Does the data behave like data?
Eleven deterministic checks over two surfaces — between deliveries and period to period. Every one is arithmetic you can redo by hand, and every finding carries the numbers it was computed from.
There is no model in it, and that is not modesty: these findings name a data owner. An anomaly detector that cannot show its working is an accusation (R-G-19).
It never looks at a customer or a transaction. Its subject is the data — batches, row counts, series, restatements — and its owners are the data owners. ACF already has the other product, and it belongs to Fraud Prevention & Investigation: AP-12 fraud and collusion detected, OP-09 double disbursement, and the merchant NST rule. Calling this module fraud detection would misrepresent it to the one department whose job that is. Q-38 asks whether ACF wants the name back.
The series here are monthly, 24 points. That is enough for a step change or a frozen run and nowhere near enough for a first-digit test, which wants of the order of 300+ observations across several orders of magnitude. Running it anyway would produce a number that looks rigorous and is not, so it reports not enough data — a distinct outcome from nothing found (R-G-20). It is the same discipline as the campaign sufficiency gate: a check that cannot run is not a check that passed.
| CHECK | CATCHES | FINDINGS | CLEAN | NO DATA | |
|---|---|---|---|---|---|
| B1 | Row-count shift between deliveries | Half a file, or a double load | 0 | 0 | 8 |
| B2 | Silent restatement between deliveries | A published value changed and nobody said so | 1 | 0 | 7 |
| B3 | Schema drift between deliveries | A producer changed their export and told nobody | 0 | 1 | 7 |
| B4 | Miss run between deliveries | The feed that died quietly | 1 | 7 | 0 |
| B5 | Out-of-window period between deliveries | A backfill arriving as if it were today | 0 | 8 | 0 |
| S1 | Step change period to period | A jump no ordinary month produces | 1 | 132 | 7 |
| S2 | Frozen series period to period | Alive and not moving — the hardest failure to see | 3 | 130 | 7 |
| S3 | Out-of-bounds value period to period | A unit or a denominator error | 7 | 75 | 0 |
| S4 | Round-number tell period to period | Hand-keyed estimates presented as measurements | 0 | 130 | 10 |
| S5 | Target-hugging period to period | The Goodhart tell | 0 | 28 | 112 |
| S6 | First-digit (Benford) deviation period to period | The canonical integrity test | 0 | 0 | 140 |
Thresholds are ours, and written down — a 6× MAD step that is also 15% of the level, 4 identical values, a 50% row shift on a batch of at least 20. Q-36 asks ACF to set the real ones.
This series has never changed. A feed that reports the same number forever is indistinguishable from a feed that is not reporting at all, and it does not raise a freshness alert because a value does arrive.
- 24 consecutive identical values of 0
- That is EVERY reported point — the series has never moved
- Certification tier: managed
- What would clear it
- The data owner confirms the value is measured rather than defaulted, or the source is fixed.
- Whose it is
- IT System Planning
This series has never changed. A feed that reports the same number forever is indistinguishable from a feed that is not reporting at all, and it does not raise a freshness alert because a value does arrive.
- 24 consecutive identical values of 0
- That is EVERY reported point — the series has never moved
- Certification tier: manual
- What would clear it
- The data owner confirms the value is measured rather than defaulted, or the source is fixed.
- Whose it is
- IT System Planning
This series has never changed. A feed that reports the same number forever is indistinguishable from a feed that is not reporting at all, and it does not raise a freshness alert because a value does arrive.
- 24 consecutive identical values of 0
- That is EVERY reported point — the series has never moved
- Certification tier: manual
- What would clear it
- The data owner confirms the value is measured rather than defaulted, or the source is fixed.
- Whose it is
- IT System Planning
This file would change values that have already been published. A restatement is legitimate; a restatement nobody announced is how two people end up quoting different numbers for the same month (R-G-09).
- FN-01 2026-08: 21.06 → 20.84 (changed — a restatement, and it will need a reason before it is stored)
- FN-02 2026-08: 7.7 → 7.82 (changed)
- FN-03 2026-08: 83 → 81.4 (changed)
- What would clear it
- The restatement is accepted with a note attached to the period, or the file is corrected.
- Whose it is
- Finance & Accounting
The clock expected something and nothing arrived, more than once in a row. Every bound metric went to not_reported for those periods — not zero, not the last value — which is the only reason this is visible at all.
- 3 consecutive missed windows, ending 2026-09-23
- 3 missed in total across 119 windows
- Breaker settings: 3 misses
- What would clear it
- The producer resumes, and the missed periods are backfilled or explicitly written off.
- Whose it is
- IT System Development
A move no other period in this series produces. It is either a real event worth knowing about, or a load that went wrong.
- Largest single-period move: 3.17 (2024-11 → 2024-12)
- MAD of all 23 period-to-period moves: 0.410
- That move is 7.7× the MAD, and 15% of the series level (20.8)
- Both gates are needed: 6× MAD alone flagged 62 of 140 metrics, which is noise
- What would clear it
- The data owner confirms what happened in that period, and the note is attached to the series.
- Whose it is
- IT System Planning
Values outside 0–100 on a percentage unit. That is perfectly legitimate for a ratio-to-plan or a variance, and impossible for a rate or a share. This check cannot tell which this is — the registered formula is a placeholder on 45 metrics — so it asks rather than asserts.
- 24 of 24 points above 100% (max 1858.6%)
- Formula as registered: events and value per month
- What would clear it
- The metric owner confirms the formula and the unit. If it is a ratio-to-plan, the unit is wrong, not the data.
- Whose it is
- IT System Planning
Values outside 0–100 on a percentage unit. That is perfectly legitimate for a ratio-to-plan or a variance, and impossible for a rate or a share. This check cannot tell which this is — the registered formula is a placeholder on 45 metrics — so it asks rather than asserts.
- 14 of 24 points above 100% (max 106.6%)
- Formula as registered: resolved POS ÷ bucket POS
- What would clear it
- The metric owner confirms the formula and the unit. If it is a ratio-to-plan, the unit is wrong, not the data.
- Whose it is
- IT System Development
Values outside 0–100 on a percentage unit. That is perfectly legitimate for a ratio-to-plan or a variance, and impossible for a rate or a share. This check cannot tell which this is — the registered formula is a placeholder on 45 metrics — so it asks rather than asserts.
- 4 of 24 points above 100% (max 105.3%)
- Formula as registered: waived amount by approval level
- What would clear it
- The metric owner confirms the formula and the unit. If it is a ratio-to-plan, the unit is wrong, not the data.
- Whose it is
- IT System Planning
Values outside 0–100 on a percentage unit. That is perfectly legitimate for a ratio-to-plan or a variance, and impossible for a rate or a share. This check cannot tell which this is — the registered formula is a placeholder on 45 metrics — so it asks rather than asserts.
- 24 of 24 points above 100% (max 126.5%)
- Formula as registered: actual ÷ budget
- What would clear it
- The metric owner confirms the formula and the unit. If it is a ratio-to-plan, the unit is wrong, not the data.
- Whose it is
- IT System Planning
Values outside 0–100 on a percentage unit. That is perfectly legitimate for a ratio-to-plan or a variance, and impossible for a rate or a share. This check cannot tell which this is — the registered formula is a placeholder on 45 metrics — so it asks rather than asserts.
- 21 of 24 points above 100% (max 128.2%)
- Formula as registered: opex ÷ operating income
- What would clear it
- The metric owner confirms the formula and the unit. If it is a ratio-to-plan, the unit is wrong, not the data.
- Whose it is
- IT System Planning
Values outside 0–100 on a percentage unit. That is perfectly legitimate for a ratio-to-plan or a variance, and impossible for a rate or a share. This check cannot tell which this is — the registered formula is a placeholder on 45 metrics — so it asks rather than asserts.
- 20 of 24 points above 100% (max 121.3%)
- Formula as registered: actual ÷ plan
- What would clear it
- The metric owner confirms the formula and the unit. If it is a ratio-to-plan, the unit is wrong, not the data.
- Whose it is
- IT System Planning
Values outside 0–100 on a percentage unit. That is perfectly legitimate for a ratio-to-plan or a variance, and impossible for a rate or a share. This check cannot tell which this is — the registered formula is a placeholder on 45 metrics — so it asks rather than asserts.
- 1 of 24 points above 100% (max 104.7%)
- Formula as registered: manual reports ÷ total
- What would clear it
- The metric owner confirms the formula and the unit. If it is a ratio-to-plan, the unit is wrong, not the data.
- Whose it is
- IT System Planning
No usable target — there is nothing to hug.
24 observations. A first-digit test needs of the order of 300+ spanning several orders of magnitude before a deviation means anything. Running it here would produce a number that looks rigorous and is not.
No usable target — there is nothing to hug.
24 observations. A first-digit test needs of the order of 300+ spanning several orders of magnitude before a deviation means anything. Running it here would produce a number that looks rigorous and is not.
No usable target — there is nothing to hug.
24 observations. A first-digit test needs of the order of 300+ spanning several orders of magnitude before a deviation means anything. Running it here would produce a number that looks rigorous and is not.
No usable target — there is nothing to hug.
24 observations. A first-digit test needs of the order of 300+ spanning several orders of magnitude before a deviation means anything. Running it here would produce a number that looks rigorous and is not.
No usable target — there is nothing to hug.
24 observations. A first-digit test needs of the order of 300+ spanning several orders of magnitude before a deviation means anything. Running it here would produce a number that looks rigorous and is not.
No usable target — there is nothing to hug.
24 observations. A first-digit test needs of the order of 300+ spanning several orders of magnitude before a deviation means anything. Running it here would produce a number that looks rigorous and is not.
No usable target — there is nothing to hug.
24 observations. A first-digit test needs of the order of 300+ spanning several orders of magnitude before a deviation means anything. Running it here would produce a number that looks rigorous and is not.
Only 1 distinct values — with no variation there is nothing for a round-number test to be suspicious of.
No usable target — there is nothing to hug.
24 observations. A first-digit test needs of the order of 300+ spanning several orders of magnitude before a deviation means anything. Running it here would produce a number that looks rigorous and is not.
24 observations. A first-digit test needs of the order of 300+ spanning several orders of magnitude before a deviation means anything. Running it here would produce a number that looks rigorous and is not.
24 observations. A first-digit test needs of the order of 300+ spanning several orders of magnitude before a deviation means anything. Running it here would produce a number that looks rigorous and is not.
Only 1 distinct values — with no variation there is nothing for a round-number test to be suspicious of.
No usable target — there is nothing to hug.
24 observations. A first-digit test needs of the order of 300+ spanning several orders of magnitude before a deviation means anything. Running it here would produce a number that looks rigorous and is not.
Only 1 distinct values — with no variation there is nothing for a round-number test to be suspicious of.
No usable target — there is nothing to hug.
24 observations. A first-digit test needs of the order of 300+ spanning several orders of magnitude before a deviation means anything. Running it here would produce a number that looks rigorous and is not.
No usable target — there is nothing to hug.
24 observations. A first-digit test needs of the order of 300+ spanning several orders of magnitude before a deviation means anything. Running it here would produce a number that looks rigorous and is not.
No usable target — there is nothing to hug.
24 observations. A first-digit test needs of the order of 300+ spanning several orders of magnitude before a deviation means anything. Running it here would produce a number that looks rigorous and is not.
No usable target — there is nothing to hug.
24 observations. A first-digit test needs of the order of 300+ spanning several orders of magnitude before a deviation means anything. Running it here would produce a number that looks rigorous and is not.
No usable target — there is nothing to hug.
24 observations. A first-digit test needs of the order of 300+ spanning several orders of magnitude before a deviation means anything. Running it here would produce a number that looks rigorous and is not.
No usable target — there is nothing to hug.
23 observations. A first-digit test needs of the order of 300+ spanning several orders of magnitude before a deviation means anything. Running it here would produce a number that looks rigorous and is not.
No usable target — there is nothing to hug.
24 observations. A first-digit test needs of the order of 300+ spanning several orders of magnitude before a deviation means anything. Running it here would produce a number that looks rigorous and is not.
No usable target — there is nothing to hug.
24 observations. A first-digit test needs of the order of 300+ spanning several orders of magnitude before a deviation means anything. Running it here would produce a number that looks rigorous and is not.
No usable target — there is nothing to hug.
24 observations. A first-digit test needs of the order of 300+ spanning several orders of magnitude before a deviation means anything. Running it here would produce a number that looks rigorous and is not.
24 observations. A first-digit test needs of the order of 300+ spanning several orders of magnitude before a deviation means anything. Running it here would produce a number that looks rigorous and is not.
24 observations. A first-digit test needs of the order of 300+ spanning several orders of magnitude before a deviation means anything. Running it here would produce a number that looks rigorous and is not.
No usable target — there is nothing to hug.
24 observations. A first-digit test needs of the order of 300+ spanning several orders of magnitude before a deviation means anything. Running it here would produce a number that looks rigorous and is not.
24 observations. A first-digit test needs of the order of 300+ spanning several orders of magnitude before a deviation means anything. Running it here would produce a number that looks rigorous and is not.
No usable target — there is nothing to hug.
Showing 60 of 822. Per-metric detail is on each metric’s integrity page.