The reported hypothesis

The project notes place the experiment on August 25, 2026. The proposed shortcut was appealing: take geographic bus and jeepney lines, force them onto 0°, 45°, and 90° segments, and let an optimizer produce a readable transit diagram without a maintainer correcting every route.

The missing piece was not another angle penalty. A transit diagram has to preserve things a geographic simplifier cannot infer safely: which direction a line serves, which parallel line is distinct, which stop belongs to which route, and which crossing is a transfer rather than a visual intersection.

What the record says happened

The historical notes report failures around Toril, San Pedro and Pichon, and parallel routes sharing a corridor. They describe distorted loops, ambiguous inbound and outbound order, false crossings, and crowded terminal areas. Those cases are useful leads for reconstruction, but this checkout contains no optimizer source, generated diagram, fixture, screenshot, or test output that can verify the exact run.

The current route files are still geographic inputs: bus-routes.geojson and jeepney-routes.geojson. They do not prove that a schematic output was generated from them.

Why the failure was more than a bad drawing

Suppose two jeepney services share most of an avenue and separate near a terminal. A geometry-only optimizer can reduce both paths to the same set of diagonal segments. The picture becomes tidy, but the rider can no longer tell which service continues to which destination. If the optimizer then moves a stop to keep lines apart, it may also imply a transfer that the geographic network never accepted.

That is the important distinction. The objective was not to preserve every meter of road geometry. It was to simplify the drawing without changing the service relationships that the planner and rider rely on. A mathematically clean line can still be wrong as transit evidence.

The correction was a change in tool responsibility

The proposed correction was an automatic first pass followed by human editing and validation. The first pass can suggest bends and spacing. A maintainer must decide whether a shared corridor should remain visually shared, whether a loop's direction remains legible, and whether a terminal's boarding relationships survive the rearrangement.

That workflow remains a research direction, not a current production pipeline. The companion experiment records the available evidence and the missing artifacts. Loom WASM is an external project that may inform future tooling. No current Davao Transit import, dependency, or deployment path proves that Loom is integrated.

Before and after output is unavailable

There is no verified before/after image for Toril, San Pedro, Pichon, or the parallel-route case. The only defensible comparison today is geographic input versus an unobserved schematic output. Publishing a made-up “failed layout” would turn a useful historical report into fabricated evidence.

If the experiment is rerun, save the exact input revision, generated output, route and stop identifiers, and a screenshot at a fixed viewport. Annotate the failure in the image, then show the manually corrected version beside it. Preserve the case even if the correction is imperfect. The point of this postmortem is to show where automatic geometry stops being a transit decision.