Plain language
What this result means
This result matters because the system treated the repository as a dependency graph, not just a list of numbers. A better source covering can imply a better derived covering. The record only counts after the derived object is actually built and checked.
- A (v,k,t)-covering is a set of k-point blocks. Every t-point subset must appear inside at least one block; fewer blocks is better.
- The four records come from the standard induction relation C(v,k,t) <= C(v-1,k-1,t-1) + C(v-1,k,t).
- This is not a claim that the source coverings are new. The win is detecting the unpropagated derived cells, materializing the witnesses, and checking them.
The reference behind the result
Where are the La Jolla Covering Repository tables?
La Jolla collects small covering designs and the bounds they establish. A table entry tells you how many blocks a known construction needs. It is an upper bound on the minimum, unless optimality has also been proved.
- Dan Gordon’s covering-design guide ↗The maintainer’s introduction and links to the dataset.
- La Jolla data on GitHub ↗Downloadable coverings, notebooks, and code.
- Covering Repository ↗A repository with later improvements, linked from Gordon’s guide.
Our comparison is dated. The four baseline counts below are those recorded for the July 3, 2026 report. Use the linked repositories to check current bounds.
How to read C(v,k,t)
- v
- Items available
- k
- Items in each block
- t
- Items in each combination to cover
- C(v,k,t)
- Fewest blocks possible
Every group of t items must occur together in at least one block. Building a valid design with N blocks proves C(v,k,t) ≤ N. A smaller construction improves that bound; it does not by itself prove the minimum.
An example you can inspect
Seven pools cover all 21 pairs.
Take seven items and put three in each pool. These seven pools contain every possible pair exactly once. Choose a pair to see where it appears.
Items 1 and 2 occur together in pool 1.
- Pool 1123
- Pool 2145
- Pool 3167
- Pool 4246
- Pool 5257
- Pool 6347
- Pool 7356
This small design is optimal: one pool covers three pairs, so covering 21 pairs takes at least seven pools. The construction reaches that lower bound: C(7,3,2) = 7.
In a suitable assay, pooling could reduce the initial measurements from 21 to seven. A positive pool still needs follow-up tests to identify the active pair. This is an illustrative design, separate from the four Numaro results.
From blocks to experiments
Are covering designs useful in practice?
Yes, when one experiment can include several items and the goal is to expose every required combination. A block becomes a pool; covering all pairs or triples means each possible interaction is present somewhere in the initial screen.
When would our smaller designs save work?
If a task fits one of our four parameter sets and each block represents one initial experiment, the smaller designs require 186 to 4,160 fewer initial experiments than the recorded baseline. Actual cost and time depend on the assay, allowed mixtures, repeat measurements, and follow-up tests. We have not measured those savings in a lab.
Covering designs and covering arrays solve different tasks.
Covering designs group items into subsets. Covering arrays assign values to variables so that every required interaction of settings is tested. NIST uses covering arrays for software and hardware testing. NIST’s explanation ↗ Our covering-array results address that related testing problem.
Visual notes
How to read the result
Result table
Four covering numbers lowered below the recorded La Jolla bounds.
| Cell | LJCR snapshot | Numaro | Delta | Note |
|---|---|---|---|---|
| C(77,25,8) | 179,223 | 175,063 | -4,160 | R4 theorem (C(76,24,7) + C(76,25,8)) |
| C(92,7,4) | 132,580 | 131,122 | -1,458 | exhaustive: 2,794,155 subsets (C(91,6,3) + C(91,7,4)) |
| C(86,23,7) | 153,683 | 153,047 | -636 | R4 theorem (C(85,22,6) + C(85,23,7)) |
| C(79,21,7) | 165,213 | 165,027 | -186 | R4 theorem (C(78,20,6) + C(78,21,7)) |
Method
How it was found
The campaign parsed the LJCR table, propagated standard covering relations to a fixpoint, and kept the four cells where the table snapshot lagged the induction construction.
- Fetched and parsed 9,482 LJCR entries.
- Applied delete-point, enlarge-block, reduce-t, and R4 induction relations.
- Built each derived covering from the two linked LJCR source coverings.
- Checked nearby mature cells with CP-SAT and simulated annealing; those attempts matched records but did not beat them.
Verification
How it was checked
Each candidate has to have the claimed number of blocks, the right block size, and points in the correct range. For C(92,7,4), every one of the 2,794,155 four-point subsets is checked directly. For the larger rows, the check rebuilds the two source counts and uses the standard induction theorem: subsets containing the new point are covered by one source; subsets not containing it are covered by the other.
The construction behind all four rows
Add one point. Combine two coverings.
The campaign found improvements already implied by the repository’s own source coverings. It used the standard induction bound:
C(v,k,t) ≤ C(v−1,k−1,t−1) + C(v−1,k,t)
- Take a covering of the first v−1 points with blocks of size k−1. Add the new point v to every block.
- Append a second covering of those same v−1 points, with blocks of size k.
- Any required subset containing the new point is covered by the first part. Any subset without it is covered by the second.
| Constructed design | Add the new point to | Append | Total blocks |
|---|---|---|---|
| C(92,7,4) | C(91,6,3) 6,765 blocks | C(91,7,4) 124,357 blocks | 131,122 |
| C(77,25,8) | C(76,24,7) 36,144 blocks | C(76,25,8) 138,919 blocks | 175,063 |
| C(79,21,7) | C(78,20,6) 28,051 blocks | C(78,21,7) 136,976 blocks | 165,027 |
| C(86,23,7) | C(85,22,6) 27,279 blocks | C(85,23,7) 125,768 blocks | 153,047 |
For the headline row, 6,765 + 124,357 = 131,122 blocks. The source coverings were already known. Numaro’s contribution was finding the propagation gap and building and checking the derived block systems.
What the evidence establishes
One exhaustive check. Three conditional constructions.
We rechecked the saved files for this page update. Every block has the required size, distinct points in range, and no duplicate blocks. The counts and the split into the two R4 parts match the recorded sources.
| Design | Required subsets | Coverage evidence |
|---|---|---|
| C(92,7,4) | 2,794,155 | Every subset checked; zero uncovered |
| C(77,25,8) | 21,042,072,975 | R4 theorem; assumes valid LJCR source coverings |
| C(79,21,7) | 2,898,753,715 | R4 theorem; assumes valid LJCR source coverings |
| C(86,23,7) | 5,373,200,880 | R4 theorem; assumes valid LJCR source coverings |
The three large designs have billions of required subsets. Their coverage follows from the induction theorem if the original source coverings are valid. The campaign checked source counts and sizes and used sampling as a smoke test; it did not independently enumerate every subset of those sources. Sampling is not a proof.
What the autoresearch record adds
Propagation worked where search did not.
The campaign reported parsing 9,482 table entries and repeatedly applying four construction relations: delete a point, enlarge a block, lower the required subset size, and combine coverings by induction. The four reported improvements came from the induction relation.
- The saved simulated-annealing sweep selected 184 cells and recorded zero improvements, with two near misses.
- The campaign report describes CP-SAT and simulated annealing matching tested records without beating them.
- A differentiable GPU approach failed its calibration on instances the annealing solver could solve.
These are outcomes of the recorded experiments, not a claim that those methods cannot work. Here the useful step was propagating known constructions before spending more effort on a new search.
Inspect and reuse the saved designs
Download the blocks and run the audit.
The archive contains all four explicit block systems, an offline Python audit, provenance, and file hashes. Each JSON file lists blocks as arrays of point labels from 1 through v, together with its parameters and source counts.
Extract the archive, open its covering-designs directory, and run:
python3 audit.pyPython 3 and its standard library are sufficient. No network access is required. The audit enumerates coverage for C(92,7,4); the other three results retain the source-validity assumption described above. It does not check current records.
Scope
What is not being claimed
These constructions improve the LJCR comparison values recorded for the July 2026 report; we have not established that they remain current records. No optimality proof or new source covering is claimed. C(92,7,4) has independent exhaustive coverage verification. The three t=7/8 constructions rely on the original LJCR source coverings being valid; their source coverage was not independently enumerated. Practical savings in a laboratory or production system have not been measured.
References
Baseline sources
- D. Gordon, La Jolla Covering Repository.
- La Jolla individual covering lookup.
- Dan Gordon’s current covering-design guide and dataset links.
- La Jolla Combinatorics Repository: downloadable dataset and notebooks.
- Covering Repository: later improvements, linked by Dan Gordon.
- Pflüger et al. (2026), A General Group Testing Strategy for Discovering Chemical Cooperativity.
- NIST: covering arrays for software and hardware testing.
- Saved covering designs, verification scope, and source counts (JSON).
Citation
How to cite
Numaro AI Autoresearch Team. "Covering-design records improving the La Jolla Covering Repository." Numaro Research Report NUMARO-2026-006, 2026.
@techreport{numaro2026CoveringDesignsLa,
title = {Covering-design records improving the La Jolla Covering Repository},
author = {Numaro AI Autoresearch Team},
institution = {Numaro},
number = {NUMARO-2026-006},
year = {2026},
url = {https://numaro.tech/research/covering-designs-la-jolla-2026/}
}