near-absorbing-state-rate-deflated-by-snapshot-residue

name	near-absorbing-state-rate-deflated-by-snapshot-residue
description	A "conversion / never-convert / progression" rate for a near-absorbing funnel state (created-not- submitted, abandoned-cart, stalled-ticket, dormant-lead, draft-not-finished) swings by orders of magnitude depending on WHEN you observe the pool, NOT on a data error. Observe each record at CREATION-time → high conversion (fast movers included). Observe the pool as it sits at a LATE SNAPSHOT date → near-zero conversion / ~99% "never move" — because the fast movers already drained out, leaving the dead-draft RESIDUE. Use when: (1) reconciling two or more "X% never convert" numbers that disagree wildly (e.g. under 1% vs ~65% vs ~90%), (2) a "99% never move / near-absorbing state" stat is quoted as the conversion baseline for an actionable cohort, (3) a denominator is "everyone currently in status S" read at one date. Fix: before calling the numbers contradictory, align them on ONE outcome and re-derive that outcome observed AT record-creation (the unconditional rate) vs at the late snapshot (the residue rate) on the SAME population — the gap IS the artifact.
author	Claude Code
version	1.0.0
date	"2026-06-25T00:00:00.000Z"
disable-model-invocation	true

A near-absorbing state's "never-convert" rate is a snapshot-residue (observation-time) artifact

Problem

You are handed several "X% never convert" numbers for the same funnel state and they disagree by 1–2 orders of magnitude (e.g. "~99% of created-not-submitted records never move" vs "~65% never submit" vs "~90% never progress"). The instinct is "one of these is wrong" or "they contradict." Usually none is wrong — they measure the rate after observing the pool at different times, and a near-absorbing state's residue rate is dominated by observation time, not by the intervention.

The mechanism: in a state where most records move fast (e.g. median time-to-submit = 0 days, half submit the same day they are created), the composition of "everyone currently in status S" shifts violently with the observation date:

• Observe at creation (every record, the moment it enters S) → most convert → high rate.
• Observe the pool sitting in S at a late snapshot date → the fast movers already left; what remains is the dead-draft residue (the slow/abandoned) → near-zero conversion / ~99% "stuck."

So "99% never move" is not the actionable stall rate — it is the near-absorbing residue measured at a late date. Quoting it as the conversion baseline for a cohort you intend to act on overstates the loss and understates the reachable opportunity.

Trigger conditions

• Two+ "never convert / never progress / never submit" numbers for the same state disagree by a large factor, and someone wants you to "pick the right one" / "reconcile before shipping copy."
• A "near-absorbing state" / "99% never move" statistic is being used as a card's conversion denominator.
• The state's time-to-next-step distribution is heavily front-loaded (median ≈ 0, big same-day spike) — the precondition that makes observation time dominate.
• A denominator is phrased "everyone currently in status S" (a late snapshot) rather than "everyone who entered S."

Solution

1. Decompose the disagreement on three axes before calling it a contradiction: different outcome (deposit vs submit vs any-forward-move — a deeper outcome is always smaller), different population (whole-population vs reachable/web-linked vs re-observed subsample), and different observation time (at-creation vs late-snapshot). Two "contradictory" numbers usually differ on ≥2 axes and are simply different quantities.
2. Prove the observation-time axis on ONE outcome, ONE population. Re-derive the same outcome (e.g. ever-submit, from an immutable submitted_at stamp) two ways on the same population:
   • at creation: of all records ever created, what % ever convert (the unconditional rate).
   • late-snapshot residue: of records still in S at a late date D (created ≤ D AND (converted IS NULL OR converted > D)), what % ever subsequently convert. A large gap (66% vs ~2% is real) is the artifact — same population, opposite numbers, pure observation time. Check it is robust across several D values (not a cherry-picked date).
3. Recommend the actionable grain, not the residue. For an intervention card, headline the reachable cohort's rate at a fixed term-relative window (e.g. advance-within-60-days), and relegate the "99% never move" population number to context, explicitly labeled a snapshot- residue artifact — never the headline conversion baseline.

Verification

• The at-creation vs late-residue rates differ by a large factor on the same population (smoking gun) and the gap holds across multiple snapshot dates.
• Each disagreeing source number, once you state its (outcome × population × observation-time) tuple, reproduces — so they were three quantities, not three estimates of one.
• An independent re-derivation on the immutable event stamp reproduces both ends of the contrast.

Example (application funnel, verified on live event data)

An "application-started, not submitted" card had 3 conflicting "never-submit" numbers. Re-derived live on the submit outcome from the immutable submitted_at stamp:

• At creation (all ~99k created records): ~two-thirds ever submit (over half of submitters submit the same day; median 0 days) → ~one-third never.
• Late-snapshot residue (~30k records still not-submitted at a late snapshot date): ~2% ever submit / ~98% never — robust across several snapshot dates (all landing in the low single digits, roughly 2–8%).
• The prior "~99% never move" figure was exactly this residue (a deeper deposit-outcome variant).
• Decision: the card headlines the reachable stalled cohort, ~28% advance@60 (reachable population, fixed window); the ~99% residue number is context only. The >30× creation-vs-residue gap is the proof it was an observation-time artifact.

Notes

• See also [[snapshot-churns-out-one-outcome-class-biases-the-rate]] — the adjacent snapshot trap, but a different mechanism: that one is cross-outcome-class retention bias (denials churn out → inflated accept rate). This one is observation-time selection within a single outcome (the residue of a fast-draining pool). Both: "don't trust a rate read off current state."
• Related: [[null-status-with-engagement-conflates-never-vs-form-stall]] (a NULL/stuck status can mean "form-stall," not "never"), [[trailing-window-population-decline-seasonal-vs-real]] (a trailing window shifts pool composition over time), [[descriptive-universe-vs-action-worklist-cohort-size]] (whole- population vs reachable-worklist — the population axis of step 1).
• The same trap hits: abandoned-cart "X% never buy" (observe at add-to-cart vs at a stale-cart sweep), support "X% never resolved" (observe at open vs at a backlog snapshot), lead "X% never engage." Whenever most records move fast and you read the rate off the residual pool, you measure the residue, not the funnel.