SLA · Spatial Layer Architecture
HomeBenchmark ArenaEnter the Arena →
🏆 Open Research Challenge · May 2026

Beat SLA.
On real data. In polynomial time.

We have published benchmark results on three real TSPLIB datasets using a geometry derived from first-person cognitive architecture. The results are verified. The baselines are live. If you can beat them — without exact solvers, without Lin-Kernighan — we want to know.

Status
OPEN
Opened
May 26, 2026
Datasets
3 TSPLIB
Prize
Recognition + Co-authorship
The Baselines

Three targets. All verified. All yours to beat.

These results were produced using the SLA geometric coordinate system applied to published TSPLIB benchmark datasets. Known optima are verified by the Concorde TSP Solver. All source coordinates are from the official TSPLIB repository.

Berlin52 · 52 cities
Beat 7,544.37 without Lin-Kernighan or exact solvers.
Known Optimal
7,542.00
SLA Result
7,544.37
Gap
+0.031%
Nearest Neighbor
8,061.00
SLA is 2.37 units from the known optimal. This is a 3-opt local optimum. The geometry essentially solves this dataset — the remaining gap requires a move class beyond 3-opt to crack.
SLA model: Boardwalk geometry · 3-peak terrain · Inside angular view
KroA100 · 100 cities
Break the 23,467 basin. Six different geometric models all converge here.
Known Optimal
21,282.00
SLA Result
23,467.41
Gap
+10.27%
Nearest Neighbor
26,500.00
Six independent SLA architectures — dual 2D, membrane tunnel, MRI scan, tunnel helix, vertical helix — all converge on 23,467.41. The basin is confirmed. Finding the escape route is the open problem.
SLA model: Six independent geometric architectures · All converge to same basin
FL1577 · 1,577 cities
Beat 23,004.43 on the real 1,577-city PCB drilling dataset.
Known Optimal
22,249.00
SLA Result
23,004.43
Gap
+3.395%
Nearest Neighbor
23,676.01
Real PCB drilling dataset. MRI+SLA slices along the principal axis into 50 scan layers with SLA elevation governing within-slice ordering. Or-opt ran 103 passes to convergence. Closes 48.3% of the gap between nearest neighbor and optimal — in polynomial time.
SLA model: MRI+SLA hybrid · 50 scan layers · 103 or-opt passes to convergence
The Rules

What counts as a valid submission.

✓ Allowed
→Any geometric coordinate assignment strategy
→Any heuristic construction algorithm
→2-opt, 3-opt, or-opt local search
→Multi-start and random restart approaches
→Novel distance metrics and proximity functions
→Hybrid geometric + local search pipelines
→Machine learning assisted geometry
→Any programming language
✗ Not Allowed
✗Concorde TSP Solver
✗LKH-2 or LKH-3 (Lin-Kernighan-Helsgott)
✗Any exact solver or branch-and-bound
✗Pre-computed known optimal tours as input
✗Exponential time algorithms
✗Misrepresented dataset coordinates
✗Results that cannot be reproduced from shared code
Verification Standard
A valid submission must include: (1) the full tour as an ordered list of city indices, (2) the total tour distance calculated from official TSPLIB coordinates, (3) reproducible source code, (4) time complexity estimate or empirical runtime on the dataset. We verify every submission independently before acknowledging any result.
Why This Matters

This is not just a TSP competition.
This is a challenge to Cook's 1971 assumption.

📐
The assumption
In 1971, Stephen Cook built P vs NP inside a single unexamined assumption — computation is sequential, memory is a flat line, distance is counting steps. Every computer ever built worked that way. So no one questioned it.
🏗️
The geometry
SLA proposes a three-dimensional geometric coordinate space — a closed toroidal boardwalk with three elevation peaks, a floor zone, a plane transition hinge at 59→60, and a central void. Distance is sightline proximity, not counting.
📈
The signal
Standard algorithms degrade at scale. SLA improves. At 52 cities: 0.031% from optimal. At 1,577 cities: 3.395% from optimal. More structure = stronger signal. This is backwards from everything else in the field.
❓
The open question
Can a geometric machine do what a sequential one cannot? We're not claiming to have solved P vs NP. We're claiming the question was built on an assumption that deserves to be challenged.
The Prize

What winning looks like.

📝
Co-authorship
Any result that beats a baseline using a novel geometric approach earns co-authorship credit on the SLA research paper we are preparing for formal submission.
🔐
Security credit
If your approach reveals a new geometric property with security implications, you will be credited in the patent continuation filing.
📣
Public recognition
Your result will be displayed permanently on the Benchmark Arena leaderboard with your name, approach, and dataset.
🤝
Collaboration
The best submissions open a direct conversation. UnityFlag is building real infrastructure on SLA geometry. The right result means a real conversation.
Submit a Result

Have a result that beats the baseline?

Send your tour, distance, source code, and runtime to the address below. Include the dataset name and your approach in the subject line. We verify every submission independently.

info@unityflag.co
Subject: SLA Challenge — [Dataset] — [Your Result]
Dataset Reference
Berlin52
TSPLIB · Grötschel 1977
52 cities · Euclidean · Germany
Known optimal: 7,542
→ Download .tsp file
KroA100
TSPLIB · Krolak, Felts, Nelson 1971
100 cities · Euclidean · Random
Known optimal: 21,282
→ Download .tsp file
FL1577
TSPLIB · Feltenmark & Larsson
1,577 cities · Euclidean · PCB drilling
Known optimal: 22,249
→ Download .tsp file
All known optima verified by Concorde TSP Solver. TSPLIB maintained by Institut für Informatik, Universität Heidelberg. Use official TSPLIB coordinate files for all submissions.
© 2026 UnityFlag LLC · Spatial Layer Architecture · USPTO 2-01026094 (May 26) · Continuation 2-01029681 (May 27)
HomeBenchmark Arena