Got it, here’s the reply draft to that message, without em dashes:

Great framing on the free flow / slowdown factor decomposition. That’s actually the cleanest way to think about it conceptually, and it maps well onto how you’d implement this technically.

The two approaches you identified are both viable, but they differ significantly in pipeline fit.

Option A: Tag the OSM source, update the Lua profile
You add a custom tag (e.g. avg_speed:car:freeflow=52) to each way and modify car.lua to prefer it over the normal speed logic. This works with both the CH and MLD pipelines, is straightforward to reason about, and keeps everything self-contained in the extract step. The downside is that conflating your speed dataset onto OSM ways is where most of the pain lives. Matching your road segments to OSM way IDs, handling directionality, splits, etc. gets messy fast. OSM ways get split at every intersection during extract, so a single way ID in your speed data may correspond to multiple internal edges.

Option B: Segment speed CSV via osrm-customize
This is the approach that was actually built and it’s part of the MLD pipeline. After osrm-extract and osrm-partition, osrm-customize can accept a CSV of from_node_id,to_node_id,speed pairs that override the extracted speeds for individual directed edges. This is OSRM’s intended traffic integration point. The catch: you need OSM node IDs as your join key, not way IDs, which makes the conflation problem slightly different but not simpler.

For a once-a-year or once-a-month batch run, Option B with MLD is probably the right architecture. You extract and partition once, then re-run osrm-customize with each speed vintage (freeflow, AM peak, PM peak) in minutes rather than re-extracting the whole graph.

On the free flow framing specifically
This is worth leaning into. If you build your freeflow model as the baseline, essentially “what’s the travel time if the road network behaves as designed, at legal speeds, with no congestion?”, you get a clean policy-relevant metric: structural access, independent of traffic. Your rush-hour model then gives you the traffic penalty as a ratio per OD pair. That ratio is interpretable, transferable across years (road networks change slowly, traffic patterns less so), and lets you separate infrastructure adequacy from congestion as distinct policy levers.
The conflation step is unavoidable either way. The central question is what identifier your speed dataset uses. If it’s road segment IDs from a national dataset (NDW? WEGGEG?), you’ll need a geometric or attribute-based match to OSM. If you already have speeds keyed to OSM node pairs, Option B is essentially plug-and-play.

What does your speed data look like structurally, and what’s the primary join key?​​​​​​​​​​​​​​​​