The brief points at a collection rather than pasting sources into itself. How much of the right material a system pulls on its own is part of what the benchmark measures, so the collection is built to be traversed rather than read front to back.
Twelve sources are encoded as a concept space in the Open Knowledge Format. A plain folder tells an agent what exists. It does not tell the agent which source to open first, which one stops being true when another changes, or which one contains the rule it is about to break.
Three fields carry that. The first holds typed edges between sources, drawn from a closed ten-word vocabulary, so following one relation reaches what must be read first while following another reaches what changes if this source changes. The second holds the acceptance criteria and names the signal each one resolves against. The third binds a comparison on those signals to a pre-configured action.
Together the second and third say, in a form a machine can evaluate, what a passing run looks like and what happens when one fails.
Every rule is paused, no rule or action node is compiled, and reading a document triggers nothing in any case. A document that causes actions when an agent processes it is structurally the same thing as a prompt injection, which is why declaration and execution are kept as separate organs. The rules exist because a scoring rubric written this way lives in the collection rather than in a reviewer's memory.
The frozen input. Twenty-six companies in five dimensions, three coordinate sets, the composite as a node weight, and the hash of the state file it came from.
The five dimensions, their weights, the score categories, and the principle that the index measures structural advantage rather than popularity.
Read for the gap ledger, the watchlist, the scoring rules and the per-company reasoning. Never read for coordinates.
Why the method applies, the baselines to reproduce, and the five things the method may not claim.
Per-construction parameters verified against real data, acceptance tests, and the traps already encountered.
The twenty-two report sections, the eight section-jobs, and the slot this work targets. A new section must say something no existing section says.
The reader is a client executive. Network vocabulary does not reach them. A correct result in unreadable prose is not a result.
The typed graph a result should join rather than stand beside, so it compounds across issues instead of expiring with the week.
The existing chart language to extend rather than replace.
The interchange format, the type vocabulary, and where the export firewall stops.
Pointer rather than payload, and the reason reads never trigger.
The frame results are recorded into, so a choice between systems rests on recorded work rather than impression.
An agent does not need all eight. It needs the survey for vocabulary and the specific paper for whichever construction it is building.
| Construction | Source |
|---|---|
| Spatial networks, the survey | Barthélemy, Physics Reports 499, 2011 |
| Random geometric | Dall and Christensen, Phys. Rev. E 66, 016121 |
| Unit disk | Clark, Colbourn and Johnson, Discrete Mathematics 86, 1990 |
| Waxman | Waxman, IEEE JSAC 6(9), 1988 |
| Threshold | Mahadev and Peled, via the survey at doi 10.37236/219 |
| Soft geometric | Penrose, Ann. Appl. Probab. 26(2), 2016 |
| Geographical threshold | Masuda, Miwa and Konno, Phys. Rev. E 71, 036108 |
| Thresholded geometric | Bradonjić et al., hdl 2117/111425 |
All eight are collected in the NetworkX guide to geometric generators, which is where the list came from in the first place.
# rebuild the frozen input from the live board
python3 data/build_point_cloud.py
# rebuild it from any archived week
python3 data/build_point_cloud.py --state \
../../microco/competitive-intel/state/archive/W25-2026.json \
--out /tmp/W25.json
# re-run the baseline
cd runs/claude-code-2026-08-17/code && python3 01_random_geometric.pyThe environment is NetworkX 3.5 with NumPy 1.26.4 and SciPy, all already present. Nothing needs installing. Demonstrations 1, 4, 6 and 7 are deterministic; the rest are seeded and the seeds are recorded.