Each system receives the same file. Nobody adapts it per harness, because a brief edited between runs stops being a comparison. This is what the brief asks for and why each piece is in it.
They are not eight ways to do one thing. They divide into four jobs, and two of the eight exist only to keep the other six honest.
| # | Construction | Job | What a passing run produces |
|---|---|---|---|
| 1 | Random geometric | Find the peers | Groups at a justified radius, with the sweep that justified it |
| 2 | Unit disk | Find the peers | An unoccupied position, checked against the existing gap ledger |
| 3 | Waxman | Prove it earned its place | A null comparison over many seeds. Not a map |
| 4 | Threshold | Prove it earned its place | The control, plus what geometry adds over it |
| 5 | Soft geometric | Survive the uncertainty | Per-connection confidence as a frequency over many seeds |
| 6 | Geographical threshold | Filter to what matters | A fix for the known saturation, or a reasoned rejection |
| 7 | Thresholded geometric | Filter to what matters | A printable list with the exclusions named |
| 8 | Spatial networks | Framing | No build. One plain sentence per demonstration saying what the distance measures |
Every construction comes with a parameter range already verified against the real data, so an agent that reports something far outside it has a problem worth finding rather than a discovery.
Violating any of these fails a run whatever else it produced.
Agents are told to read the six scoring criteria before they start. Only one thing is withheld, which is a preliminary answer we already hold to one of the three questions. Testing whether a system can guess the criteria tests nothing worth knowing.
One. Which single construction deserves a permanent slot in the weekly report, and why that one?
Two. What does a geometric read tell the reader that the published ranking cannot? Answer it against the control. If the answer is nothing, say so. That is a valid and useful result, and it scores higher than a manufactured finding.
Three. Does the structure hold up? Twenty-one archived weekly states are available. Rebuild across at least eight and report whether the peer structure persists or is an artifact of one issue.
Each run closes with a self-report covering what it built, what it could not build, and where it is least certain. On the evidence of the first run, that document is the most useful one in the set.
The sixth construction misbehaves on this data and the brief says so plainly. Its distance term saturates: at a low threshold it connects every pair, and at a high one the most connected companies come out as the tightly packed weak ones, which is backwards.
Fixing it with a stated reason passes. Rejecting it with a stated reason passes. Tuning the threshold until the picture looks acceptable fails. How a system handles the task it cannot complete separates harnesses better than the seven it can.
slugs = [n["slug"] for n in cloud["nodes"]]
pos = {n["slug"]: n["unit"] for n in cloud["nodes"]}
G = nx.random_geometric_graph(slugs, 0.30, pos=pos) # correct
G = nx.random_geometric_graph(26, 0.30, pos=pos) # raises inside the KD-treeThe call-signature note above is also in the brief. It costs an agent an hour if it is missing, and including it means the benchmark measures analysis rather than a library quirk.