Admin
// signed inAPI-cap reminder
Loading subscriber monitor…
How subscribers are counted
Unique active paying Stripe customers on SparkVibeAI paid plans, counted once per customer. Excludes trials, free accounts, manual grants, and Stripe test-mode subscriptions. Checked every five minutes; counts older than 15 minutes are marked stale.
Customers
Every account, newest first. Click a row for plan, credits, keys, and the support actions.
| Username | Plan | Status | Renews / ends | Verified | Interviews | Joined | |
|---|---|---|---|---|---|---|---|
| Loading… | |||||||
Create an account
Support-created accounts skip email verification. Role "spark" is a customer.
Unit economics
$29 Pro · 8 credits · observed LLM spend and AWS billing kept separate
Historical LLM cost
Daily mean for clean, current-stack completed runs
Full-utilization scenarios
All 8 monthly credits consumed
AWS account billing
Actual Cost Explorer UnblendedCost, not the run ledger's Lambda estimate
Definitions and limitations
Stripe webhook receipts
The last 50 events the billing webhook processed and what each did to the ledger.
| When | Event | Status | Grant | Detail |
|---|---|---|---|---|
| Loading… | ||||
Signup clusters
Free accounts sharing a signup IP or device fingerprint. Report only; decide per account from the customer drawer.
Rate-limit overrides
Raise or clear one operator's limits for one class. Takes effect on the next request. Defaults shown for reference.
1 · Committee convergence
Is the committee getting better? Scores are grouped by metric era and never blended across the 2026-08-23 metric change.
Loading…
2 · Cost anatomy per run
Where the LLM dollars go: stage, vendor, model, cache share, and the tail. Ledger rows are configured-dollar observations.
3 · Interview lift
Does the interview produce better specs? Funnel plus a paired comparison of interviewed vs not-interviewed builds.
4 · Agent funnel
Discovery → get_started → credential → first tool call → checkout → commission → paid.
5 · Outcomes
Customer-reported "did this spec ship?" 14 days after a build, plus GitHub pushes as a proxy. This is the calibration input that replaced the verifier.
6 · Rate-limit pressure
Denied and would-deny counts per hour from the limiter's own metrics.
Calibration tools (legacy)
Export the calibration set or recompute rubric weights. Recompute needs at least 10 (spec, rubric, outcome) tuples; outcomes now come from widget 5.
Your recent actions
Every write from this console is logged with a snapshot. Snapshot-restorable actions can be undone here; ledger moves are reversed with the opposite adjustment.
- Loading…