Billo Research — our institutional research collections, maintained under shared conventions.
Individual labs are able to pursue unrelated questions while sharing naming, navigation, provenance, retention, and evidence rules. We adopt an archival posture without adding undue ceremony or requiring individuals to learn archival vocabulary just to start an experiment. Our approach is designed to help future readers and experimenters reconstruct our experiments, view the results directly, verify or contradict them, and determine what remains unproven or unexamined.
Billo Research is our collection of lab setups, experiments, notebooks, protocols, evidence, analyses, and reports. We follow the best practices of the scientific-methods and laboratory-recordkeeping guides provided in our ai-engineering repository.
Start with the lab index. Each lab provides its own scope, notebook entries, setup procedures, and experiment index. Each experiment connects its question and protocol to run records, evidence, analysis, and findings.
Our first lab, Local LLMs, documents an initial local inference setup and a coding-model evaluation protocol. Setup commissioning and experiment execution are pending; no results are claimed yet. The collection also supports symbolic and numerical investigations, replications, and exploratory work.
A thought or observation can begin as a dated notebook entry. A bounded investigation gets an experiment record; an execution attempt gets a run record. Directories grow with the work. See the creation decision guide and record templates.
labs/
lab01-<slug>/
README.md
journal/
procedures/
experiment01-<slug>/
README.md
protocol.md
inputs/
code/
runs/
run0001/
run.md
raw/
derived/
analysis/
report.md
This is an illustrative layout, not a requirement to create every directory. The lab index records the conventions adopted here.
Our records aim to preserve the path from a finding through its analysis and run records to the exact inputs, code, configuration, and environment used. Failed attempts, deviations, inconclusive results, and unanswered questions belong in that record. A documented recipe is distinguished from an actual rerun, and an author's checks from independent verification.
Native computational files can accompany readable exports so readers can inspect both the method and the displayed results. Large or externally held evidence is represented by an inventory with retrieval and access information.
These practices draw on CERN's analysis preservation and archival guidance, adapted for our purposes. The skill's source register explains the sources and adaptations. This project is not affiliated with CERN.
Billo Systems maintains the collection. Use repository issues to raise a question, missing dependency, or conflicting result; use a pull request to contribute records. See CONTRIBUTING.md for the first steps.
Our original code and documentation are available under the MIT License. Third-party inputs and separately licensed datasets retain their own terms, recorded with the material or its inventory.