Pavo System Knowledge
Knowledge systems that learn from every investigation, experiment, and outcome.
Pavo builds a verifiable System Book from code, data, traces, experiments, and expert knowledge, giving every person and agent the context to improve the system intelligently.
- 01Reconstruct the system
- 02Verify what is true
- 03Compound every learning
System Knowledge
How notification eligibility works
Who is allowed to receive a lifecycle message, and the policies that gate every send. Reconstructed from code, warehouse queries, policy docs, and one expert rule.
Key policy claims
The worldview
Knowledge systems should continuously verify what they know, learn from outcomes, and invent new knowledge through work.
Knowledge is not static documentation. It is continuously strengthened through expert verification, investigations, experiments, interventions, and production results.
Propagate every update →
A complete system for building and improving production knowledge.
Reconstruct the system. Verify critical claims. Apply trusted knowledge across improvement work. Learn from every decision and production outcome.
The System Book, in full
Every claim keeps its sources, its conflicts, and its verification.
Scattered evidence resolved into a book people and agents can both read, and audit.
Detailed product case
From scattered evidence to verified system understanding.
Question
How does the notification system decide who should receive a message?
Scattered evidence
- CodePolicy logic
- WarehouseEligibility data
- ServiceFatigue rules
- ExperimentsTreatment evidence
- DashboardsOutcome metrics
- ExpertsDocumented exceptions
Before Pavo
6 sources · unreconciledThe same decision is described six different ways, with conflicts unresolved and no single source people or agents can trust.
Pavo reconstructs and verifies
System Book updated- 01Reconstruct the full decision flow
- 02Surface a conflicting claim
- 03Ask an expert to verify it
- 04Update the System Book
- 05Make verified context available to people and agents
- 06Write the later experiment outcome back into knowledge
Verified answer
6 sources · 1 conflict resolvedA user receives a message when eligibility clears the fatigue rules and an active treatment applies, unless a documented exception overrides it, the conflict Pavo surfaced and had an expert verify.
FAQs
System Knowledge is how Pavo learns one production system: a search ranker, a recommendation feed, a notification engine. It reads your code, warehouses, dashboards, and past experiments, and extracts typed facts, each carrying a citation back to the exact file, query, or record it came from.
- Indexed facts, each with its citation
- A System Book, written as chapters
- A Knowledge Bench that scores the book
- A hub where new findings are written back
Every system your team works on gets its own knowledge base, so each investigation builds on the last instead of starting over.
A wiki stores documents and enterprise search retrieves them. Pavo reads the sources themselves and extracts atomic, typed facts: how a metric is defined, what is true about the data, why a choice was made, how a component works. One connected system commonly yields tens of thousands of them.
Every fact links back to the code, query, or table it came from, so you can open any claim and see its evidence rather than trusting a summary. That grounding is also what lets a claim be caught when its source no longer holds.
The Knowledge Bench measures it. It is a set of pointed questions about your system, scored by a judge that is separate from whatever produced the knowledge, so the system is not grading its own work. Every verdict carries its reasoning, so a score is never a black box.
It scores coverage and correctness apart: whether the knowledge can answer at all, and whether the answer is right. When the judge cannot settle a question, it escalates to a person rather than inventing a score. You own the question set, and it should encode the mistakes your team has already seen people make.
Work writes back into it. Every deep dive, experiment read, and new definition discovered while doing the work lands in the knowledge hub as a finding, linked through an evidence chain to the run, the queries, and the reasoning that produced it. Re-scanning refreshes the index and recovers findings from past tasks.
Each artifact carries a state, validated, unvalidated, or to verify, so you always know how far to lean on it. The book itself is versioned, so a claim can be traced back to the moment it was true.
Yes, and not only Pavo's. The extracted facts are exposed as a searchable index. Inside Pavo, the book, the bench, and every agent draw on it directly; outside, your own coding agents can query the same index, so they answer from your system's real facts instead of guessing.
Because every fact carries its citation, an agent can hand over the evidence with the answer, rather than treating a retrieved document or an old decision as unquestioned truth.
One production system to start with, and read access to at least one source close to it:
- Warehouses: Databricks, BigQuery, Snowflake
- Semantic layer: dbt
- Code: GitHub, Bitbucket
- Dashboards: Tableau, Power BI, Superset, QuickSight, Redash
- Experimentation: Eppo, Statsig
- Docs: Confluence, Google Docs
Every connector is read-only and scoped to the catalogs, repositories, or projects you choose, and your credentials stay inside your own Pavo instance. Where a live connection is not possible, you can upload anonymized files instead. Plan for the first pass to take a few sessions: indexing runs on its own, and your time goes into reviewing the book and vetting the bench.