Ranking the roster, without one permanent winner.

A review of the distributor book and a design for ranking, recognition and spiffs. Figures are from the live Airtable base Shopify Shop unless labelled otherwise, measured across all 39,082 commission rows joined to order totals and deduplicated by order. airtable is live production. dev fixture is the June snapshot on the dev Medusa cluster, illustrative only. vrx sheet is the Verified RX commission sheet.

Three people are 65% of the business.

Trailing-twelve-month revenue per distributor code, from live Airtable. Deduplicated by order — a naive per-rep sum double-counts multi-rep orders and overstates the book by 22% ($34.68M against a true $28.43M).

Top 3 = 65%. Top 10 = 83%. Top 20 = 91%. airtable

The distribution, in percentiles

The 90th percentile rep does $258k. The 100th does sixty times that. airtable

The book is growing, not stalling

Monthly revenue has risen every month bar two since launch: $244k in Aug 2025 → $3.95M in Jun 2026 → $3.18M in Aug 2026 (partial). 118 of 141 codes were active in the last 30 days. Whatever else is wrong, the selling motion works. airtable

Not a person, not a house account — an entire sales organisation.

The Verified RX commission sheet settles it. Code 3716 is the roll-up for a 68-person selling organisation, three managers deep. Jake's $15.57M is VRX's number, not one man's. vrx sheet

And it is deeper than two levels. Three reps are marked under jaq (beneath Jacquelyn, who is herself under Jake) and one under Fab (beneath Fabiana). So the real org is at least three tiers. The portal models parent_agent_id one level deep everywhere it is used — roster scope, sub-agents, team grouping. It cannot currently represent this shape.

Why this changes the whole design

• Ranking 3716 against individual reps compares a 68-person org to one person. It is not a close call or a tuning problem; the comparison is meaningless.
• The VRX sub-reps have no attributable identity. Their sales roll into 3716, so today they cannot be individually ranked at all — you cannot gamify people the system cannot see.
• The rate ladder is already a tier system. 15% / 20% / 25% / 35% is a hierarchy someone maintains by hand. It is the closest thing to an existing tier scheme in the business, and it is in a spreadsheet.
• This is now a third source of commission rates, alongside Distributors.Commission Rate and the 6,011-row Customer-Distributor Splits. Three places, and they do not have to agree.

Weight classes, drawn on the real distribution.

These are not invented brackets — they are the buckets the roster already falls into, from live Airtable trailing-twelve-month revenue. Within a cohort the rank key is movement, not magnitude: a rep who grows 40% beats one who grew 4%, whatever the absolute dollars.

Two planes, not one

Because 3716 is an organisation, a single list cannot hold both kinds of competitor. The portal already has the right shape for this — the master view renders Top Teams and Top Individuals separately. Use it.

How the brackets hold

• Redrawn quarterly, not monthly. Monthly redraws let people yo-yo between cohorts and make streaks meaningless. Announced at quarter start, frozen for the quarter.
• Sandbagging does not pay. The pool scales with the cohort's own revenue, so winning a lower bracket is worth materially less than placing in a higher one. Dropping a class to win it is a losing trade by construction.
• Tiers only go up. Rank is periodic and volatile; tier is cumulative and earned once — lifetime revenue, customers acquired, months active. A floor that never drops underneath a race that resets is what makes this work.

Revenue is live. Tenure barely exists. Payment is fiction.

There is one real scoring system already in the base, and it is worth copying: Customers carries Order Frequency Score, Order Value Trend Score, Product Breadth Score, Health Score, Zone and Score Delta. That is a working composite-score methodology built in-house — it just scores customers rather than reps. Reuse the shape.

Ship the status game before the money game.

Phase one ranks on revenue, orders and customers — never on commission. Revenue is the one number here that is unambiguously live and trustworthy. Rules v2.0 cuts over 1 September, two days out, and its rate logic does not exist in the schema yet; everything after that date is stamped isProvisional: true. A recognition board can live with a provisional number. A payout cannot.

• Per-cohort boards — trailing 90 days, growth against your own prior 90.
• Tiers — cumulative, earned once, never lost.
• Streaks — consecutive months with an order; consecutive months of growth.
• Milestones — first order, tenth customer, first $100k, twelve months active.
• Most improved — a single cross-cohort award, the one place the whole roster competes on equal terms, because percentage growth is scale-free.

There is no working payout ledger to attach money to.

This is the single biggest finding for phase two, and it is not a design problem — it is a missing system. Commission accrual is alive and correct. Commission payment died in September 2025.

Commissions Payable is a formula over Total Collected from Customer. That input is empty on 98.5% of orders, so Commissions Payable and Commission Status evaluate to $0 and "No Commission" across essentially the whole base. Any "commission paid" number pulled from Airtable today is fabricated. The only surviving signal is Commission Payout Status, a manual single-select with 3,113 rows marked PAID, maintained by hand.

So phase two is gated on rebuilding payout tracking, not on ranking design. Spiffs cannot be paid against a ledger that says everyone is owed zero. When that exists, the mechanics:

• Frozen periods. At close, snapshot each rep's aggregate to an immutable record stamped with its rules version. A closed period is never recomputed.
• Auditable to the order. A rep opens their number and sees the orders behind it. One wrong payout poisons the whole scheme.
• Written tie-breaks, in order — growth %, then new customers, then order count, then earliest first order. The current board breaks ties alphabetically by name, which is indefensible once money is attached.
• A dispute window — fixed days after close, final thereafter.
• Pool per cohort, a percentage of that cohort's own revenue, so it scales without anyone re-deciding it each quarter.

Assume it will be optimised against.

• Order splitting. Breaking one order into five to inflate order count. Rank on revenue and distinct customers; cap how much order count can contribute.
• Month-boundary timing. Holding or rushing orders across a close. Freeze on order date, not entry date, so a pushed order still lands in its true period.
• Self-dealing. A rep ordering through an account they control. Flag where the customer domain matches the rep's — flag for review, do not auto-exclude.
• Cohort sandbagging. Handled by the quarterly redraw and pool scaling above.
• Split abuse. Splits are keyed by customer and sourced from Airtable, entirely outside the parent/child tree. Whoever edits that base can move commission attribution. Once money rides on it, that base needs a change log.

Fix these before any rep sees a rank.

Extend what exists. Almost none of this is new code.

The all-distributors console (GET /distributor-overview) already computes per distributor: orders, revenue, commission, distinct customers, active months, split orders, flagged count and AOV. It is all-time and unbucketed — the single axis it lacks is a period window. Add periods and cohorts to that endpoint and the ranking falls out of data that already exists.

One attribution rule this design must not break. Revenue and order count belong to the owning distributor once, however many parties share the commission — only commission divides. Every board must state which of the two it ranks on. Mixing them is how a leaderboard starts disagreeing with a payout.

Rank on movement, inside your weight class.

Cohort decides who you are measured against. Score decides where you land inside it. Nothing in the score is an absolute dollar figure — that is deliberate, and it is the only way a $13k rep and a $250k rep can be judged by one formula.

Unit and window

• Rank the distributor code, never the person. Eleven names hold two or three records each; person-level ranking would double-count them.
• Trailing 90 days, recomputed monthly, against the immediately preceding 90. A calendar month makes the last day of the month decisive and invites timing games — and at a median of $13,469/yr, one month is too lumpy to rank on at all.
• Order date, not entry date. Frozen at close, so a held or rushed order still lands in the period it belongs to.

Attribution — where the 22% error lives

Revenue and order count belong to the owning distributor once, however many parties share the commission. Split participants get an assisted counter that is displayed but never feeds the score — otherwise co-repping becomes a way to farm rank, and 5,583 orders already carry more than one rep.

The four components

Acquisition must be measured globally, not per rep. If "new" means new-to-you, moving a customer between reps mints an acquisition out of nothing. Dedupe customer identity on normalised email, falling back to company name.

Growth is a log-ratio, not a percentage

Why log. It is symmetric: doubling scores +0.693, halving scores −0.693. Percent growth is unbounded above and floored at −100%, which is precisely why every percent-growth leaderboard ever built is won by whoever had the smallest base.

Why k. It damps the denominator. Without it, $200 → $2,000 is +900% and beats every real performance in the bracket. With k drawn from the cohort's own median, a tiny base cannot manufacture a large ratio.

Then shrink it — the part that makes it honest

The classic failure of a growth board is a rep with two orders posting +300% and topping the table. Every component is therefore pulled toward the cohort midpoint in proportion to how much evidence the rep actually has:

A rep at the cohort median gets w = 0.5 — half their own signal, half the midpoint. At four times the median, w = 0.8. One order against a cohort median of twelve gives w ≈ 0.08, so they sit near 50 whatever their ratio claims. This is empirical-Bayes shrinkage, and it is the difference between a board people trust and one they laugh at.

Each component is a within-cohort percentile rather than a z-score — the distributions are skewed even inside a bracket, and a percentile survives that and can be explained to a rep in one sentence: you are in the 70th percentile of your bracket for new customers.

The composite

Tie-breaks, written down and deterministic, because money will eventually ride on them: acquisition, then growth, then owned order count, then earliest first owned order. Never alphabetical by name — which is exactly what the current board does today.

A floor that never drops, under a race that resets.

Rank alone is punishing: lose once and you are visibly demoted. Tier alone is inert: reach it and stop. Running both is what makes this stick.

• Rank — the score above, per cohort, recomputed monthly. Volatile by design.
• Tier — a separate monotonic axis on lifetime owned figures: revenue, customers ever acquired, months with an order. Earned once, never lost. A bad quarter costs you rank, never status.
• Streaks — consecutive months with an order, consecutive months of positive growth. Aimed at the 64-rep Starter band, where the problem is that selling is sporadic rather than small.
• Most Improved — highest shrunk growth across the entire roster. Log-growth is scale-free, so this is the one board where all 141 codes can legitimately compete together.

The rate ladder already in use — 15% / 20% / 25% / 35% on the Verified RX sheet — is the closest thing the business has to a tier system today. It is hand-maintained in a spreadsheet. Worth folding into the tier design rather than inventing a parallel one. vrx sheet

Pay on incremental revenue, never on total.

Each cohort's pool is a fixed percentage of that cohort's growth against the prior period, not of its revenue. Three consequences, all of them good:

• It self-funds. You are only ever paying out of money that did not exist last period.
• If a cohort shrinks, its pool is zero. No participation trophies drawn from flat revenue.
• It scales without anyone re-deciding the number every quarter, and it makes sandbagging into a lower cohort a losing trade — a smaller cohort generates a smaller pool.

Distribute 50 / 30 / 20 across the top three, or smoothly across everyone above the cohort median. Freeze at close, stamp the rules version, and keep it auditable down to the individual orders that composed the number — a rep who cannot see what made their figure will not trust it, and one wrong payout poisons the whole scheme.

None of this can run yet. Commissions Payable evaluates to $0 base-wide, Payout Items has zero rows, and Commission Paid So Far is populated on 76 of 39,082 records. Phase two is gated on rebuilding payout tracking, not on ranking design.