Diagnostic report for Northloop

Evidence

Last 30 days: Sign Up 1,840 · signup_completed 620 · user_signed_up 210 · Signup Completed 95. Same product moment; four chart definitions.

Why it matters

Activation and acquisition reports silently disagree depending on which name an analyst picks. Funnel step one is not stable.

Fix

• Pick one canonical, e.g. Sign Up Completed (past-tense, Title Case).
• Stop firing the other three in the client SDK and any server path.
• In Amplitude, merge or map historical aliases to the canonical (or document a chart-level mapped view until history is backfilled).
• Update the tracking plan row and owners the same day the code ships.

Evidence

Workspace Created 520 and team_created 340 both fire in the same 30-day window. Your notes say the rename shipped August 2025; mobile (or an old build) still emits the old name.

Why it matters

Workspace creation is undercounted by ~40% if charts use only the new name; double-counted if someone unions both without deduping users.

Fix

• Find every remaining team_created call site (likely older mobile build or a forgotten server job).
• Ship Workspace Created only; keep a temporary dual-fire only if you need a measured cutover week, then kill the old name.
• Map historical team_created → Workspace Created in the taxonomy.

Evidence

Volumes: 9,200 · 3,100 · 180. Near-dup clustering catches Log In/login; User Logged In stays separate and still dribbles in.

Why it matters

DAU/WAU proxies and “returning user” segments drift by source file.

Fix

• Canonical: Log In with property method (password · sso · magic_link — already present on many rows).
• Remove login and User Logged In producers; alias history.

Evidence

Check In Submitted 5,400 · checkin_submitted 980 · Check-In Submitted 140. Hyphen variant clusters with the Title Case name; checkin_submitted does not and is easy to miss in a UI-only pass.

Why it matters

Check-in is your core product action. A 15%+ shadow stream breaks retention curves.

Fix

• Canonical: Check In Submitted.
• Require workspace_id (see #7 for key spelling).
• Alias both variants; delete producers.

Evidence

Upgrade family: Plan Upgraded 410 · plan_upgraded 155 · Subscription Upgraded 60. Purchase pair: Purchase Completed 290 vs purchase_completed 275 — nearly equal volume, classic client + server double-fire shape. Keep Refund Processed separate.

Why it matters

Revenue and conversion charts can read ~2× on purchases if both names are summed. Upgrade reporting is split three ways.

Fix

• Canonical upgrade: Plan Upgraded. Canonical purchase: Purchase Completed.
• Confirm whether purchase fires from both browser and webhook; keep one path (prefer server) and pass a shared insert id if you must dual-send.
• Do not merge refunds into purchases.
• Require numeric amount + currency + plan_type on both (see #6, #8).

Evidence

Sample rows for Sign Up variants and Invite Sent carry email and full_name. Your own notes flag a 2025 growth experiment as the likely source.

Why it matters

Event properties are the wrong place for direct identifiers: wider export surface, harder retention/deletion, and easy leakage into shared charts.

Fix

• Stop sending email and full_name on events immediately.
• Identify users with Amplitude user properties / identity API (or your warehouse profile table), not on every event.
• Scrub or drop the properties on historical data per your retention policy.

Evidence

Plan is split across plan, plan_type, and planType (each populated on about a third of sample rows). Workspace id appears as workspace_id, Workspace ID, and sometimes org_id with the same ws_* values.

Why it matters

Breakdowns by plan or workspace look empty for large slices of traffic. Required-field checks on the plan fail even when “a plan field” was sent.

Fix

• Canonical keys: plan_type, workspace_id. Snake case everywhere.
• Reserve org_id for a real organization id if you have one; do not overload it as workspace.
• Alias the old keys in the warehouse or taxonomy map during transition.

Evidence

Values pass shows pro/Pro/PRO, pro_annual/Pro Annual, country US/U.S./us/USA/United States, status active/Active/ACTIVE. At least one plan value is a free-text deal note, not a category.

Why it matters

Segment and cohort filters under-count. Free text in a category key makes every new deal a new “plan.”

Fix

• Allow-list plan_type: free · pro · pro_annual · team · enterprise. Put deal notes in a separate optional string property if you truly need them.
• Store country as ISO-3166 alpha-2 (US, CA, GB, DE).
• Lowercase status enums at the producer.

Evidence

In the sample of upgrade/purchase-like rows, roughly one in four arrives with an empty amount. Populated values arrive as digit strings (49, 49.00, 99) rather than a consistent number type. currency mixes USD and usd.

Why it matters

Revenue sums drop nulls; type mixes break some warehouse casts.

Fix

• Validate at the producer: number amount, uppercase ISO currency.
• Quarantine failures to a side stream instead of silently omitting.

Evidence

Page views: 128k + 44k + 8.9k under three names. Buttons: 22k + 6.7k + 1.2k. Both families cluster cleanly once casing/tense is normalized.

Why it matters

High-volume noise crowds the event list and makes taxonomy reviews harder. Rarely used for product decisions in SaaS if you already have session replay or a warehouse pagefeed.

Fix

• If you keep them: one Page View, one Button Clicked with button_id (not Button ID).
• If you do not use them in Amplitude charts, stop sending and reclaim the noise budget.

Evidence

experiment_flag_exposed 45,000 and Feature Flag Evaluated 12,000 — neither appears in the tracking plan. Combined volume rivals your intentional product events.

Why it matters

Orphan high-volume events inflate bills and bury the taxonomy. If experimentation needs them, they should be documented and named once.

Fix

• Decide: product analytics vs experiment tool. If the experiment platform already logs exposures, drop these from Amplitude.
• If you keep one, pick a single canonical and add it to the plan with an owner.

Evidence

Plan lists 22 events; inventory has 36. Plan rows with zero fires in the window: Onboarding Completed, Activation Milestone, NPS Submitted, Logout, Password Reset Requested. Your notes confirm onboarding “never shipped.”

Why it matters

Analysts trust the spreadsheet; production diverged in 2025. New events ship without owners.

Fix

• Mark never-shipped rows as deprecated or remove them.
• Add every production name you intend to keep (after the merges above).
• Assign an owner column that is a real person or team, reviewed quarterly.

Evidence

Sign-up rows alternate between source and utm_source for the same idea, with casing drift (google vs Google).

Why it matters

Channel reports need a single field.

Fix

• Prefer utm_source (and friends) when the value is campaign traffic; use a separate signup_source for product-defined origin if needed.
• Lowercase at ingest.

Evidence

legacy_pulse_v1 last seen 2025-12-02 (3 events) · Old Dashboard Viewed last seen 2025-08-19 (1 event).

Why it matters

Clutter and confusion for anyone browsing the full event list.

Fix

• Block/hide in Amplitude; remove any leftover producers; note as deprecated in the plan.

Evidence

About 8% of sample rows have an empty user_id while other identity fields may still be present.

Why it matters

Breaks user-level funnels and inflates “anonymous” counts after login.

Fix

• After auth, ensure the SDK has the user id before product events fire.
• Allow truly anonymous pre-auth events; do not mix the two without a merge path.