Decision-grade analytics · Ontology as IP

churnOS

A governed judgment layer for agentic products, which accounts and capabilities to ship, throttle, or kill, priced by cost of leaving live.

Built for teams shipping assistants, workspace agents, and ops automations. Not another success-rate dashboard, auditable GrowthDecisionRecords with YAML-governed policy.

Request repo access →
Weekly meeting, rendered

What your agent review could look like

Each row is a GrowthDecisionRecord: what's wrong, what it may cost to leave live, and what to do, with policy provenance from semantics.yaml.

Capability Risk Radar
What to ship, throttle, or kill
Ranked decisions priced by cost of leaving live.
At-risk account list
Synthetic demo Static preview; Streamlit simulator available on request (teaching warehouse and authored priors, not production telemetry).
What is in the repo

START → DECIDE → LEARN

Synthetic teaching warehouse, not production telemetry. Request access to run the Streamlit simulator locally.

START
  • Product Profile presets (assistant-heavy, CRM workspace, ops mission)
  • Synthetic warehouse generator (seats, runs, approvals, connectors)
  • Reproducible seeds for demos and tests
  • Profile switches ontology vertical and synthetic priors
DECIDE
  • Capability Risk Radar (account + capability decisions, health score, tornado)
  • Activation & Habit, Trust & Approval, Run Economics, Connector Blast Radius
  • Version Compare for agent version regressions
  • Override decisions with audit trail and JSONL persistence
LEARN
  • Experiments and Agentic Flags (CPSO, TTFV, HITL under synthetic treatments)
  • Outcome Flywheel (retention write-back on the same record)
  • Semantics Console, Taxonomy Browser, Record Inspector
  • Legacy analytics archive (retention, unit economics, CRO, marketplace)
Who this is for

Builders shipping agent capabilities every week

Shipping new capabilities weekly

Run success rates look fine, but nothing tells you which capability is quietly hurting retention, or which one to kill.

Owning retention for agent seats

Trust incidents and approval fatigue pile up. You can't prioritize which problems to fix first without a ranked list.

Watching inference burn climb

Long reasoning loops eat margin. You see aggregate token spend but not which capabilities are uneconomic per seat.

What if

What if your weekly agent review looked like this?

Agentic products fail on verified outcomes, cost per successful outcome, and trust after multi-step runs, not another trace explorer.

What if
Paying looked healthy, but nobody got a first win?

Revenue metrics stay green while accounts never hit a verified agent outcome. The “AI churn wave” is often activation failure, not classic feature adoption.

What if
One intent burned a dozen metered steps?

CPSO and retry amplification show when “success” is margin-negative, especially for power users under flat pricing.

What if
The agent “worked” until humans took over?

Rising human-in-the-loop, dismissals, and scary side effects erase trust faster than usage dashboards admit.

The problem

You ship agent capabilities from traces and run logs, but traces don't tell you which ones to kill. Success rates hide retention harm until seats churn.

When you throttle a capability, nothing writes back to retention. You never learn if the call helped.

The solution

churnOS emits auditable GrowthDecisionRecords: ranked accounts and capabilities with dollar impact, YAML-governed verdicts, and an outcome flywheel that closes the loop.

Override when you disagree. Write back retention impact; even in synthetic mode, you see how decisions connect to churn.

Ontology as IP

Machine-readable policy, not hardcoded heuristics

Taxonomy → semantics → schema → record. Change verdict rules in YAML; the Radar reclassifies without a code deploy.

The ontology/ package defines exception categories, vertical-specific semantics.yaml action maps, and JSON Schema contracts for GrowthDecisionRecord. Explore the repo via LinkedIn.

Before → After

The shift isn't more dashboards; it's a weekly decision ritual with dollar impact and an audit trail.

Area Before After
Weekly review Run success rate dashboards Ranked ship / throttle / kill list
Harm signal Anecdotal support tickets Dollar impact per capability
Policy Hardcoded Python heuristics YAML-governed verdicts & actions
Audit trail Slack threads & spreadsheets Saved decision records you can revisit
Account risk CSM spreadsheets and gut feel Ranked account decisions with $ impact
Retention loop Decisions never revisit churn Outcome flywheel writes back (demo)
Proof in production

Moovez / Quotely case study

ChurnOS-style decision-ops patterns applied to last-mile quoting after CV, OR, and SMS shipped.

Read the case study →