Forty-two decisions.
D-01 through D-42 with their amendments, each condensed in full operational substance, and a crosswalk to the rule each one became in the coding agent's instructions.
docs/decisions/decision-register.mdAbout
A reference implementation, kept honest in public.
MetricMine is a contract-driven medallion pipeline that runs end to end on one machine, built in the open so that every claim it makes can be checked by someone who has never talked to its author.
The project
What it is, in one paragraph.
Raw data lands in bronze exactly as received. A deterministic profiler describes it. Machine-readable ODCS contracts, hand-written or drafted by two narrow AI agents, shape silver and gold behind a human approval gate. An auto-modeling engine emits the entire gold layer from the approved contracts, and a read-only MCP server serves the result with a context registry one join from any payload, so an assistant answers business questions grounded in both the data and the meaning behind it. The current tagged release is v1.1.3; the changelog records what each release landed.
Origin
2023, then 2026 in the open.
MetricMine began as a private prototype in 2023. The current codebase was rebuilt in the open in 2026 from a written specification, with the clean-room rule recorded in NOTICE: the repository was built fresh from an independent written specification, and no code from the 2023 prototype was read, imported, or adapted.
The name and the mark
MetricMine. One word, capital M twice.
The symbol is three connected cubes. The cubes stand for data modeled as cubes; the connections stand for sources joined into one star, which is what the engine builds.
The name and the logo are not covered by the Apache-2.0 license. Forks and derivative works may not use them in a way that suggests endorsement by, or affiliation with, this project. The brand system, with its vector masters and the normative standard, lives in the website repository.
How it is kept
The register is the project's memory.
Every design judgment is a numbered decision in a versioned register, minted by its own documentation pull request before any implementing change lands. Every measurement that contradicted a plan is a numbered finding. The stable line changes only through that machinery.
Sixty-six findings.
F-01 through F-66: toolchain behavior observed live, never inferred, from the first scratch gate proof to the toolchain rung, each one with the evidence it cites.
docs/verification/gate_proof_findings.mdOne stability rule.
From v1.0.0, the pipeline, the engine, the serving layer, the two proposers, and the gates change only through the register, and a clone of main at any tag gets a working demo.
CLAUDE.md, rule 19Who
Built by Justin Goff.
MetricMine is a personal open source project, built and maintained on its author's own time. The commits, the pull requests, and the register carry the record; the profile is Justin Goff on GitHub.
Contact
One address, one organization.
Questions, findings, and pull requests belong on the repository. Anything else reaches the project at github@metricmine.ai. The GitHub organization is metricminellc, a namespace and nothing more.
From the builder
Why I built it.
The pattern
My 18 years of work span software development, application development, data intelligence, data science, machine learning, and AI. One thing has held since college and in every organization since: everyone's data is messy, and everyone's data is fragmented to some degree. No exceptions, whatever the company's size or maturity.
That fragmentation is always the blocker. It blocked transactional applications. It blocked analytics. It blocked predictive analytics and machine learning. Today it blocks generative AI and agentic workflows. The workload changes every few years. The bottleneck does not.
The fix
The fix has stayed the same too. Start with governance, embed it in standard operating procedures, and make it easy. Easy is the part most organizations skip. If governance depends on how much any one person knows about the process, it will not hold. Automation that applies governance by default is what gives it a chance to stick. Funding to clean this up is almost never there, one more reason the automated path is the only realistic one.
2023, then a fresh start
By 2023 I had watched the same cycle repeat for years: fragmented data, a real appetite for unified analytics on top of it, and no clean path between the two. I developed the concept of a unified, automated event star and built a private prototype on my own time. I planned to release it and stopped short. Family time was limited, and there were real gaps in the solution I could not close then.
Advances in AI closed that gap, specifically embedding agents into the workflow itself. The 2023 version handled sources statically. I started over from scratch and rebuilt it as a dynamic solution, paired with agentic Claude Code and a proposer agent workflow. Data contracts carry the governance up front, and the same contracts drive deterministic generation of the models. Predictable and inspectable.
Trust first
If data cannot be trusted, it does not get adopted. Even correct data sits unused when nobody can verify it. Software and workflows that nobody adopts do nothing. Trust is the prerequisite.
What MetricMine is
MetricMine is a downloadable pipeline anyone can run. It is my 18 years of hard-earned pattern recognition, extracted, and the thing I wish I had at the start. It would have saved me months. Probably years.
Industry standard tools sit at every gate. Airbyte at the door, with its catalog of 600+ connectors (Airbyte's own count) for your own sources. dbt for transformation. DuckDB for local storage. The contracts and the compiled context are warehouse-agnostic files in git, and dbt profiles carry the execution plane. The goal is a short path from a new source to a governed model that agents and downstream applications can use.
Thanks
Thank you to my wife and kids for their patience through the late nights and weekends. Thank you to God for the wisdom and energy to keep at this through long work hours and family demands, off and on from 2023 through 2026 and beyond.
Justin Goff
Commit your work to the LORD, and your plans will be established.
Proverbs 16:3 ESV