Chill 45-minute playlist
A track sequence from a music library, paced for the listening window and the mood.
corpus: librarycost: secondsbudget: 45m
engine · v0An engine of Project Flânerie
Saunteur.
An engine for finding your way.
Saunteur is the engine underneath Urban Gallery. Given a collection of things worth attending to, a way to compare them, a cost between them, and a budget, it produces an ordered itinerary that maximizes total value within that budget.
01 · The shape of the engine
Four inputs and an output. The same shape regardless of what’s in the corpus.
01 · corpus
A collection of things
sculpturestracksessaysbreweriescorridors
02 · comparator
A way to compare
value(a, b)relevancefit
03 · cost
A cost between
meterssecondswordsminutes
04 · budget
A budget to spend
2h afternoon45m playlist4-day traverse
→ output
Ordered itinerary
max valuewithin budgetwalk-friendly
02 · Domain tests
If a domain encodes as a corpus and a configuration, the abstraction is real.
Saunteur started with one problem: a public-art walking tour. Items are sculptures; distance is meters between them; budget is two hours of an afternoon; value is “pieces seen.” None of that is built in. We tested the engine against unrelated problems to find out whether the abstraction held, or whether the picker would have to fork.
Focused reading itinerary
An ordered path across a corpus of essays. Same shape; cost is now words; value is fit.
corpus: essayscost: wordsbudget: 90m
Brewery walking tour
A neighborhood loop on foot. Walking minutes, ratings, and which spots you've already visited — all in the cost.
corpus: breweriescost: minutesbudget: 4h
Four-day Highland traverse
A sequence of single-day corridors, with state threaded between days. No multi-day machinery.
corpus: corridorscost: minutesbudget: 4 days
The same algorithm runs all four without a fork.
03 · What the work has taught us so far
Three things we’ve learned, pushing the engine against unfamiliar problems.
findingI
A "tour" is more general than it sounds.
The same algorithmic shape — score against distance, walk down possible tour lengths until one fits, then improve it locally — applies whether the cost is meters of walking, seconds of audio, or attention spent on a paragraph.
findingII
Locality bias matters more in some domains than others.
At walking distances it’s a feature; at driving distances it becomes a bug. The right escape mechanism depends on whether the user wants coherence (a focused reading session) or spread (a multi-neighborhood brewery trip).
findingIII
Multi-day composition needs no special machinery.
A four-day traverse is a sequence of single-day corridors with state threaded between days. The engine doesn’t need to know it’s multi-day.
04 · Honest measurement
Benchmarked against the academic Orienteering Problem literature.
We benchmark Saunteur against the corner of optimization research closest to this kind of problem. It runs near-ceiling on the classical proven-optima suites and shows documented headroom against modern specialized heuristics on the larger OPLib suite.
We run the benchmarks to check our own work: to know whether the engine is sound before pointing it at problems where the optimum isn’t known.
Suite · Result
Tsiligirides 1984
classical, proven optima
near-ceiling
Fischetti 1998
classical, proven optima
near-ceiling
OPLib
modern, specialized heuristics
+ headroom
Detailed methodology & per-instance numbers available on request. The engine is unmodified between suites.
05 · Status
A Back Room Strategies project for now. Not generally available yet.
Saunteur lives inside the firm for now. Urban Gallery’s walking tours run on it in production.
It will be available to others eventually, most likely as integrations built into existing apps for particular industries, and not as a general-purpose library.
Done
Internal engine, benchmarked, deliberate domain tests
Now
Urban Gallery running on Saunteur as its production engine
Next
Selected industry partnerships & engine integrations
06 · Inquiries
Partnership, evaluation, or licensing.
For partnership, technical evaluation, or licensing, write directly. A short note about the problem and the shape of the data gets a reply from Seth.
seth@backroomstrategies.com