partial historical coverage, modeled data, and unpriced general-purpose connections. I-66 Inside the Beltway is the exception outside its published tolling windows, when the authoritative price is $0 rather than a missing price.

The route oracle uses 330 origin-destination (OD) pairs for I-95 and I-495, but VDOT history has prices for only 314. The missing 16 appear on 107 of the oracle's 685 routes, mostly where a trip crosses from the I-495 Express Lanes onto I-95/395. For each one, TollChat copies the price from the observed VDOT OD pair that matched it most closely and labels the result as modeled.

In 1,200 held-out comparisons from five days of matched operator data, this direct proxy method had a mean absolute error of $0.106, compared with $0.154 for ordinary least-squares regression. The test period is too short to show seasonal accuracy, and the largest error was $8.05, so these modeled prices are ballpark estimates rather than operator quotes.

Why doesn't TollChat cover I-66 Outside the Beltway?

VDOT's public feeds do not include the I-66 Express Lanes Outside the Beltway. Without a public feed or historical dataset it can validate, TollChat cannot provide reliable current prices or annual ballparks, so it leaves that facility unsupported instead of guessing.

Does TollChat hallucinate?

Yes. TollChat can make mistakes, but we observed no fabricated costs in this 1,000-answer test. Pricing tools calculate its toll and take-home-pay figures in code. GPT-5.6 Luna handles the conversation and explains those results.

In one frozen annual-ballpark evaluation for Springfield-Franconia to Westpark, 996 of 1,000 responses (99.6%) made no quantitative claim outside the supplied tool evidence under the strict policy.

One response changed the supplied 5:31 PM time to 1:31 PM. Three responses correctly calculated differences between scenarios, but those calculations were not in the payload. The arithmetic was correct, but it still violated the strict evidence policy. Core semantic compliance was 97.1%. Adding the requirement to attribute the answer explicitly to TollChat and applying the strict quantitative policy produced the conservative 93.1% end-to-end result.

This repeated-generation test measures reliability for one production-shaped fixture. It does not measure TollChat's overall hallucination rate, route coverage, source accuracy, or production traffic. Read the complete report and methodology.

Why is TollChat free?

TollChat is free because production AI experience is comically hard to get without already having production AI experience. It's the industry's favorite little paradox. TollChat was born to break that loop: you get a free commute-cost experiment, and we get to find out what happens when an AI project meets real users instead of another demo. It may work. It may not. Either way, that's production experience.

What does TollChat collect?

New public usage counting and daily publication have stopped. Historical aggregate and snapshot data remain retained on AWS and are not being purged in this release; they do not include chat text, trip details, salary figures, IP addresses, or advertising identifiers.

The privacy notice explains the retained historical data separately from the secure session cookie used to keep a chat together for up to one hour.

TollChat keeps the active conversation in the ephemeral microVM's memory. We make an effort to protect your personal information by automatically redacting detected PII from diagnostic traces before storage. Automated redaction can miss information. Please avoid sharing sensitive personal details. Diagnostic traces may include redacted conversations and tool results; trace archives expire after seven days. OpenAI receives prompts, prior conversation context, and responses with Responses storage disabled. OpenAI still keeps abuse-monitoring logs, which may include that content, for up to 30 days by default. We accepted that tradeoff to keep TollChat free to use. Read how OpenAI handles API data. A random credential in a secure, HTTP-only browser cookie keeps public follow-up messages together for up to one hour. TollChat stores only its one-way hash and does not attach the credential to traces or logs.

The map loads tiles from OpenFreeMap, which receives the requests needed to provide the map. Read the privacy notice and terms for details. The project is open source under Apache-2.0. Please report bugs or privacy questions to contact@tollchat.ai.