Start a new study
Choose an exact grid or a seeded sample. Compare design choices within each traffic and vehicle scenario. Keep the baseline, seed lists, parameter ranges and valuation assumptions attached to every result.
Open the workbench ↗RESEARCH CHECKPOINT · 3 NOVEMBER 2026
The experiment comes before the conclusion.
The new simulator exposes all 46 physical, behavioral and observation inputs, plus three valuation inputs. It can search settings for any of eight designs, keep sensitivity cases separate, and test each frozen screening choice on fresh seeds. The earlier timing study below is a preliminary result from a narrower question.
We have not identified Michigan’s least costly intersection or a universally best Michigan-left setup. The model prices fuel and vehicle time. It does not optimize crash risk, construction, maintenance, land, pedestrians or network spillback.
Download the protocol
These editable JSON files run in the browser or with the included Node command. The broad preset samples four of 86,093,442 assumption combinations; it does not exhaust that domain.
Timing preset · Broad sensitivity preset · Small verification preset
What a complete experiment records
- Demand and boundaries: vehicle arrivals per approach, intended destinations, observation windows and fixed trip endpoints within each scenario.
- Intersection choices: cycles, green allocation, crossover distance, control, offsets and turn-on-red settings; roundabout radius and protected-left allocation apply only to relevant designs.
- Assumptions: headways, gap acceptance, clearance intervals, speeds, turning geometry, fuel use and storage diagnostics. These produce separate sensitivity cases and cannot win by changing the conditions of comparison.
- Coverage: explicit values, exact or sampled mode, actual distinct candidates, sampling seed, invalid combinations and skipped inactive settings. Sparse grids do not search every real number between values.
- Selection and validation: the lowest screening operating cost among complete, storage-screened settings is frozen for each design/scenario/valuation. Baseline and selected settings then use disjoint validation seeds. Validation does not choose a new winner.
- Cost accounting: full-trip fuel × fuel price + full-trip seconds × vehicle-hour value / 3,600. Queue time and restarts are already included. Annual all-traffic differences scale by annual intended lefts × observed total-entrant / intended-left ratio.
Eligibility requires all observation-period entrants to finish, positive exposure, and no sampled queue-storage exceedance in any screening replication. Storage is checked every ten seconds using approximate lane capacities. Passing is not proof of real-site feasibility. Invalid parameter combinations remain in the audit. Cancelling preserves completed candidate records but suppresses a validated conclusion. The browser limits plans to 10,000 effective design/scenario candidates and 100,000 screening simulations per batch; large experiments can be split into explicitly recorded batches.
Preliminary timing study: the result and its limits
Before the expanded build, 1,710 Michigan-left settings were tested with 12 shared screening seeds. Of these, 1,679 passed the completion and storage screen. The selected setting was frozen, then tested on 100 fresh seeds. The baseline used a 90-second cycle, 50% E–W usable green, 660-ft crossover spacing, yield control and both red-turn options off.
| Quantity | Baseline | Frozen screening choice |
|---|---|---|
| Cycle / E–W usable green | 90 s / 50% | 60 s / 60% |
| Crossover distance / control | 660 ft / yield | 560 ft / yield |
| Right on red / crossover left on red | Off / inactive | On / inactive |
| Validation operating cost / 1,000 external trips | $778.44 | $734.85 |
| Paired cost difference ± 95% seed interval | $-43.59 ± 1.07 per 1,000 trips | |
Conditions: 300 vehicles/hour/approach on each axis; 20% intended lefts, 10% rights; 35 mph; 30 mpg; $3.50/US gal and $20/vehicle-hour, assumed 2024 dollars. Full-trip time fell by 7.51 seconds per external trip. All measured entrants completed and no storage exceedance was observed in the 100 validation replications.
The choice hit three searched boundaries: shortest cycle, shortest spacing and largest E–W green share. It is a best-tested setting in that grid, not a continuous optimum. Its right-on-red benefit is conditional on modeled gaps and the absence of pedestrian conflicts. Three stress cases flagged storage; a minor-road-heavy case made this setting more costly than its baseline.
| Fresh stress case (30 seeds) | Paired Δ USD / 1,000 trips ± 95% | Selected setting diagnostic |
|---|---|---|
| light | -47.33 ± 2.50 | No observed exceedance |
| balanced400 | -43.02 ± 1.86 | Storage/completion flagged |
| arterialHeavy | -75.90 ± 2.77 | Storage/completion flagged |
| minorHeavy | +25.02 ± 5.61 | No observed exceedance |
| left10 | -33.42 ± 1.68 | No observed exceedance |
| left35 | -62.40 ± 2.06 | Storage/completion flagged |
| finerStep | -45.57 ± 1.91 | No observed exceedance |
| longerLoading | -43.39 ± 1.29 | No observed exceedance |
The stress tests hold the selected setting fixed; they do not reoptimize for each traffic pattern. At the assumed 550 million intended-left maneuvers/year, the balanced base scenario scales to approximately $119.6 million/year less operating cost than this Michigan-left baseline (seed interval half-width $3.1 million). That is a conditional extrapolation from one scenario, not a statewide estimate or a comparison with standard left turns. The 200 million and 1.1 billion sensitivity volumes scale this arithmetic proportionately.
The eight-design comparison bundled with that study used common settings and 30 seeds, without equal optimization of every design. It is descriptive, not an engineering ranking. A separate 30-seed check with the original two-layout engine agreed with the direction of the base-case operating difference; its outputs are kept separate.
Preliminary report JSON · All 1,710 screening rows CSV · Replication records and stress tests
Source audit
| Source | What it supports | What it does not establish |
|---|---|---|
| MDOT GEO-670-E, revised 2022 | The preliminary 560–760-ft spacing domain follows the cited 660 ± 100-ft placement guidance. | A universal optimum, permission at a particular site, or capacity of every crossover. |
| FHWA Signalized Intersections guide, archived 2004, chapter 10 | Median U-turn geometry and operational context. | Current measured statewide volumes or a crash response for every parameter. |
| FHWA timing manual, archived 2008, chapter 6 | Coordination and offset concepts; travel time and phasing matter. | A universally correct −7-second offset. This is an archived reference, not current design authority. |
| FHWA-SA-25-021, October 2025, Table 1 | National per-crash economic and comprehensive values in 2024 dollars, by maximum KABCO severity. | Design-specific crash rates, severity distributions or a cost-minimizing signal plan. Comprehensive costs already include economic costs. |
| User-supplied Michigan scaling assumptions | 550M intended Michigan-left maneuvers/year; 200M–1.1B sensitivity; approximately 1,000 crossovers. | An official current inventory or measured statewide turn count. |
Preparation for the Popular Rationalism series
This is a research outline, not a drafted or published article. Preserve the frozen protocol and complete data before choosing a narrative.
- Follow the whole maneuver. Explain the extra distance, U-turn, return signal, two-stop cases and fuel ledger. Show matched drivers and common endpoints.
- What changes when we change the intersection? Compare all eight layouts under the same demand. Show where operating assumptions and storage limits change the comparison.
- Searching without choosing the answer. Explain exact versus sampled grids, separate assumption scenarios, common seeds, frozen choices and untouched validation. Report reversals and failed cases alongside improvements.
- What does a crash cost—and can this model predict one? Separate crash frequency, crash type, maximum severity and valuation. Distinguish assumed event rates from a calibrated safety prediction.
- From one intersection to Michigan. Trace the annualization denominator and expose the uncertainty in crossover inventory, turn volume and the mix of traffic conditions. Do not label a conditional operating subtotal “total savings.”
Before making policy claims, obtain observed approach and crossover counts, site-specific signal plans and turn restrictions, queue observations, consistent whole-system crash exposure and severity data, and construction/maintenance estimates. Define how site types and hours are weighted statewide. The expanded simulator makes these assumptions testable; it does not supply the missing measurements.
Run it independently
node experiment-tests.cjs node tests.cjs node run-experiment.cjs complete-protocol.json my-result.json # Or create the preset directly: node run-experiment.cjs --preset complete my-result.json
The runner writes a protocol file before simulation. Browser results and Node results share the same engine. Open experiments.html locally for the offline UI; it falls back to main-thread execution where local workers are unavailable. Keep the tab open until the run finishes and export results before leaving.