Share:

Debt collection — everything we know

The evidence pack a new dashboard gets designed on top of. What the process actually is end to end, how big the book is, why payments fail, what we have tried and what it returned, what is broken today, where every number lives, which numbers are safe to quote and which are not. Written 23 September 2026.

Static by design. Every figure is quoted from a dated measurement and the report that produced it is named beside it, so re-measuring produces a NEW number with a NEW date rather than silently editing this page. The only live values are four committed counts-only summaries. Companions: https://reports.iconicbyai.com/money (the one operational page), https://reports.iconicbyai.com/debt-plan (the plan), https://reports.iconicbyai.com/Studio1_First_Miss_Handoff.html (first payments in depth).

The book being chased
$16.79M
10,038 accounts owe it. 11,605 accounts in the debt GHL.
Answered a text in 30d
8.3%
968 accounts. NEVER quote the 58% lifetime figure.
First instalments miss
~47%
~102 a week, ~21 a working day. The break is instalment #1.
Never paid a penny
3,460
$7.40M — 44% of the book, nothing after the deposit.
Placed with the agency
17,919
$31.9M of balance. Running ~300-500/month, unbroken.
Automated retries
ZERO
Nothing re-charges a failed payment. Ever. A person does.

Eight rules. Every wrong answer this area has produced was a DEFINITION problem rather than an arithmetic one — the sums were right and the basis was not. A subagent once reproduced a table exactly and was still wrong, because the interpretation was. Read these before quoting anything on the other tabs.

1

NOTHING re-charges a failed payment. Every second attempt is a person

Verified straight from Square on a live series (00440822, Texas, 17 invoices): only the SCHEDULED invoice carries automatic_payment_source: CARD_ON_FILE and a card_id; every past-due one reads NONE. Square charges the card once, on the due date, and never again, then the series mints a fresh invoice next cycle. Measured across the zombie pile at 0.90 failed charges per missed cycle — under one. So there is no dunning engine to tune and no automated baseline exists anywhere in this history. The words retry, retried, re-presented, retry rate are banned about our own book — in replies, page copy, table headers, CSV columns, variable names and commit messages. Say “chased by a collector”, “never chased at all”, “when a collector got hold of them it worked 44% of the time”. Neil has corrected this twice.

2

Therefore every conversion rate on a second attempt is SELECTED — never quote one as an expected rate

A collector only puts a card through after the customer has agreed to pay, so of course it converts. The day-2-3 cell reading 90.3% measures a conversation went well, not day 2 is a good day to charge a card. The same confound runs through the step-down evidence: charges for MORE than the failed amount convert at 73.5% — just as well as the smallest step-down — which is the giveaway that the amount is not the mechanism. Anything automated is a NEW capability with NO baseline: write UNKNOWN. A trial needs a no-retry control arm or the automated arm gets judged against a human conversation rate and looks like a failure however well it does.

3

Lead with the things that did NOT happen — those are the only unselected numbers

Nothing selected an absence, so an absence can be quoted flat. Safe: “58.5% of soft declines were never chased within 30 days — $12.4M never asked for again” · “77% of failed openings are never chased at all” · “4,831 accounts have never once answered a text”. A rate measured on the people somebody reached is not safe. Build the dashboard’s headline tiles out of absences.

4

NEVER quote a blended rate on this book — it is always a mix effect

This has produced a wrong headline three separate times. Reply rate: blended fell 48.5% → 21.1%, but a first-month conversation answers at 51-57% and has not moved in a year; the re-chased pile answers at 12-13% and grew twelvefold. Neither segment changed. Engagement: 6,706 accounts have replied at some point = 58% “engaged”; the 30-day figure is 8.3%. Collection rate: 55.9% → 39.8% was not a collapse — LIVE payers held at 70.6% → 65.8% while the DEAD pile (no payment 3+ months, still auto-billed) doubled from 36.6% to 53.7% of billing. Every one of these is the same error: the denominator changed, not the behaviour. Split the population before you draw the line.

5

Say which BASIS you are on before quoting any collection number — five bases, all correct

Neil’s own ~63% is collected / (finance written − written off) on SETTLED cohorts = 64.0% on live cities 24+ months old. Every other basis lands somewhere else and reads as a contradiction: collected/written on settled cohorts 48.8% · collected/written on the whole live book 42.2% · share of instalments attempted each month that were taken 39.8% · (deposit + collected)/sale value 58.6% · collected against what has actually FALLEN DUE since Apr 2026 43.7%. The last one is the honest one for young cohorts — collected ÷ finance written is a measure of AGE, not of collection. A five-month-old contract on a twelve-payment plan cannot have paid more than 5/12.

6

Split never-paid from stopped-paying before you design anything

They are not the same customer and they will not answer the same call. 3,460 of the 10,038 owing accounts never paid a single instalment after the deposit — $7.40M, 44% of the book. On the zombie pile, 4,687 never paid a cent ($6.69M that only ever existed as an attempted charge) against 5,114 who paid and went quiet. Never-paid-a-cent climbs monotonically with the miss run — 28.5% of payers, 41.5% at one miss, 43.3%, 46.5%, 53.4% among zombies. Ever-paid-anything out-predicts the decline reason: every reason family clusters at 10-31% next-cycle recovery, while never-paid alone reads 41.5% at one miss. Sort on that first.

7

The damage is done at the FIRST miss — everything after is arguing over the remains

Given a series has just missed N in a row, does the next cycle get paid? 1 miss → 20.8% · 2 → 12.9% · 3 → 7.9% · 4 → 5.0%. (Charge-attempt view of the same thing: 44.0% → 23.5% → 16.2% → 14.6% — higher because it cannot see cycles where no charge was ever attempted. Name which view you are quoting.) The collapse is between miss 1 and miss 2. A cut-off at 4 misses is a decision about what is left after the loss has already happened.

8

Age-match, and check the population is LIVE, before aggregating anything

A month built mid-month stays short. A cohort under 12 months old has not had time to pay out. A rate off a pile that includes 47,571 long-dead series inflates outstanding to $33.9M — 88% dead history. Three separate reports have published a wrong number by skipping this. The stillBilling flag (at least one invoice scheduled ahead, not CANCELED/REFUNDED/DRAFT) is the scope for every headline.

A finance agreement is a deposit at the till then a monthly series charged against a stored card. Those are two different events with two different failure modes. Everything below is about the second one. The chain has seven links and three of them are a human doing something by hand.

StageWhat happensWho / what does itThe detail that matters
S0 The charge fails Square, on the due date A finance agreement is a deposit at the till then a monthly series charged against a stored card. Square fires the card once on the due date. It fails, and that is the end of Square’s involvement — no reminder, no second attempt. 577 of 600 recurring invoices carry reminders: [], and of the 23 that have any, every one is a no-card-on-file invoice. 0 of 557 auto-charge invoices carry a reminder. Exactly backwards.
S0b Nothing warns them beforehand Nobody — verified, not assumed No proactive pre-payment message exists anywhere. Scanned the whole CRM note corpus — 439,694 lines, 43,274 distinct bookings, 2.05 BILLION characters including pasted email correspondence — and every hit is reactive. delivery_method: EMAIL on 600 of 600 recurring invoices; exactly 2 invoices in the whole business are set to SMS. Square does not text about an upcoming payment. A customer said it plainly in 2022: “This is the first email or contact I have received in regards to the contract. I was under the assumption that the bill was being automatically debited monthly.”
S1 The miss reaches a file Our own daily build The /square-declines report pulls /v2/payments per location and writes declines.json (a rolling 5-day window, never a history). /chase-daily turns it into one CSV per day, deduped within the day, and records everyone handed over in a Supabase ledger. Push the day BEFORE — Texas recurring charges land ~18:00 UTC, long after the 7am cron, so today’s file is always partial.
S2 A human loads it into GHL Hamad, by hand, once a day The file is downloaded and uploaded into the US Debt Collection GHL sub-account. This is the single biggest structural weakness in the whole process. It is a daily batch, so day 0 does not exist: the fastest text ever sent left 23.2 hours after the charge failed, median 24.4h, nothing inside 12 hours ever. Against the measured curve — day 1 is 48% of everything a missed opening ever returns inside ten days — the first text lands at the END of the most valuable day.
S3 GHL texts them Workflows, UI-only New + no owner → Assign states for debt lead → update business name from reference code → Debt not debt (first-time missers: offers 3 options to find out why, handles the conversation with AI, assigns a team member). Already owned → 2nd Missed payment Automation. The opening text is “1 card issue · 2 money is a bit tight · 3 something else”. The decline reason never reaches GHL at all — the outbound text is identical whatever the issuer said.
S4 A collector works it Nine desks, GHL boards + GoTo phones Each collector has their own GHL pipeline. Cards get assigned, conversations happen by SMS in GHL and by phone through GoTo. GoTo does not write back to GHL, so a phone-worked card reads as unchased — 0 calls appear in the GHL thread and that is a source limit, never evidence nobody rang. Money taken on the phone shows in Square as KEYED.
S5 A reduction is offered Caet, late A credit line is posted against the CRM finance agreement. This is the most expensive tool we have and it is used first, not last — 600 of the 945 accounts reduced in 2026 had no failed charge at all. The ladder Neil approved puts Caet LAST, not at the front door.
S6 Placed into collections The CRM flag, then Monterey The booking screen’s “Confirm collection ?” dialog calls BookingUIClass.collection(id), which stamps collected='Y' + a collectiondate. The /application/diary/reports/csv-incollections.php export selects on it — client ref, name, address, date of birth, both emails, balance. That export is BROKEN (HTTP 500, header row only) and we rebuild it byte-for-byte instead. Customer-facing: a warning email (“10 BUSINESS DAYS TO CONTACT BEFORE ACCOUNT IS PLACED INTO COLLECTIONS”) then a text naming the agency and 800-456-2225.

The process facts nobody has written down anywhere else

1

The CRM calls it COLLECTIONS, never Monterey

Searching the CRM for “Monterey” does not find the handover. The agency’s name is written nowhere in any field — only in free-typed notes and finance references. When looking for a third-party process, search for the FUNCTION, not the vendor’s name.

2

Nothing records the file being SENT

The flag means placed into collections, not sent to Monterey. No send log, no schedule, no FTP or email trail exists in the CRM. If it leaves by email or through Monterey’s portal, that happens outside this system. Five live Monterey logins exist inside our CRM (all @montereyfinancial.com, all created by member 281, all carrying BOOKER/DESKTOP/FINANCE/MANAGEMENT/PUBLIC/SUPERFINANCE/USER/VAULT), so they may now pull it themselves. Unresolved.

3

Money stopped coming back from the agency

Nothing has been keyed back as “Paid to Monterey” since 31 Aug 2026 — and that is not a stale source: August holds 6,153 ordinary payment lines against July’s 6,368. We keep placing accounts and nothing comes back.

4

The escalation ladder Neil approved, S0-S5

S0 Square miss + decline reason (day 0) → S1 automated self-serve recovery (d0-7) → S2 automated persistent chase (d7-21) → S3 collector calls (d21+) → S4 Caet’s reduction offer (d45+) → S5 Monterey. Payment at any stage drops the account out; if it had already gone to Monterey that is a red alert. Caet is the last human stage, not the front door — today accounts reach her with no chase history, so we may be discounting people who would have paid at S1. Timings 7/21/45 are still open.

5

The full 22-step contact ladder is drafted and UNSENT

Day −5 to collections: 22 steps · 6 forks · a 4-rung discount ladder · 6 guard-rails · 8 pieces of copy · 6 metrics, on /Studio1_First_Miss_Handoff.html → The full cycle. Shape follows the measured curve: 0+2h, 1, 2, 3, 5, 7 → 10, 14, 21, 25, 29 → day 30 fork → 31/33/35/38 → day 40 escalation → 45, 60, 75-90 Monterey. Nothing sent, nothing scheduled. The aim is CONTACT, not collection — the day-7 message is three buttons and “can’t this month” is a SUCCESS that routes OUT of the ladder into an arrangement.

6

Why day 30 is not a policy choice

Square charges once on the due date and never again, then the series mints a fresh invoice. Day 30 is the next invoice falling due — not a re-charge, and not something we trigger.

How big it is, on every basis anyone uses. Say which basis you are on before quoting any of these — they are all correct and they all disagree.

FigureValuePeriodSourceRead it this way
The whole finance book, all time $126.5M written 2019 → Aug 2026, 59,597 agreements /finance-cohort, 7 Aug 2026, off the CRM Debtor Ledger. $53.6M collected (42.4%) · $21.2M written off · $51.6M outstanding. The identity finance = paid + credits + balance holds to $6,034 on $126.5M (0.005%). Worst cities on collected %: Chicago 21.8%, Nashville 34.9%, Houston 38.1%. Best: Boston 47.8%, Florida 44.5%.
The debt GHL sub-account 11,605 accounts As at 4 Sep 2026 /debt-status (pcTyOadbzBqZO7v3cbwn). 13,870 open cards = 11,605 accounts. 10,038 owe $16.79M · 1,080 paid off · 456 written down · 16 never financed · 15 no join. $16.79M is a BALANCE, not arrears — landing here means missing at least one payment, but plenty went back on plan. Say “the size of the book being chased”, never “overdue”.
Contact state of that book 968 live ≤30 days, as at 4 Sep 2026 /debt-status, lib/debt_unassigned.status(). Five-rung ladder: 968 live (≤30d) · 1,185 cooling (31-90d) · 4,551 cold (90d+) · 4,831 silent · 70 untexted. The live figure is 8.3%. 6,706 have replied at some point — quoting that reads as 58% engaged and is the single most misleading number available here.
Ownership 1,217 with nobody As at 4 Sep 2026 /debt-status, /debt-unassigned, /debt-handout. 10,388 accounts carry a collector. The ownerless pile was worked down on 4 Sep — 589 cards assigned, 0 failed, ownerless on the Unassigned board went 641 → 52, and every one of the 52 is ownerless for a filter reason. 197 more cards are owned by three DELETED staff accounts — worse than unassigned, because they still LOOK owned.
Placed with the agency 17,919 agreements As at 23 Sep 2026 /monterey, via the CRM collected flag. $31.9M of balance placed · 17,912 still owe over $10 · most recent placement 17 Sep 2026. By year: 2021 135 · 2022 2,266 · 2023 3,515 · 2024 3,666 · 2025 5,678 · 2026 2,659. A live process running ~300-500/month, unbroken.
Money the agency returned $923,969 980 agreements, 3,017 finance lines /monterey, 23 Sep 2026 — the only place the word appears in the CRM. 923,969.46 keyed back as “Paid to Monterey” across 923,969.46 — plus $204,336 written down (179 credits), less $4,511 of reversals and fees. Nothing since 2026-08-31. 64 of those agreements returned nothing at all.
The zombie pile 2,656 series 4+ consecutive misses AND still billing, 7 Sep 2026 /zombie-reasons, off a 416,258-invoice cache. $4.35M outstanding · $544,275 invoiced every month that never lands · 53.4% (1,417) never paid a cent. Depth: 4-5 954 · 6-8 798 · 9-11 599 · 12-17 300 · 18+ 5. Texas 898 worst, then NY 667. A wider count — accounts with a trailing 4+ fail run, whether or not still billed — is 9,801, $14.6M. Name which you are quoting.
The live early window 2,751 series 1-3 misses, 7 Sep 2026 /zombie-reasons/early.csv. $932,466 owed on the runs, $536,951/mo still billed. 1 miss 1,313 · 2 misses 862 · 3 misses 576. This is where the intervention belongs — the hazard curve collapses between miss 1 and miss 2, so a programme aimed at the 4+ pile is aimed at the wreckage.
The recurring series book 10,781 still billing As at 23 Sep 2026 /paid-off + lib/recurring_series.js. 64,127 series read from 422,333 invoices. A series only ever holds ONE invoice ahead — 10,777 of 10,781. That is the shape of the whole problem: exposure is one cycle at a time, and a settled agreement is only visible between settling and its next charge. Today: 8 owe nothing and are still billing, 38 are set to charge MORE than the balance ($4,244 over).
Written off / reduced $19.7M all time 11,511 agreements, 2021 → Aug 2026 /reductions, assembled from the Debtor Ledger + each booking’s finance schedule. The CRM has NO reductions report — this is built from two places. July 2026 was the biggest write-down month in twelve: $201,329 / 173 credits. The credit’s date is whatever the person posting it typed; the CRM holds no created-at for a credit line, so never present it as an audit timestamp.

~85% of every failure is a money code, and the bucket mix (soft 82.6% / hard 10.5% / structural 6.9%) is flat across every month and every city. So the reason code is a routing fact, not a customer-wealth fact — it tells you what to say, not who can pay.

WhatThe numberBasisRead it this way
~85% of it is a MONEY code 55% over a limit · 23% refused no reason · 10% insufficient funds · 7% issuer blocked recurring · 5% technical 1 May - 31 Aug 2026, 3,733 first instalments, recurring only, deduped by payment id (/money#proposal, re-derived for the board). Bucketed structurally it is soft 82.6% · hard 10.5% · structural 6.9%, and that mix is FLAT across all three months and every city — a property of the book, not a spike or a team problem.
TRANSACTION_LIMIT is NOT a per-transaction cap The biggest single code Square’s own wording; measured on $25 and $50 charges. It means the card is at its limit. Telling those customers to “split the payment” fails every time. Action is MAXED (different card / after payday), not SPLIT. And it is a money code: it recovers at 20.9%/30d against INSUFFICIENT_FUNDS 21.3% — identical. The customers agree: of those who picked a number from the GHL text, 7 of 11 chose “money is a bit tight”, and 7 of 8 money answers had a TRANSACTION_LIMIT decline.
CARDHOLDER_INSUFFICIENT_PERMISSIONS is the one with a 30-second fix 24.0% of early-access first misses vs 5.4% traditional — 4.4× /first-miss-week, /decline-buckets. The issuer has fenced the card off from card-not-present recurring charges and only the cardholder can lift it, usually a toggle in the banking app. ~8 people a week. This is the only place in the whole area where “contact your bank” is correct — and nobody is told, because the decline reason never reaches GHL. ISSUER_INSTALLMENT_ERROR is similar in kind: the bank refused the card specifically as an instalment.
CARD_EXPIRED reaches almost nobody 30 of 11,475 first misses /opening-payment, 16 Sep 2026. Boston Aug 2026 sample: zero of 41. The “update your expiry” framing that everyone reaches for first is wrong. So is the plain insufficient-funds framing. Dead-card codes are a REGULAR-decliner signature — on the payment test, all 9 INVALID_ACCOUNT, both PAN_FAILURE and both CARD_EXPIRED sat in the chronic group and none in the other two.
The economy is real but roughly a THIRD of it INSUFFICIENT_FUNDS +168% per 1,000 attempts Jan-Jul 2025 vs Jan-Jul 2026, normalised per attempt (attempts grew 23%). Overall fail rate 36.8% → 43.1%; declined value $8.21M → $12.29M. Of the excess beyond volume growth: blocked cards ≈48%, no-funds ≈32%, invalid account ≈12%, generic ≈11%, maxed-out −13% (offsetting). Maxed-out cards, the single biggest affordability code, got slightly BETTER — which a broad economic squeeze would not do.
The dominant driver is BLOCKED CARDS, and it is not economic CARDHOLDER_INSUFFICIENT_PERMISSIONS +594% Per 1,000 attempts: 5.1 (2025) → 35.2 (2026). Not in Square’s published enum; our own note says “seen live on instalment-flagged cards” — which points at how charges are submitted, not at customer finances. Timing shows a climb with an inflection in Sep-Oct 2025: 0.2 (Jan 25) → 8.0 (Jul 25) → 18.9 (Sep) → 22.9 (Oct) → 39.7 (Jul 26). INVALID_ACCOUNT and INSUFFICIENT_FUNDS step in the same two months. Whatever changed around Sep-Oct 2025 is the single most valuable unanswered question in this whole area. NOT diagnosed.
6.5% of SCHEDULED invoices hold NO CARD ON FILE 714 of 10,973 /first-miss-nocard, 16 Sep 2026, independently rebuilt from 30,000 invoices. They cannot auto-charge and produce no decline code, so no bucket and no list ever sees them. They recover 6.3% next cycle against 28.2% — worst on the book. And only 14 of the 100 soonest had never paid a penny: 86 were paying normally and their card came off and nobody noticed. That is silent attrition, not a new-customer problem. Never tell them their card was declined.
detail is worthless It just echoes the code "Authorization error: 'GENERIC_DECLINE'" There is no human-readable sentence to surface to a collector. The code → message mapping has to be ours, and it is: report_square_declines.js maps ~30 codes to 7 actions — MAXED / RETRY / NEW_CARD / RE_KEY / SPLIT / CALL_BANK / CHASE, unknown falls to CHASE, never to silence.

The card the customer hands over is the strongest single signal on the book — and it is chosen, not assigned. Best card class to worst is 38 points on collection rate; kids vs adults is 5.8. A sixfold difference in the size of the two axes.

Collection rate by card class × kids/adults — 21 Sep 2026

Card classKidsAdultsAllNever paid
Traditional credit 62.7% (376) 54.5% (553) 57.6% (929) 22.8%
Traditional debit 49.9% (820) 46.9% (1,613) 47.9% (2,434) 29.4%
App-bank credit 23.1% (37) 23.3% (103) 23.3% (140) 56.4%
App-bank debit 23.2% (150) 18.6% (487) 19.6% (637) 62.5%
No card on file 7.7% (33) 14.7% (86) 12.6% (119) 75.6%

“No card on file” here is 119, not 384. The first build read a card feed that stopped at 31 Aug, so every agreement whose first instalment fell in September looked card-less — and 124 of those 384 had already PAID. Every agreement now carries a cardStatus where only none may be read as an absent card.

1

IT IS THE CARD, AND IT IS NOT CLOSE

Best card class to worst is 38 points. Kids vs adults is 5.8. A sixfold difference in the size of the two axes. Re-scoring kids on the ADULT card mix moves 47.7% → 43.7%, so ~4 of the 5.8 points is the customer and the rest is that kids customers hand over better cards (adults take app-bank debit 15.8% of the time vs kids 9.7%).

2

IT IS THE BANK, NOT THE PREPAID CARD

A normal, non-prepaid app-bank card fails 76.1% against 39.7% for a normal high-street card. Prepaid adds ~10 points on top (87.3%) and ~20 at a traditional bank (61.8%) — a separate, smaller effect.

3

CREDIT vs DEBIT, app banks stripped out

So it is not the same finding twice: credit 27.2% fail vs debit 44.7%. Best-to-worst card is 27.2% vs 82.5% — threefold, decided at the till.

4

App banks are 19% of instalments and 32% of the misses

81.1% fail vs 40.1%. And they recover worse — back within 10 days 18.8% vs 25.2%, ever 28.3% vs 43.3%. Failing more AND recovering less compounds. Per person on the recurring book over 12 complete months: 19.8% of clients (3,500 of 17,646) hold at least one; 15.7% (2,765) have nothing else. Nashville 27.2% high, Florida 13.9% low.

5

Named banks beat the class

Sutton (Cash App / Chime) 308 / 87.3% fail · Green Dot/Bonneville 95.7% · Bancorp 112 / 76.8% · Stride 99 / 74.7% · SoFi 82.6% — against Navy Federal 55.7%, Wells Fargo 44.5%, BoA 40.1%, and Chase DEBIT 382 / 36.6% vs Chase CREDIT 124 / 6.5%. On the collection-rate matrix: SUTTON 318 agreements at 16.8% collected, 64.8% never paid, against Chase 138 at 75.7%.

6

THE CARD IS CHOSEN, NOT ASSIGNED

Someone able to offer a high-street credit card is already the better payer, so most of the 37-38 points is the customer showing through the card. It is an indicator at the till, not a lever of that size. Asking for a different card does not turn one customer into the other. The real, smaller prize is customers holding both who hand over whichever is nearest. Any rollout must be a trial with a control arm.

7

App-bank is TWO populations, not one

Of 2,589 app-bank customers on the recurring book since April: 853 (32.9%) have ever had a charge succeed and run at 43.5% — basically traditional-debit territory, and their failures are the most funds-driven of any group. 1,736 (67.1%) have NEVER succeeded once — 0.0% across 7,398 attempts. More attempts on that card is not the missing ingredient; the card does not work for us at all. The fix is a different rail (a different card at the till, or ACH), not a different schedule.

8

WEEKLY CHIME PLANS HAVE NO MECHANISM

Weekly rests on “they cannot find $250 but could find $60 four times”. That only bites if the failures are money-shaped. They are not. App-bank failure mix is TRANSACTION_LIMIT 44.2% / PERMISSIONS 21.2% / GENERIC 16.9% / INSUFFICIENT_FUNDS only 10.3% — and the limit reason is FLAT across amount bands ($0-50: 37.2%, $250+: 40.2%). A per-transaction size limit would thin out at $50. It does not. So it is structural to the account, and a weekly schedule would hit a count cap four times as often. Also unpriced: every report in this repo models instalments as MONTHLY.

9

THE EARLY-WAGES THEORY IS DEAD

Do not move charge dates on it. Across 78,016 charges both classes peak Friday and trough Sunday; only the SIZE of the swing differs (1.78× best-to-worst day vs 1.14×). These accounts run empty faster, they are not on a different clock. Worst cell is Monday 12.9%.

The clock is the most actionable thing in this whole pack. Day 1 is 48% of everything a missed opening ever returns inside ten days, and our fastest text leaves at 23.2 hours. The first text lands at the end of the most valuable day, and that is structural, not sloppy.

WhenWhatBasisRead it this way
Day 1 48% of everything a missed opening ever returns inside 10 days 1,099 of 11,475 people, $338k of $638k. /opening-payment, 16 Sep 2026.
Day 3 71% cumulative Day 3 is the realistic call queue — 86 of the ~102 weekly misses are still out at day 3.
Day 7 92% cumulative Days 8-10 add only 180 people. A process starting on day 5 has forfeited 84% of it.
Day 10 → day 30 +20.1% → 26.4% the long tail Still worth a ladder, but an order of magnitude below the first 72 hours.
Day 21 the cliff 1 miss → next cycle pays 20.8%; 2 misses → 12.9% That is why the drafted ladder puts its hardest message at day 21 — before the second miss lands.
What a caught miss is worth $922 per agreement by day 365 $1,165 vs $243. Recovers 88% of a clean payer ($1,327). The gap survives stripping the opening charge out of both sides (0.39× at 365 days), so it is behavioural: a missed opening is not one lost payment, it is a different agreement.
Our fastest text 23.2 hours median 24.4h, nothing inside 12h ever Structural — the feed into the debt GHL is a daily batch. Day 0 does not exist and cannot exist until the feed is event-driven.
The worst day of the month to charge the 31st 43.3% success against 50.0% overall And it is simultaneously the most-chosen charge day (9.1% of agreements). 28th+30th+31st ≈ 19% of agreements, all bottom-four. Best: 2nd 54.4%, 1st 53.5%, 16th 53.4%. Measured across 451,285 CRM charges. Evidence supports moving month-end charges to the 1st-3rd — but charging AHEAD of a stated payday did NOT hold (34.5% vs 37.5%, heavily confounded). Not actioned; rescheduling real charges is outward-facing.
Month-end lull is REAL, not a broken cron 29th/30th near-empty, 31st dumps Jun: 57 / 868 / — · Jul: 14 / 15 / 870 · Aug: 1 / 2 / 929. Two near-empty chase files at month-end are correct. undefined
The week is not flat Friday 29.9% of arrivals Thursday only 8.4% Mon+Fri = 48.3% of the week. 17 complete weeks: mean 97.3, median 98, range 43-144 — a 3.3× spread. A rota on the average is underwater one week in four.

They are not three problems. The decline-reason mix is ~85% soft in all three. They do not differ by why the card failed — they differ by how long ago anyone asked, and whether anyone asked at all. So it is one contact engine with three entry points and three economics, not three collection strategies. And priority is not size.

PopulationDefinitionHow manyHow muchRead it this way
1 — New The opening instalment 5,127 fail of 10,943 a year (46.9%) $1,136,232 missed at the door Fresh misses, agreement ≤ 9 days old. ~102 a week, ~21 a working day. Worth $922 more per agreement by day 365 if caught inside ten days. This is the priority and it is not the biggest number.
2 — Early life Opened 10-60 days ago 849 of 1,599 have missed $254,182 Mostly Population 1 arriving late. Opening failed → 72.4% still on a miss streak, avg banked $84. Opening paid → 8.1%, avg $320. Nine times the failure rate and a quarter of the money. 503 (31.5%) have never paid a cent — $161,930 attempted and refused, 84.9% soft. These are sales that never started, not debtors. Studio voice, not collector voice. 353 sit at exactly ONE miss right now — the only intervention point that matters.
3 — Zombie 4+ consecutive fails 9,801 accounts (2,867 still charged) $14.6M; $611,841 a cycle into the void The biggest number and the WORST place to spend the next hour. Splits into 5,114 who paid before going quiet (banked $3.06M, avg $599 — but 1,906 made exactly ONE payment, a first-cycle failure in zombie’s clothing) and 4,687 who never paid a cent ($6.69M that only ever existed as an attempted charge). 1,469 of the live-and-still-charged have a payment history — that is the settlement campaign.

Measured 17 Sep 2026, square_attempts*.jsonl, data to 16 Sep. Non-overlapping definitions: new = opening failed AND agreement ≤ 9 days · early = age 10-60 days AND streak 1-3 · zombie = streak ≥ 4, any age. Two hazard curves exist for the same thing and they differ by level, not shape — the charge-attempt view (44.0% → 23.5% → 16.2% → 14.6%) and the invoice view (20.8% → 12.9% → 7.9% → 5.0%). The invoice view is lower because it counts cycles where no charge was ever attempted (13.3% of the zombie pile). Name which you are quoting.

Nine desks, GHL boards, GoTo phones. The single most important thing on this tab: Square records WHO EXECUTED a transaction and never WHO CHASED it, and calls do not write back to GHL at all. So every per-collector number in the estate is a floor, an upper bound, or both.

1

The account is NOT quieter — the SPLIT moved

Inbound texts per 30-day window: 2,184 → 2,374 → 2,203. Flat. Underneath: Maria R 151 → 610 (+304%, now 27.7% of every reply) while Caet S 258 → 149 (−42%) and Charles C 367 → 238 (−35%). So the collectors saying they are quiet are right about their own inbox and wrong about the account. Never answer “are we getting less work” with an account-level total.

2

Money is flat and the AUTOMATION absorbed the fall

Jun→Aug 2026 total collected $846,875 → $808,619. The automated run +19.7% ($385,656 → $461,743); every human desk −24.8% ($461,219 → $346,876). Desk logins fell hardest: Nashville Collections 121 → 49, Boston 118 → 57, Studio1 388 → 227. Maddy Brener took ZERO collections payments Jun-Aug — she is reporting on the team, not from her own numbers.

3

What counts as “automation” is BEHAVIOUR, not a name

A hardcoded 13-first-name allow-list was filing $1.4M — more than half the book — as “automation”. The collector is now the Square team-member name, full stop. An account is the batch run only if it took ≥20 payments inside a single minute on ≥10 separate days — both halves required. Over 20 months only Natalia Accounting (49/min on 25 days) and Studio1 Collections (37/min on 15) clear it. The busiest human peaks at 12/min, 40/day, and takes days off.

4

Caseload and live conversations are near-unrelated

Caet holds 3,491 accounts / 98 live · Maria 957 / 277 · Kimberly 148 / ZERO. The average balance barely moves across the five contact rungs ($1,866 on a live conversation vs $1,720 on one that has never answered) — there is no balance threshold to sort the book by. It sorts by contact and by nothing else.

5

Read a collector’s book by STAGE, never by headline rate

Emily’s own Debt(Emily) stages — accounts already in conversation — are 115 accounts and she rang 68 (59%). The Debt AI → No reply backlog behind it is 465 accounts and she rang 48 (10%). And 236 are settled, dead or placed with the agency — she rang almost none, which is the right instinct. “She works 13% of her book” is a real number and a useless one.

6

What a fortnight of dialling actually looks like

Emily, 8-18 Sep 2026, 9 working days: 256 dials to 163 people, 2.9h talk, ~28 dials a day. Only 15 of the 163 stayed on past two minutes. 99 got to 30 seconds; 23 never picked up. The shape is the finding: 87 dials — the single biggest block — ended between 30 and 59 seconds, which is a voicemail greeting plus a message. 30 seconds is the honest ceiling on what call data can tell you — there are no transcripts (org is on recording v1, API access OFF). Whole desk: 20.2 conversations ≥2 min a day, 3,395 calls over 18 active days.

7

DID THE MONEY FOLLOW THE CALL? No — and entry method proves it

Of the 8 people who paid after Emily rang: 7 were ON_FILE, 1 was KEYED. Median wait 4.1 days, nothing inside the first hour. A desk taking cards on the phone produces a cluster at 0-30 minutes, keyed. There is no such cluster. The control agrees: rung 9/116 = 7.8%, never rung 49/531 = 9.2%. Where the groups differ is the ATTEMPT column — a card was tried on 53.4% of the rung group vs 40.9% of the rest. She rings people whose charge date is near — sensible targeting, and exactly why their payments land soon after a call whether or not the call did anything.

8

Same finding on the zombie pile

Of money landing after a 4th miss: 852 (68.7%, $148,950) was the schedule firing on the stored card — nobody touched it. 389 (31.3%, $55,089) was a person putting it through. The invoice is what collects it, so cancelling removes the mechanism, not just the paperwork.

9

Square records WHO EXECUTED, never WHO CHASED

Proved and closed: Square invoices do carry creator_team_member_id, and crediting by invoice sender changes totals identical to the cent for every collector — because creator == processor in this workflow. Do not build it. Fixing attribution needs a source of truth for “who chased”: a CRM activity log, or crediting by assignment. Handle sensitively — the data cannot distinguish “her work happened on the phone” from “her list is accounts she manages”.

10

KEYED is the reliable human signal

card_details.entry_method: KEYED = a human typed the card in; ON_FILE = stored card. Automatic charges also carry a team id, so the team id alone does NOT mean human. July 2026: 613 KEYED payments into Recurring = $138,105, 16.5% of all recurring money; all 32 locations 4,002 keyed = $517,444 = 22.8% of July. That $138k is a FLOOR — a collector charging a card already on file shows as ON_FILE.

11

The invoice-based collection rate cannot see phone payments

A collector taking payment over the phone creates a separate order, so the original invoice stays UNPAID and the rate never credits it. July 2026: billed $1,981,902, marked paid on invoice $763,317 = 38.5%, but actual cash into Recurring $839,461. $76,143 the rate cannot see; true rate ≈ 42.4%. OPEN: if the mix has shifted over time from auto-charge toward phone payments, part of the 55.9%→39.8% “collapse” is an artefact of the measure. Only July is measured — do not assert it is an artefact until the month-by-month gap is computed.

What has actually been run, and what it returned. Read the 21 Aug sweep first — it is the only clean, unselected conversion number in the whole area, because it swept a near-complete population rather than the people a collector had already talked round.

WhatResultRead it this way
The 21 Aug sweep — the only clean conversion number we have 25% converted, $97.90 each 95 cohort members swept, 84 of whom had not paid since their decline (the real target state). 21 converted = $2,055.86, $24.47 per account swept. Near-complete population, so no selection bias — this is the one unselected conversion rate in the whole area, and it was executed badly (day 9-11, naive amounts), so it is a FLOOR. By original reason: TRANSACTION_LIMIT 39% / $36.21 per account · INSUFFICIENT_FUNDS 29% · GENERIC_DECLINE 8% · dead-card codes 0 of 7.
CUTTING THE AMOUNT MADE IT WORSE — monotonically full 57% → half 25% → quarter 4% 188 ON_FILE attempts on 100 people. Full amount 21 attempts / 12 taken / $5,275.65. About half: 95 / 24 / $2,358.64. A quarter or less: 68 / 3 / $209.69. The quarter attempts went to people who had just been refused at half — that ordering is clean evidence, not selection. 56 people were refused at half or less and still paid nothing. The reason mix is IDENTICAL across bands, which is the tell that the amount was never the binding constraint.
The ONE exception: step down only on TRANSACTION_LIMIT 3 of 23 = ~1 in 8 60 people refused at half were tried again at a quarter. 3 went through — all three had failed the half on TRANSACTION_LIMIT. GENERIC_DECLINE 0 of 24 · INSUFFICIENT_FUNDS 0 of 6 · dead-card codes 0 of 7. Only step down when the half failed with TRANSACTION_LIMIT, and know it is ~1 in 8. Every other code is 0 for 37.
The do-nothing baseline 34 of 99 self-cured in 9 days Before anything was sent: 17 of 33 long payers, 16 of 33 first-ever decliners, 1 of 33 chronic decliners. Half the long payers self-cured with nothing sent; the chronic group does not move. Any programme measured without a control arm will claim credit for this.
REDUCING A BALANCE DOES NOT UNLOCK PAYMENT $0.12 back per $1 945 accounts reduced since Jan 2026, $999,512 written off, $118,179 collected afterwards. 60.4% paid nothing at all. And it runs the wrong way: $0-250 off returns 96.7%; $2,000+ off returns −0.3% (chargebacks land after the credit) and 75.8% pay nothing. Depth of the cut is the strongest signal: under 25% of balance → $0.74 back · 25-50% → $0.41 · 50-75% → $0.14 · 100% → −$0.03.
Half the write-off is a credit that CLEARS the balance 396 of 945, $511,431 51.2% of everything written off, and those can never show recovery by construction (99.7% paid nothing after). Blending them into the headline is what makes recovery read as a flat failure. The 549 part-cuts wrote off $488,082 and collected $135,351 = $0.28 back per $1 — more than double the blended figure. Worst combo: $2,000+ off with a balance still owing — 82 accounts, $195,393 given up, $12,566 back.
The money was already on the table 200 of 381 paid the SAME DAY Of the 381 that ever paid again after a reduction, over half paid the same day as the reduction. Almost nobody starts paying later. And 600 of the 945 had no failed charge at all — we reduced accounts that were never fighting a card (72% of the money, $0.08 back per $1). The ones with a real card problem recover best: insufficient funds $0.52, needs a new card $0.42, call the bank $0.46.
A DEEP RUN IS NOT DEATH — the one encouraging number 19.7% later paid again Of the 11,622 accounts that have ever hit a 4+ consecutive-miss run, 2,286 later paid again — $1,024,771 banked, avg $448 over 2.7 payments. Every one of those was a person getting hold of them. But it does not stick: of 803 recoveries with 6 further cycles billed, 30.1% paid 0 of the next 6, only 8.2% paid all six, and 33.3% went on to hit another 4-in-a-row run.
Where a shut-off threshold should sit NINE misses, not four Cut at 4: 2,656 series, $544,275/mo of billing stopped, 10.2% of those runs later saw money, $391,740 refused. Cut at 9: 904 series, $193,628/mo, 4.5%, only $27,306 refused, $4.67M of dead billing killed. 68.4% of all recoveries land at exactly 4-5 misses, while a collector is still working the file. Counterweight: 264 series (9.9%) took a card payment during the run that never touched the invoice — $66,499 off-rail. Nothing has been cancelled.
Outreach had NO lift on first payments and HALF the chargebacks The separate outreach programme moved neither first-payment success nor collection. See /outreach-report. Do not fund a chase programme on an outreach analogy.

Thirteen things. Four are BLOCKING — nothing built on top of this reaches its stated value until they are fixed — and the rest are live defects or data traps that will produce a wrong number for anyone who does not know about them.

WhatSeverityThe detail
The decline reason never reaches GHL at all BLOCKING The outbound text is identical whatever the issuer said. There is no decline-reason field in the debt sub-account — 23 contact custom fields, none reason-related; of 881 tags the only card-ish ones are card_issue and card handling - <person> (human routing). The 11 reason-specific messages are written and sit in code (push_square_declines_ghl.js, field sq_decline_message) — they need ONE SMS step in the GHL UI whose body is {{contact.sq_decline_message}}. GHL’s API cannot create or edit a workflow, so a human must add it.
Day 0 cannot exist while the feed is a daily batch BLOCKING Fastest text 23.2h. Day 1 is 48% of the return. An event-driven feed is the single highest-value build item in this area and every value estimate for the ladder depends on it.
A dropped day is PERMANENT — nothing backfills LIVE RISK Over 6 Jul - 13 Sep, 20 of the 25 absences came from TWO DAYS. Fri 4 Sep: 15 of 15 still-owed new people, ZERO created in GHL, $2,988.56, and 14 of the 15 have never paid anything since — still absent 13 days later. Sat 12 Sep: 5 of 6. Every other day in ten weeks: 0 absent. On both days our file was built and held those customers — what never happened was the LOAD into GHL. The failure mode is a lost day, not a rate. Current 28-day window (2026-08-26 → 2026-09-22): 1 whole day lost, 20 people never reached GHL ($3,841).
The /pay portal promises a bonus credit it never applies LIVE DEFECT Tiers: Full +15% · 75% +10% · 50% +5% · 25% 0 · Goodwill(10%) 0. The screen says “$X credited to your balance” and the balance does not move — it is written into the Square payment note and a GHL custom field, with no CRM write, no ledger write, no Square credit. Latent, not an existing liability: scanned all 83,623 recurring payments taken in 2026 across six accounts — zero carry the AI Payment Portal note. Fix before any traffic. The portal also shows the customer a miss-1→5 escalation ladder promising “legal action and impact to your credit rating” at miss 5 — verify that is true and compliant.
notifyGHL points at the WRONG sub-account LIVE DEFECT It reads ghl_config.json, which holds XcjF95QScPf2rxRSgp2e (“AI Debt Collection”) — a near-empty account with 1 draft workflow. The real one is pcTyOadbzBqZO7v3cbwn with 55 workflows, 12,657 opportunities, 881 tags. Check the location id before writing anything. The working token is ghl-inventory/.token or GHL_USDC_TOKEN; ghl_config.json’s token gets a 403 on the real one.
The CRM’s own In Collections export is broken LIVE DEFECT /application/diary/reports/csv-incollections.php returns HTTP 500, header row only — verified against every shape its own form can produce. Not the session, not the parameters: Debtor Ledger, Default Ledger and Credit Control all return rows through the same helper in the same session. We rebuild it byte-for-byte instead (build_in_collections_file.js, md5 83c8e011… on the header).
GoTo call data does not write back to GHL STRUCTURAL 0 calls appear in any GHL thread. A phone-worked card reads as unchased. Also: cr.v1.read started 403-ing on 17 Sep 2026 and is still 403 — goto_call_pull.js and /Studio1_GoTo_Call_Activity.html are frozen at 16 Sep. Leading hypothesis is a lapsed Advanced Reporting add-on; check billing before re-running OAuth. A different endpoint still answers: GET /call-history/v1/calls, per user, and it carries the dialled NUMBER.
GHL assignedTo is NOT maintained DATA TRAP 219 opportunities sit in a stage literally called “Assigned to you” and not one has assignedTo pointing at the collector — 212 still point at Caet S. So every collector’s book is UNDERSTATED and every “% of book worked” is an upper bound. It is also current state at snapshot date, not ownership on the day of the call.
The filing is wrong in BOTH directions DATA TRAP 611 accounts filed on Debt AI(Collected) still owe $744,080 — 34% of that board, and it is accelerating (88 cards / $79k when last measured on 28 Aug). Nobody chases them, because the board says they are done. The reverse too: 155 settled accounts sit on a Collected board ownerless, 170 sit on a Problems board owing ~$292k. Whatever moves a card to Collected is not reading the ledger.
1,448 accounts cannot be texted and the tag only knows 432 DATA TRAP GHL dndSettings.status: "active" means the DND is ON — the channel is BLOCKED. Reading it the natural way round said 162 against a real 1,448 (11.5% of 12,637). 1,301 SMS:active + 107 SMS:permanent (replied STOP, cannot be undone) + 40 master dnd:true. The dnd enabled contact tag holds only 432 of them — count the FIELD, never the tag, or suppression logic silently texts ~1,000 people who said stop. Nobody is call-blocked or email-blocked by a channel setting.
Square cannot end a recurring series through ANY API HARD LIMIT POST /v2/subscriptions/search returns {} on all six recurring locations — this billing is not Square Subscriptions. The Invoices API has no recurring-series object at all. Ending one is the Square Dashboard, by hand. The API can cancel ONE invoice (the series mints a replacement next cycle, so it is a nightly sweep forever) or disable the card (permanent, no re-enable endpoint, and it takes the card off the collections team). Neither is switched on.
Two outward pushes are BUILT but still DRY RUN AWAITING NEIL 1. The daily Square-declines → GHL push (push_square_declines_ghl.js, ~176 texts from ~236 declines a day, runs a day behind deliberately). Blocked on the GHL SMS step above; going live now would enrol people into workflows sending the OLD miss-count copy. 2. The CRM → Square customer import (square_customer_import.js) — verified on 409 bookings, 0 failures, idempotent on reference_id; needs "args": ["--live"]. Running that import live is the one thing that would unlock source attribution on Square money (today all 45 import logs are live:false).
The chase handover ledger CANNOT be rebuilt IRREPLACEABLE chasedaily_file_handover_ledger_json in Supabase is the only record of who was handed over and when. declines.json is a rolling 5-day window, so a lost ledger means re-chasing thousands within a week. It keys on id + phone + email — the same person exists under two Square customer ids when they have paid at two city accounts.

Sixteen claims that are wrong, withdrawn, circular, or true-but-misleading. Several of these were published before they were caught. If a number on a new dashboard matches one of these, the dashboard is repeating a known error.

The claimWhy it is wrong
“58% of the debt book has engaged with us” 6,706 have replied at some point and the flag never expires. 4,551 of them last spoke over 90 days ago. The live figure is 8.3% (968 accounts).
“Reply rates have halved” They have not. First-month conversations answer at 51-57% and have not moved in a year. The blended figure fell because the re-chased pile grew twelvefold and answers at 12-13%. A mix effect, not a decline.
“The collection rate collapsed from 56% to 40%” LIVE payers held at 70.6% → 65.8%. The DEAD pile (no payment 3+ months, still auto-billed) doubled and went from 36.6% to 53.7% of billing. Nobody performed worse. We now bill MORE to non-payers than to payers.
“Step-down payments convert at 77% vs 43%” Charges for MORE than the failed amount convert at 73.5% — just as well as the smallest step-down. If shrinking the amount were the mechanism, asking for more could not work equally well. What separates the groups is that a changed amount means a human agreed it. And the controlled sweep found smaller was monotonically worse.
“Worked accounts recover 72% vs 3% untouched” Circular. A customer’s recovery payment is itself “an attempt within 21 days”, so worked→recovered is near-tautological. Use the 21 Aug sweep (25%) for the conversion rate and the historical decay only as a RATIO.
“87% of first missers are not broke” WITHDRAWN 16 Sep 2026. TRANSACTION_LIMIT recovers 20.9% at 30d / 38.8% ever against INSUFFICIENT_FUNDS 21.3% / 39.5% — identical. The big code is a money code. What survives is the soft/dead split (87.2% soft, 1.5% dead), which is a routing fact, not a customer-wealth fact.
“We see TRANSACTION_LIMIT on $25 charges while on the phone” WITHDRAWN. 248 of the 253 $25-and-under TRANSACTION_LIMIT declines in 2026 are entry: ON_FILE; only 1.6% of all TRANSACTION_LIMIT declines are KEYED. entry_method is the field that answers “was a human there” — check it before describing a charge as something a person did.
“Chasing adds close to nothing” WRONG and corrected. That came from a 5.6% recovery rate measured on the original invoice only. Collectors take payments over the phone as SEPARATE payments, so they never touch the failed invoice and are invisible to an invoice-status view. 77-83% of collector money lands on accounts with a missed invoice = genuine recovery, ~$149k/month. Stopping the team WOULD lose ~$150-178k/month.
“Most post-4-miss money is collectors rescuing accounts” WRONG and corrected. 68.7% is the schedule firing on the stored card on a LATER cycle. This does not break the no-retry rule — nothing re-charges a failed invoice; what succeeds is the next invoice.
“384 agreements have no card on file” WRONG — the real figure is 119. The card feed defaulted to the last three COMPLETE calendar months, so in September it stopped at 31 Aug and every agreement whose first instalment fell in September looked card-less. 124 of the 384 had already PAID, which is proof the card existed. The tell was the SHAPE — 141 of 384 were August shoots; a genuine capture failure is flat across months. “No card FOUND” is not “no card ON FILE”.
“Declines doubled year on year” Nearly published. The file that looked right had a 2025 side under-pulled tenfold (~1,400/month vs ~14,500). A tenfold step in one month is a pull artefact, not the business. Same family of error: both Square feeds are APPEND-ONLY and MUST be deduped on payment id — unduped, Aug 2026 read 22,117 attempts against a ~15,500 baseline and turned a +32.3% rise into “+81.8%”.
“Miss #2” from the R-number 004720-R-0001’s digit is the recurring period sequence — invoice 12 of the plan — not how many payments were missed. Agreed with real history on only 12 of 41 Boston rows and is wrong in both directions. The correct count is that customer’s payment_requests with an earlier due_date whose status is UNPAID/FAILED/CANCELED/PAYMENT_PENDING, +1.
Any total that includes CANCELED invoices A cancelled plan’s remaining invoices sit at CANCELED and are not debt. Counting them added $1.52M of fake arrears and 834 fake misses to one July figure. Miss/debt status set is UNPAID + FAILED only.
Any exposure summed per ROW Exposure must be per CUSTOMER, counted once. Summing “arrears before this miss” per row re-adds the whole balance for anyone who missed twice in the month, and each later miss’s arrears already contains the earlier one. Same error overstated the ownerless book by $327k by summing per CARD instead of per ACCOUNT.
“Negative balance” as over-collection 107 agreements go negative purely because a credit larger than the finance was posted ($34,773), and 10 more carry payments against an agreement with no finance. Bookkeeping shapes, not money taken. And of 15 exact-double cases, Square corroborates only 3 — the other 12 are CRM-side double postings.
A build stamp, an HTTP 200, a file mtime, or a green push All banned as proof of freshness. /debt-trend stamped itself with today over 10 Sep data for eleven days, because it is owner:"laptop" and read a feed written by an owner:"render" job into a gitignored, non-Supabase directory. When a report is owner:X, every feed it reads must also be refreshed on X.

Everything a dashboard would read, and every key it would join on. The join keys are exact — no name matching and no fuzzy matching is needed or wanted anywhere in this area.

The join keys

From → toThe keyHit rateThe detail
Square recurring invoice → CRM agreement invoice.title IS the CRM ref 99.9% (3,597 of 3,600) Probed 600 invoices × 6 recurring locations. "D3845375" is the Debtor Ledger’s ref, repeated in description and primary_recipient.company_name. The invoice ALSO carries phone and email, so a recurring invoice pull needs no /v2/customers lookup to reach a person. No name matching and no phone matching is needed anywhere between Square invoices and the CRM — do not add any.
GHL card → CRM agreement companyName IS the order reference 1,849 of 1,852 GHL has no reference field. The update business name from reference code step of Assign states for debt lead writes it into companyName, and /opportunities/search returns the whole contact inline — one opportunity sweep gives both the card and its money key.
GHL contact → Square customer PHONE only never name, never email Re-probed on raw digits, last-10, email and surname — email and name have never added a single extra match. Bulk index: GET /contacts/?locationId=&limit=100 paginated on startAfter+startAfterId pulls all ~12,491 debt contacts in 125 pages, giving phone → {id, owner, dnd} in one pass. Reuse that, never probe per person.
Phone → Square customer POST /v2/customers/search with phone_number.exact 46% → 96% coverage Run against all six city accounts. That search alone doubled coverage over the local caches.
Decline → invoice order_id, never customer+date 4,950 of 5,165 exact Over a 3-day window date matching is safe; over a month a customer has several misses and several second attempts and it attaches the wrong reason to the wrong invoice.
Square card → issuing bank BIN → data/chase_lookups/bin_issuers.json 99.9% e:1 IS the app bank flag. A quarter of customers used more than one card, so an agreement takes its MODAL bin from the shoot date on. The attempts file includes the 140k FAILED rows, which is what stops any rate off it being circular.
Studio → Square account Derived from data, never the ref prefix 1:1 Studio 1 = FLORIDA (M) and Orlando (F) → florida · Maps A (P) → newyork · Dallas (D) + Houston (H) → texas · Boston = T · Nashville (N) · Chicago (C). Icon + Soul = ATLANTA, closed 2024, no Square account. Boston being T is the one nobody guesses right.

The feeds

FileWhat it holdsWritten byThe trap
Reports/finance_cohort/ledger.json CRM Debtor Ledger — finance, paid, credits, balance per agreement pull_debtor_ledger.js (finance-cohort-ledger, daily, laptop) 14 studios × 2019→now = 112 posts, ~5 min. studioid is REQUIRED and takes ONE studio at a time — the <select> is name='studioid' in SINGLE quotes, so a name="…" regex silently skips it and it 500s with a header row. This must run FIRST; everything downstream reads it.
Reports/finance_cohort/schedules.json Per-agreement dated payment lines pull_finance_schedules.js (daily, laptop) PaymentUIClass.loadByBooking. bookingid == the Order Reference minus its letter prefix (M2908606 → 2908606). Counts payment_debit; payment_declined never collected; payment_order is the deposit at sale. Dated sums MUST include the NEGATIVE payment_debit lines — positive-only reads 3.5% high, and that gap is refunds. Cold backfill is 59k fetches (>1h); incremental after, keyed on paid+credits moving.
Reports/finance_cohort/square_attempts.jsonl EVERY Square card attempt since 2025-01 WITH the reason code pull_square_declines_history.js (square-attempts-history, daily, laptop) 264,022+ attempts, all 6 cities / ~30 locations, ~45 min cold. APPEND-ONLY and it rewinds a day per location, so it MUST be deduped by payment id on read. amt is DOLLARS, not cents. Refresh it before quoting any recent week — a pull moved one day from 540 to 1,077 attempts.
Reports/card_types/card_attempts.jsonl The same attempts PLUS card_type (CREDIT/DEBIT) and prepaid_type pull_square_card_types.js (card-types-pull, daily, laptop) square_attempts.jsonl CANNOT answer debit-vs-credit — it never kept card_type. This purpose-built feed does. It defaulted to the last three COMPLETE calendar months and that produced a wrong headline; extended to today. Also append-only, dedupe by id.
Reports/square_declines/declines.json A rolling 5-day window of declines with reasons report_square_declines.js (square-declines, daily, render) Never a history of anything. The permanent record is square_attempts.jsonl. Reports/square_declines/ is gitignored (PII), so per-day files never reach Render by push; Render’s cron rebuilds only the last 5 days.
chasedaily_file_handover_ledger_json (Supabase) Who was handed over to GHL, and on which day report_chase_daily.js (chase-daily, daily, render) THE ONE FILE THAT CANNOT BE REBUILT. Keys on id + phone + email. owner:"render" deliberately — a laptop copy would write the ledger twice and empty the days the first copy claimed. It no longer CUTS anyone (fixed 3 Sep 2026): a day’s file is everyone who missed that day. A day still carrying skipped was built under the old rule and is short by that many.
The debt GHL — pcTyOadbzBqZO7v3cbwn “US Debt Collection”: 55 workflows, ~12,657 opportunities, 881 tags pull_ghl_debt_workload.js / pull_ghl_debt_unassigned.js / pull_ghl_debt_sms.js ghl_config.json points at the WRONG place (XcjF95QScPf2rxRSgp2e, near-empty, 1 draft workflow) and its token 403s on the real one. Working token: ghl-inventory/.token or GHL_USDC_TOKEN. Workflows are READ-ONLY via API — no create/update/delete, every detail path 404s, the canvas is UI-only. Cards and boards ARE fully writable, including creating and deleting a board. Contact enrolment works: POST /contacts/{id}/workflow/{wfId}.
goto_calls_daily.json Per-agent per-day call aggregates goto_call_pull.js (FROZEN at 16 Sep 2026) The agent keys are the SHORT names — Nick, Emily, Maria, Charles, Collections 1, Collections 2 — NOT the long GoTo LINE names. Using the long names returns zero for every day and reads as “the team made no calls”. Day-bucketed in America/New_York. Per-call detail with the dialled NUMBER is GET /call-history/v1/calls, still working: direction is INVERTED, there is a silent 100-item cap with NO page marker, and duration is milliseconds of talk time.
Reports/note_issues/raw.jsonl CRM notes — the free-text record, incl. pasted email correspondence report_note_issues.js 439,694 lines but only ~43,274 distinct bookings — APPEND-ONLY, so any count must dedupe by booking (last line wins) first. A payday scan read straight through and came out 8× inflated. Notes live in the CRM’s allergies field. Note date stamps are DD/MM before 2022 and MM/DD after — it is the ERA, not the author; ~40% are ambiguous.
data/ghl_debt_dnd.json Who cannot be texted scripts/ops/ghl_dnd_count.js (read-only) Paginates all ~127 pages, refuses to report a short pull as complete, and dumps rawStatusValues so the polarity can never quietly flip again.
data/collection_rate_segments.json Per-agreement card class, issuer, kids/adults, city pull_collection_rate_segments.js (daily, laptop) Chain: ref → Square customer (temp/recurring_invoices_cache.json, whose invoice titles carry the CRM ref, 95.2%) → attempts → BIN → issuer. Only network call is a read-only diary pull (8 studios, ~30s). The invoice cache is READ, never refreshed — refreshing it is a ~35 min Square sweep owned by paid-off / zombie-reasons. Must run between the ledger pull and the report.

Square API facts that cost time to learn

1

/v2/payments returns the DEFAULT LOCATION ONLY

Pass location_id explicitly and loop every location. Not doing so undercounted 47k against a real 151,318.

2

Match the recurring location on the NAME with /^Recurring/i, never ===

New York’s is called Recurring NYC, and an exact compare silently drops the biggest city. Ten reports did this. Location ids: texas LE6G3FYRFNVFB · boston L3TZS37RCY0QY · newyork F29PDT5T9ZPB7 · florida D22KRWPAJQ3EE · chicago LKH1CNCTW9PNN · nashville LR0G5P980H0KE.

3

Paging /v2/invoices SILENTLY TRUNCATES

A 100-page cap = exactly 20,000 invoices per city, and five of six cities hit it dead-on — 20,000×5 + 8,238 Chicago = 108,238, the exact total the old report always printed. Real totals: NY 165,682 (it was reading 12%), Texas 87,342, Florida 84,581, Boston 31,576. Square returns newest-first, so truncation ate the oldest history and long-lived payers looked never-paid. A suspiciously round Square total means suspect a page cap FIRST. Use throw on a leftover cursor, never a silent break.

4

POST /v2/invoices/search takes ONE location and ONE customer

2+ gives VALUE_TOO_HIGH. So it is one query per customer; no batching. ~1,100 calls per week’s payments, ~1 min across six parallel city tokens.

5

23.1% of FAILED payments carry NO customer_id

39,450 of 170,908 — 74.7% of 2023 failures, 28.6% 2024, 15.0% 2025, 11.3% 2026. So any “never attempted” figure is a CEILING, not a count. COMPLETED payments are fine (44 blank of 285,691), so recovery and entry-method analysis is unaffected.

6

Square can issue TWO identical invoices for the same period

Same customer id, same amount, same day, two invoice numbers. De-dupe on customer + due_date before chasing. Days are bucketed in the studio’s timezone, never the server’s.

7

Reports under Reports/ that hold PII are gitignored and ride a Supabase shelf

Render’s disk is ephemeral and wiped on every deploy, so a laptop-built page reaches Render only through Supabase (<slug>_file_* keys) and a restore path in serve_reports.js. Supabase upserts NEED ?on_conflict=key and the status must be checked — an unread one binned 10 days of builds. Only counts-only data/*_summary.json files reach git.

Twenty-one entry points across ~45 manifest slugs. Start at /money — it is the operational page and everything else is depth behind it. Key-gated pages use ?key=studio1-sales-2026; that gate exists only where a page carries names, phones or balances.

What it answersRouteURLGateThe detail that matters
THE ONE PAGE /money https://reports.iconicbyai.com/money open Six tabs: The board proposal · This week · Did we chase them · The conversations · The plan · Who to ring. It is a COMBINER — report_money.js does no data work; it lifts the panes out of four generators’ HTML with depth-counted <div> matching and re-wraps them. It is the LAST manifest entry — re-run it after any of the four or the page shows stale panes. The old URLs 302 here.
The board paper /money#proposal https://reports.iconicbyai.com/money open A decision paper, not an analysis. Asks for 3-4 weeks of one developer, 1.0-1.2 FTE, 12 weeks in 4 phases and seven named approvals: engineering · headcount · settlement capped at 25% opening day 40 · stop charging 2,867 dead agreements · pre-payment contact · a card-quality step at the till · a 10% control group. Case: $57k-$236k a year on the first-payment work. Exempt from the staleness banner — a decision paper does not go stale with age, it goes WRONG when a figure underneath it is re-measured. Re-run it and then report_money.js whenever a quoted number moves.
The plan and its evidence /debt-plan https://reports.iconicbyai.com/debt-plan open Static. The “ruled out” section is the most useful part — three of the first draft’s five actions did not survive contact with the data, and the item now at the top was not in it at all. Every one of those ideas is plausible enough that somebody will suggest it again.
The first-payment evidence pack /Studio1_First_Miss_Handoff.html https://reports.iconicbyai.com/Studio1_First_Miss_Handoff.html open 10 tabs, 113 rows, including the full day−5-to-collections contact ladder (22 steps, 6 forks, 4-rung discount ladder, 6 guard-rails, 8 pieces of copy). Static by design.
Is it getting worse? /debt-trend https://reports.iconicbyai.com/debt-trend open Volume → recovery → texts, every rate maturity-gated. Carries the first-month vs carried-over split that kills the blended reply rate. It went silently stale for 11 days behind a fresh build stamp — the cross-owner feed trap.
Where the desks are /debt-workload https://reports.iconicbyai.com/debt-workload open Boards, caseloads, replies per desk per 30-day window. assignedTo is CURRENT state, not history — a rise is trustworthy, a fall can be a reassignment.
The whole sub-account /debt-status https://reports.iconicbyai.com/debt-status?key=studio1-sales-2026 key-gated Three independent complete states per account — money, contact, owner — so any two cross without a residue. The only page that ever said how big the account is.
Ownerless cards / hand-outs /debt-unassigned · /debt-handout https://reports.iconicbyai.com/debt-unassigned?key=studio1-sales-2026 key-gated Join on companyName, count ACCOUNTS not cards, and only ever-replied debtors get handed out (Neil, 27 Aug: “if they have never given a reply they stay in unassigned”). The writer is move_debt_cards.js — dry run by default, every applied move logged so --undo puts it back. Those undo logs are gitignored and exist only on the laptop.
Collection rate, age-matched /collection-rate https://reports.iconicbyai.com/collection-rate open Collected ÷ what has actually FALLEN DUE, every cohort compared at the same age. Reads 43.7% where the lifetime basis reads 9.9% — same money, same contracts. Carries the kids × adults × card matrix.
Billed vs collected /collections-health https://reports.iconicbyai.com/collections-health open Two known structural blind spots, both making collections look WORSE than it is — collector-vs-auto decided by first name, and an invoice-based rate that cannot see phone payments. Not altered, on Neil’s instruction.
Who collected what /collections-monthly https://reports.iconicbyai.com/collections-monthly open daily:false — a twenty-month all-city /v2/payments pull is ~10 minutes and the answer moves once a month. Its HTML is TRACKED IN GIT and that commit is the only way it reaches Render — the one report where committing generated HTML is correct.
Why the charge failed /decline-buckets · /decline-trends · /square-declines https://reports.iconicbyai.com/decline-buckets?key=studio1-sales-2026 key-gated Structural / hard / soft, plus what was charged afterwards. /square-declines is the one that carries the issuer’s own code and the 7 action buckets, and it feeds the (dry-run) GHL push.
The card / the bank /card-types · /card-issuers · /payer-types · /due-next-7 https://reports.iconicbyai.com/card-issuers open Debit vs credit, BIN → issuer, never-paid / intermittent / always-pays buckets, and the FORWARD list of everything falling due in the next 7 days — the only forward-looking page in the area.
The opening instalment /opening-payment · /first-payments · /instalment-recovery · /first3 https://reports.iconicbyai.com/opening-payment open The three outcomes, the 3-month vs 12-month-prior comparison, the 1st-vs-2nd instalment cross, and the first-three-payments cohort. /first3’s cache FREEZES statuses at pull time and does NOT self-heal — audit P1 AND P2 AND P3.
The zombie pile /zombies · /zombie-reasons https://reports.iconicbyai.com/zombie-reasons?key=studio1-sales-2026 key-gated /zombies counts the pile; /zombie-reasons carries the REASON, the hazard curve and the shut-off mechanics.
Settled but still billing /paid-off https://reports.iconicbyai.com/paid-off?key=studio1-sales-2026 key-gated Two actions, not one: END the series (owes nothing) and REDUCE the charge to the balance (owes less than the charge). Every KPI card is a filter into the list, and /paid-off/handover.csv is the one button — it hands over the CSV and records the hand-over, so two people do not work the same list twice. The sweep is built and DELIBERATELY NOT ARMED (three gates, dry run default, max 10).
Write-downs /reductions · /reduction-recovery · /reduction-groups https://reports.iconicbyai.com/reductions?key=studio1-sales-2026 key-gated The CRM has no reductions report — assembled from the Debtor Ledger plus each booking’s finance schedule.
The agency /monterey https://reports.iconicbyai.com/monterey?key=studio1-sales-2026 key-gated Plus /monterey/in-collections.csv — the CRM export rebuilt byte-for-byte because the real one 500s.
The daily hand-over /chase-daily · /decline-supply · /missed-chase · /chase-coverage https://reports.iconicbyai.com/money#landed mixed /chase-daily writes the file, /decline-supply keeps “the pull broke” and “the cut rule ate them” apart forever, /missed-chase is the recovery list of everyone the old cut rule dropped (1,020 people, $1,254,900.83 past due), and /chase-coverage checks the whole handover end to end. The verdict is PER DAY on purpose — a blended rate reads 97% and hides an outage.
Calls /collector-calls https://reports.iconicbyai.com/collector-calls open The numbers a collector actually rang, named against four local snapshots (88% matched). --agent= and --days= — a per-collector page is config, never a new file.
The customer-facing portal /pay https://reports.iconicbyai.com/pay?c=<square_customer_id> per-customer link Live since Mar 2026. Bonus-credit tiers, Cash App Pay, scheduled payments, autopay, callback. The bonus credit is never applied — fix before any traffic. chase.csv already emits the per-customer link.

Manifest order matters and is already correct. finance-cohort-ledger → collections-flag-pull → in-collections-file → monterey; and collection-rate-segments must sit between the ledger pull and collection-rate or it reads yesterday’s money. money is LAST. chase-daily is owner:"render" on purpose — a laptop copy would write the handover ledger twice.

Twelve design rules, every one of them paid for by a wrong number that shipped. If a new dashboard follows only these, it will not repeat any published error in this pack.

1

Lead with absences, not rates

Every rate on this book is either selected or a mix. Counts of things that did not happen are the only numbers nothing selected, and they are also the ones that point at work. “X people missed yesterday and nobody has contacted them” is both safe and actionable; “our recovery rate is N%” is neither.

2

Make the unit an ACCOUNT, never a card and never a row

2,123 debtors hold more than one GHL card and 2,110 hold cards on more than one board. Summing per card overstated the book by $327k. Summing arrears per row double-counts anyone who missed twice in a month. One row per order reference, always.

3

Show the DAY, not the blended rate, for anything operational

The handover reads 97%/94% blended and looks healthy. Per day it shows two total outages that cost 20 people permanently. A monitor that scores a missing data pull as green is worse than no monitor — “no payment data pulled yet” must never render as a clean day.

4

Split the three populations on the face of it

New (opening miss) · Early life (10-60d) · Zombie (4+). They are not three problems — the decline mix is ~85% soft in all three — but they have three economics and three voices. Population 2 is mostly Population 1 arriving late, and Population 3 is the biggest number and the worst place to spend the next hour.

5

Put ever-paid-anything on every row

It out-predicts the decline reason. Never-paid reads 41.5% at one miss and climbing, while every reason family clusters 10-31%. “Never paid a cent” and “paid four and stopped” are different conversations with different scripts.

6

Surface the decline REASON, and route on the three that matter

CARDHOLDER_INSUFFICIENT_PERMISSIONS → “there is a toggle in your banking app” (the only 30-second fix anywhere in this area, ~8 people a week, and nobody is told). Dead-card codes → ask for a new card, never chase. No card on file → never say “your card was declined”. Everything else is a money conversation.

7

Carry a freshness stamp that is the DATA, not the build

The newest real data row, per feed, on the page. A build stamp lied for eleven days. And show the feed’s owner beside it — a laptop page reading a render-owned feed is the specific way this goes wrong.

8

Count the DND FIELD, never the tag

1,448 blocked against 432 tagged. Any suppression keyed on the tag silently texts ~1,000 people who said stop, 107 of whom replied STOP to a carrier and can never be undone.

9

Show contact as a LADDER, never a yes/no

Live (≤30d) · cooling (31-90d) · cold (90d+) · silent · untexted. A lifetime yes/no makes a mostly-cold book look half-alive — 58% against a real 8.3%.

10

Do not build a metric on who chased

Square records who executed the transaction and nothing else; the invoice creator is always the processor. Proved, and do not rebuild it. If per-collector credit matters, it needs a new source of truth — a CRM activity log, or crediting by assignment. And calls do not write back to GHL at all.

11

Every list needs a hand-over log, not a lock

The /paid-off pattern: a button that hands over the CSV and records who took it and when, so two people do not work the same list twice. Pressing twice is safe. It is a LOG, not a lock, and it writes nothing outward.

12

Assume a control arm from day one

We have never run an automated pre-notice, an automated cliff message, a weekly plan or an automated second attempt. Every one of their expected rates is UNKNOWN. Half the long-payer group self-cures inside 9 days with nothing sent — any programme without a held-back 10% will claim credit for that.

Live counts as at this build — the only computed numbers on this page

The handover, last 28 days — built 2026-09-23T13:02. 415 failed openings 2026-08-26 → 2026-09-22 · 53 paid it themselves inside 24h · 362 needed chasing · 342 reached the GHL (94.5%) · 321 texted (88.7%) · 20 never reached the GHL ($3,841, 19 still unpaid) · 21 in the GHL and never texted · 1 whole day lost (2026-09-04, 15 people, $2,989).

Last complete Mon-Sun week — built 2026-09-23T13:00. 2026-09-14 → 2026-09-20: 261 agreements opened, 123 failed the first instalment (47.1%, $24,717) · 109 in the GHL, 104 texted, 50 replied (48.1% of texted) · 14 never reached the GHL · fastest text 23.1h, median 25h · 93 still unpaid ($17,803) of which 37 replied and still owe — the warmest names in the business. Reasons: TRANSACTION_LIMIT 58 · GENERIC_DECLINE 31 · INSUFFICIENT_FUNDS 24 · PERMISSIONS 7 · CARD_EXPIRED 2.

The recurring series book — built 2026-09-23 11:17. 64,127 series from 422,333 invoices · 10,781 still billing, 10,777 of them holding exactly one invoice ahead · 8 owe nothing and are still billing · 30 are about to charge more than the balance ($4,244 over, 14 inside 7 days) · CRM re-check agreed on 38 of 38, 0 disputed · alignment: CRM ahead on 10,134 ($7,145,137, EXPECTED — collectors key cards elsewhere), Square ahead on 2,091 ($689,704, the direction worth investigating).

The agency — built 2026-09-23 12:07. 17,919 agreements placed into collections carrying $31,893,775 of balance, most recent 2026-09-17 · $923,969 ever keyed back across 980 agreements, nothing since 2026-08-31 · $204,336 written down · 5 live agency logins inside our CRM.

Twelve things nobody has answered. The first one is the most valuable question in this whole area — it is the largest single driver of the rise in declines and it points at our own submission behaviour, not at the economy.

The questionStatusWhat is known
What changed in Sep-Oct 2025? HIGHEST VALUE CARDHOLDER_INSUFFICIENT_PERMISSIONS per 1,000 attempts: 8.0 (Jul 25) → 18.9 (Sep) → 22.9 (Oct) → 39.7 (Jul 26), a +594% climb. INVALID_ACCOUNT and INSUFFICIENT_FUNDS step in the same two months. It points at how charges are submitted, not at customer finances — and it is the largest single driver of the decline increase. NOT diagnosed.
Is the “collection rate collapse” partly a measurement artefact? OPEN Only July 2026 is measured — $76,143 of phone money the invoice-based rate could not see, true rate 42.4% not 38.5%. If the mix has shifted over time from auto-charge toward phone payments, part of 55.9%→39.8% is the measure, not the business. The month-by-month gap is NOT yet computed. Do not assert it either way until it is.
Why does Boston miss 34.3% when everywhere else is 40-47%? OPEN Nobody has explained it. Worth an afternoon before building anything, because whatever Boston does differently is the cheapest available lever.
Does the agency still receive anything? OPEN Nothing keyed back since 31 Aug 2026, against a source that is demonstrably still writing 6,153 ordinary payment lines in August. Unresolved: whether Caet’s team still produces the file, or whether the five Monterey logins mean they now pull it themselves. Nothing in the CRM records the file being transmitted at all.
Does cancelling an unsent SCHEDULED invoice email the customer? BLOCKS THE SWEEP Square notifies a customer when an invoice they have been sent is cancelled. These are scheduled and undelivered, so the expectation is silence — an expectation, not a verified fact, and verifying it means cancelling a real invoice. Doing 904 at once would send 904 emails to people who owe money.
Does the CRM keep saying the $4.35M is owed after a cancel? BLOCKS THE SWEEP Cancelling an invoice cancels the REQUEST, not the agreement. And collectors would need an ad-hoc invoice or payment link to replace the rail they lost.
Is the /pay portal’s miss-5 threat true and compliant? LEGAL It promises the customer “legal action and impact to your credit rating”. Verify before any traffic reaches that page.
What feeds the Debt AI G-Sheet, and is it still on? OPEN Neil switched OFF the Google-Sheet feed on 6 Aug 2026, making the Square-declines push the intended entry point — but that push is still dry run, and the chase file is still loaded by hand. If both ever run, everyone gets texted twice.
Which CRM field carries the shared ID? BLOCKS THE SPINE Neil confirmed Square / IES / CRM share an id, which removes the collections spine’s one real technical risk. Nobody has said which field it is.
Seven ledger spreadsheets, one per market — only New York is mapped BLOCKS THE SPINE ledger_config.json is hard-coded to New York’s exact geometry. The risk that could corrupt IES rather than merely delay it: wrong column letters do not throw, they return a plausible number. Needs the seven spreadsheet IDs, each shared with studio1-automation@studio1-automation.iam.gserviceaccount.com, and a per-city reconcile against Square before migrating.
Ladder timings, reduction bands, and where S3 sits AWAITING NEIL 7/21/45 are placeholders. Open: whether S3 is the wider team or S2 goes straight to Caet (this sets her workload), the reduction bands + floor + approver, follow-up and expiry intervals, Monterey’s file format, and the GHL sender identity for collections.
The repeat-miss risk of a reliable discount WATCH A dependable pay-now discount trains people to miss. On a recurring base, customers learn that letting the card decline produces 15% off, and a one-off cost becomes a permanent price cut. Five rules agreed: never on the expired-card path, once per customer per rolling 12 months enforced by the portal, “one-off” stated in the copy, first two misses only, and track repeat-miss rate among discount takers vs everyone else as the tell.