Your agent should only rely on what your team has decided is true. Currently.
Propperly maintains what your analytics organization has concluded from the work it has already done: which definitions are current, which were superseded and why, which conflict, and what still needs a decision. Your team approves every change by design. Your agent reads the result.
active_user counted on a 30-day window.
finance model · Q1
Dashboard v1 still reports the 30-day count.
dashboard v1 · 3 places
active_user is the rolling 28-day definition for product reporting.
human decision · analyst · Jun 12
Which definition governs historical restatements?
awaiting review
Your agent can reach every source and still not know which answer is current.
Retrieval finds the documents. It doesn’t tell the next analyst, or the agent answering an executive, which definition is current, which decision replaced another, whether the numbers still conflict, or what was never resolved. That is not primarily a retrieval problem.
Which definition is the one to use right now.
What replaced what, and why it changed.
Whether sources still disagree with each other.
What the team has never actually settled.
Propperly maintains the conclusion, not just the documents.
Retrieval can hand your agent every file that mentions active_user. Propperly tracks how that one definition actually evolved, so the agent reads today’s answer, with the history behind it.
- superseded
A 30-day window was the working definition.
finance model · Q1 - challenged
A new analysis argued the 30-day window overstated active users.
analysis · May - current
The rolling 28-day definition was approved for product reporting.
human decision · analyst · Jun 12 - conflict
Three dashboards still report the old 30-day count.
dashboard v1 · 3 places - unresolved
Which definition governs historical restatements is still open.
awaiting review
The agent reads which definition of active_user currently governs and the evidence behind that governance.
Test it on one analytics domain before you integrate.
You don’t need a data migration to see the value. Point Propperly at the work your team already has for a single domain, and it reconstructs what looks current, superseded, conflicting and unresolved, for your team to review.
- metric definitions
- dashboards
- analyses
- SQL & notebooks
- decisions
- docs & tickets
The Local Evaluation Kit runs on your side, on work you choose, offered as a hands-on evaluation, not a public download.
Start with the understanding. Keep it current.
Two stages of working with Propperly, first build the reviewed understanding your analytics agent relies on, then keep it current as the organization changes. Start on one bounded domain.
Maintained Current State
Propperly maintains the state of the work.
Read what exists
Propperly reads the work already there, documents, analyses, code, AI sessions, and rebuilds the project's current understanding.
Your team decides and attributes
Someone confirms, corrects, or rejects it. That decision is attributed and persists, downstream AI inherits governed state, not raw inference.
◆ human authority requiredBuild on it
People and AI agents read the approved state directly: decisions, evidence, open questions, ruled-out paths, sources.
Keep it current
When sources change, recommission the domain. Propperly surfaces relevant differences and routes what requires renewed human governance, the governed state updates without rewriting its history.
We measure this and publish what held.
Propperly's maintained-state approach has been evaluated on sealed research corpora. We publish what the evaluations showed, including the results that went against us.Read the research record.
Evidence-led writing on maintained understanding.
How a project loses its own history, why AI memory is not a maintained state, and what analytics agents need beyond retrieval.
What context graphs still get wrong about organizational knowledge
Analytics practiceWhy your analytics agent was accurate at launch and wrong six months later
Product thinkingThe Picture Keeps Moving
Project Continuity, keep long-running Claude Code projects from losing their history.