Every decision,and what it was worth.

Approve an action, and when the window closes the ledger writes what it was worth — with the comparison the number rests on.

Impact ledger · demo network

window closed 25 Aug · 06:40

beta
FOUND
8596 400 € projected
IN FLIGHT
45NOT MEASURED YET
PROVEN
4031 200 € measured

Reprice 62 dairy SKUs

28 Jul · City Center · gross margin

PROVEN+3 700 €

4 matched controls · α = 0.10

Win-back campaign, 2 140 guests

22 Aug · network · repeat rate

PROVEN+1 350 €

randomised holdout

Bakery reorder, 9 343 SKU at zero

04 Aug · North Park · stock-out share

MEASURING+1 850 € est.

NOT MEASURED YET

Cut discount leakage

16 Aug · Westside · average basket

NO EFFECT0 €

NO VERDICT

Move two shifts to Friday peak

19 Aug · Harbour · labour cost %

INCONCLUSIVE— no money

NO VERDICT

every row carries its own currency and its own methodology versionExport CSV / JSON

A measurement that is allowed to come back empty.

Nothing is measured that was not committed, and nothing is called proven unless both the data and the work behind it hold up.

You approve, and the baseline freezes with it

It leaves the Inbox with a target metric, a predicted effect and a measurement window, and in that same moment the metric and the money base are captured, scoped to the location the action targets. A strong month later cannot inflate an older result, and until the window closes the line is a projection that the ledger labels as one.

The window runs, and the work has to happen

If the action carried an execution task that was never completed, the line is closed as inconclusive and writes no money, whatever the metric did in the meantime.

gate · execution

Inconclusive — no money written

A verdict, or no money at all

Proven, no effect, worse or inconclusive. A line becomes proven only when the improvement clears the significance test against comparable locations — not when the number happens to move.

gate · significance

No effect — too small to matter, even when the statistics like it

Where a proven number comes from.

City Center4 matched locations
window 28 Jul – 25 Aug

− 4 weekscommit · baseline frozen+ 4 weeks

+3 700 €

PROVEN · 4 matched controls · α = 0.10

Placebo check on the four weeks before the commit: the two groups tracked each other to within 0.4 pp. Had they already been drifting apart, this line would not say proven.

Built to survive a hard question.

The ledger exists to be checked. Sometimes by you, sometimes by a board or an investor running diligence. Every part of it is built for that reading.

Frozen money base

The revenue or stock value at commit time is the base every later figure is computed on, and the page shows the arithmetic.

You can redo the multiplication instead of trusting the total.

Strict significance

The effect is judged against a small-sample prediction interval rather than a flat cutoff: at least four comparable locations as controls at a one-sided 10% level, or the location's own volatility where a chain has no peer group. An improvement smaller than half the one predicted — or than a tenth of a percentage point — is recorded as no effect even when the statistics like it.

Fewer lines say proven, and the ones that do hold up.

Proof needs execution

A decision whose task was never marked done is stamped inconclusive rather than credited with whatever the metric happened to do.

The ledger measures work, not drift that arrived on its own.

randomised holdout

A share of the target segment held back from the send, assigned deterministically. Campaigns only.

claimed
matched-control DiD

Comparable locations, grouped by revenue level, format and region, over the same period.

claimed
own-volatility

The location against its own variance, where there are not enough comparable peers.

claimed
legacy control

The same comparison on the old bar: a looser test, and as few as two peers. It stays on rows measured before the current method.

claimed
trend-adjusted

A trend adjustment and nothing subtracted. Not proof, and the row says so.

none
pre / post

Before against after, with no comparison at all.

none

Whatever the grade, it is written on the line and carried on the row in the export, next to the methodology version.

one movement · one metric · one location · one window

City Center · approved by Ana3 700 €
City Center · approved by Radu2 900 €
City Center · approved by Ion2 100 €
City Center · approved by Elena1 400 €
City Center · approved by Mihai900 €
largest, not sum

written to the ledger

3 700 €

the ledger refuses to write 11 000 € — approving a group of recommendations cannot multiply the money

confounder guard · every window

promo covered 71% of the windowdiscount spikematerial price movesupplier / stock-out eventcalendar shock vs learned sensitivity

Bundle promo on soft drinks

Old Town · gross margin · window 12 Jul – 09 Aug

flagged

not claimed · promo cover

A promotion over the window or a discount spike takes the verdict away outright. A price, supplier or calendar shock is written on the row, and takes it away too once the newer guard is on.

A projection counts as found, never as proven.

This is the part of marql we are least willing to oversell. Its entire value is that a number in it survives being questioned.

  • A line that has not been measured yet is a projection.

    It counts towards what was found, never towards what was proven.

  • Metrics with no money base, such as inventory turnover or average rating, are measured as an effect and never converted into money.

  • The beta mark on the card above is ours, not the product's: the ledger ships, and we would rather under-claim it than call it finished.

How a proof number gets attacked.

Against comparable locations over the same period, so the movement everyone shared is subtracted before your result is counted — the section above draws it. A placebo check on the weeks before the decision confirms the two groups were tracking each other; if they were already drifting apart, the line is not called proven. Where there are not enough peers, the yardstick becomes the location's own volatility, and the line says so.

Then there is no proven money. An action that carried an execution task which was never completed is closed as inconclusive, whatever the metric did during the window. The window opens when you accept and the task is due at the end of it, so a verdict taken without checking execution would have been crediting ordinary drift.

No — see the figure above. Counts stay per action, because 'five actions proven' is a true statement about work done, but proven money is a statement about money and must not multiply. Two different quarters of the same metric are two movements, and those do add.

Your network's. Each ledger row carries its own currency, and a multi-country network aggregates to the base currency you set. A Romanian network reads its ledger in lei; the export carries the currency field on every row.

Yes, as CSV or JSON. Each row carries the date, the status, the action and location, the target metric, the predicted and actual change, the predicted and measured money, the currency, the measurement period, the methodology version and who approved it. The ledger page shows the same arithmetic on screen, with the confidence range where the row has one.

No. Campaigns get a stronger method, because they address people rather than locations. A share of the target segment is held back from the send as a control group, assigned deterministically, and the difference between the two groups is the effect. Below a minimum group size the campaign still goes out and is reported as unmeasurable rather than given a number nobody should trust. Campaigns contribute only to proven money and never inflate what was found.

Every price point includes the whole platform. €200 a month per location for the first 9, then €185 and €160 as the network grows, with each rate applying only to the locations in its band.

Ready to see
the euros it found?

The walkthrough is built around the tills and accounting you already run. Leave your details and pick a time.