How Kelynto explains variance to plan.
Kelynto does not ask a language model for causes or for numbers. A rules-based engine reads the week from your own files. Every figure reconciles to the week's variance to plan, a cause gets an amount only when your own records show it, every amount is labelled an estimate and opens to the record behind it, and what the records do not explain is stated. This page sets out each step and how the method is tested.
The method, in order.
Every weekly run goes through the same steps, on your own data and no one else's.
-
Under plan and over plan, store by category
Variance is actual sales minus plan. The week is taken one cell at a time: one store, one category, its actual sales against its plan. The review calls a cell under plan a shortfall and a cell over plan a lift. The shortfalls are added up on their own and the lifts on their own, for the company and for each region, store and category, and every level adds up to the one above.
The net variance to plan is stated beside them. A week $12K below plan can hold $77K of shortfalls and $65K of lifts; the review states all three, so a strong store never hides a weak one.
When most of a store's categories move together, something acted on the whole store, so the review lists it once for the store and not once per category.
Setting: the share of categories that must move together.
-
The week in three parts
Every level of the review has the same three parts.
Recorded this week, with estimated impact. Causes read from your own records: a stock-out, a store closure, a promotion that ran. Each carries an amount, labelled an estimate, and the record behind it.
On file this week. Weather, local events, road work and competitor openings near a store, listed with place and dates as context for the planner.
Still open. Every shortfall no recorded cause accounts for, stated as an amount.
One list in the product names the causes that may carry an amount. Every screen, the drafted commentary and every export read that list.
-
A stock-out is read from your in-stock file
The engine sets this week's in-stock rate for a store and category against that store and category's own recent weeks. A drop that clears the threshold is recorded, with the stores it touched and each store's rate beside its usual one.
The amount beside it comes from a stated rule: the plan scaled by the size of the drop, and never more than that store and category fell short of plan.
The record shows that availability dropped, store by store. The amount is labelled an estimate.
-
A promotion is stated as the calendar has it
A promotion is stated with the stores and the category your promotions calendar gives it. An estimate of its lift is stated only when it clears a gate: it must point the way a promotion would push, stand clear of the ordinary variation of that week, and be material by your own thresholds.
When the gate is not cleared, the review lists the promotion as on the calendar with no estimate.
-
A recorded cause is credited where its records sit
The review credits a recorded cause's estimate only in the store and category cells where its records sit, and only against the movement of its own direction there: an estimated loss against the shortfalls of its own stores and category, an estimated lift against their lifts.
A cause is never credited with more than moved in its cells. When two causes sit on one cell they share that cell's movement, in the order the engine ranks them.
A cause is never credited with a fall or a rise that did not happen where it sits.
-
The record behind a recorded cause, one click away
A recorded cause opens to the records it rests on: which record, which dates and which stores. For a stock-out that is your in-stock file, with each store's in-stock rate that week against its usual rate. For a promotion it is your promotion calendar and the stores it covered.
The planner judges each cause from that record.
-
Context on file, with place and dates
A weather alert, a local event, road work or a competitor opening is matched to stores by distance and by week. Each is listed with its place, its dates and the source it came from, in order of date. Opening one shows the stores it is on file for.
A planner who knows what happened at a store can say that a record mattered, that it did not, or that something else did, and that answer is kept as the planner's judgement.
Weather, alerts, events and road work can be pulled from outside sources. Competitor openings come from a file you send.
-
What the records do not explain is always stated
Every shortfall that no recorded cause accounts for is stated as an amount: the sum of those shortfalls, store by category. Its share is a share of the shortfalls: $62K (81%) of $77K of shortfalls, in a week $12K below plan. The movements the engine found inside it are named first, such as three stores below plan in every category, then what is spread over other stores and categories.
It is on the page at every level, for the company, a region, a store and a category, and every level adds up to the one above. A shortfall is never assigned to the nearest plausible story.
The open amount is useful. It is a missing record, a threshold to tune, or a cause your planner already knows and can name.
-
The noise guard
Inside the open amount, the review names the movements that stand out, such as "every category below plan together at 8 stores". Store and category numbers move every week for no reason at all, so a movement is listed only when it clears three bars: your percentage threshold, your minimum amount (sized to your business when your first sales file loads, which an admin can set by hand), and a multiple of that category's ordinary spread between stores that week.
The spread is measured from your own data each week, so the guard needs no tuning to start.
Settings: both thresholds and the multiple.
-
What planners decide, name and add
A recorded cause is a candidate until a person rules on it. The planner confirms it, rejects it, or corrects it to a different cause. Each verdict is recorded with who gave it and when. Verdicts are kept as labels in your workspace and counted in planner agreement on Score and learn. The planner also edits the drafted commentary, and every version is kept with its history.
For a movement inside the open amount, the planner can name the cause: a pricing change, a plan or forecast error, store operations or a data issue. The review then says beside the amount which cause planners named, and at which stores.
A planner can also add a cause the review missed, in their own words, with the stores, the dates and their own estimate, on a draft or on a published review. It is listed under "Recorded by planners", with who recorded it and when.
-
Records ahead, as context
Watch lists what is on file near each store for the weeks ahead: weather forecasts and alerts, local events, road work and competitor openings, each with its place, its dates, its source and when the source issued it.
Impact Memory keeps what was on file at each store beside what sales and plan were in those weeks, as a record the planner can look up the next time the same condition is on file.
Setting: how far ahead the weekly list looks (1 to 8 weeks, 3 by default).
-
Every figure is printed from the week's evidence
The engine computes every number. The words are written by Kelynto's built-in writer, or by Claude when the AI writer is switched on for the workspace.
The built-in writer writes the draft from the three parts. Every amount and percentage in it, and in a planner's edits each time they are saved, is matched against the figures of the week, and a figure that matches none is named on the review, in the publish step and to whoever sends the documents.
With the AI writer, Claude, by Anthropic, writes from the week's evidence, in which every figure, estimate and record carries its subject, period, direction and source. The model cites an entry for each figure and never types one, and Kelynto prints the figure in its place. Each claim is checked against the entries it cites, and is dropped when its figure is about another place, when it moves a record to a place the record is not on file for, when it turns a direction round, when it states an estimate as a fact, when it sets an amount beside a record on file, or when it reports an action as approved. A text that does not hold as a whole gives way to the built-in draft.
The planner reads, edits and publishes the review, and the review shows how the draft was written, claim by claim.
-
When the data changes, the published review is kept
A review keeps the figures it was produced with, and names the files it was produced from. When the sales or plan of its week are loaded again, it says so, and shows its own figures with what is on file now beside them.
Running the week again states the new data and carries over everything people did: labels, causes planners added, follow-up actions and every version of the commentary. A published review stays published as it was until a planner publishes the restated one, which says what it replaces, when it was published and by how much the figures changed, and offers it to the people who received the earlier review.
What an amount means.
A method is easier to trust when it says exactly what each number is.
- A record first, then an estimate A record confirms that something happened. The amount beside it is an estimate from a stated rule, labelled as one, and a person confirms each cause.
- Three kinds of recorded cause carry an amount Out of stocks, store closures and promotions that ran. Weather, local events, road work and competitor openings are listed as context. Anything else is in the open amount, where your planner can name it.
- Rules you can read Each cause is stated by a rule with a threshold, set out on this page. The labels your planners create are yours, and can be exported at any time.
- Thresholds are settings They are fixed for your workspace, and frozen before your eight sealed weeks are run.
How the method is tested.
By automated tests on generated weeks, where the right answer is fixed in advance.
- The arithmetic. Shortfalls and lifts are added up separately, a cause is credited only where its records sit and never with more than moved there, and every level adds up to the one above.
- The three parts. Only a recorded cause may carry an amount, on every screen, in the draft and in every export. The tests fail if any other record does.
- A week whose lifts hide its shortfalls. A promotion's lift at many stores beside shortfalls at a few: the review must state those shortfalls as still open.
- Every load. A file is read back from the database before it is kept, and a recurring export is reconciled to its control totals to the cent.
- The words. Figures in the commentary are matched against the week's data on every save, and the AI writer's claims against the entries they cite.
Then on your own history.
You can run Kelynto on your own past weeks before you buy, under rules fixed before anyone sees a result.
- Sealed by you. Someone on your side outside the planning team keeps your last eight weeks, and the commentary your team wrote for them, until Kelynto's set-up is frozen and recorded.
- One run. The eight weeks are run once, in order. No setting is tuned on the weeks under test.
- Scored against your planners and simple rules. Each kind of cause is scored against what two of your planners judged separately, beside simple baselines computed from the same files: no cause, the plan gap alone, a seasonal naive rule, a historical anomaly, and availability alone.
- Every week reported. All eight weeks are in the result, each restated in three parts and reconciled to your own totals.
- Each result names its engine version. The version is stored on every weekly run, so a figure can always be traced to the code that produced it.
The measures of a live pilot.
The three measures are the ones on the product's own Score and learn page, counted from what your planners did. Each is reported with its sample size.
| Measure | Definition | Pilot target | Minimum sample |
|---|---|---|---|
| Draft edit rate | The word-by-word difference between the draft and the text the planner published. 0 means sent unchanged, 1 means rewritten. Reported as the median of up to the last four reviews. | 20 percent or less rewritten | 2 reviews with planner edits |
| Planner agreement | Planners' labels that confirmed what the review recorded or had on file, out of all such labels. Also reported by type of cause. | 60 percent or more | 10 labels |
| Would use every week | The planner's weekly answer on a scale of 1 to 5. | 4 of 5 or higher | 3 answers |
The targets are set before a pilot starts, and your own bar can replace them in the pilot agreement. Reported beside them: the share of the shortfalls credited to recorded causes, the share still open, the correction and rejection rates, how often Watch items were new to the planner, and the minutes a review took, as the planner reported them.
The weeks around the review.
Each of these keeps every rule above: an amount in the review comes only from a recorded cause and is a figure of this week's statement, and a planner's own input is kept apart and attributed.
- What this means for next week Under the three parts, a panel states in words what carries into next week, built only from this week's statement and the records on file: a stock-out still open or ended by the in-stock file of the week after, a closure that ends or continues by its record's dates, a promotion that continues or ends by the calendar of the week after.
- Memory across weeks At the top of the review, this week's statement is set beside the week before's, measure by measure, computed from the two statements. The review says when a recorded cause, or an open amount above your threshold, has been on record at the same place for several weeks running.
- Notes from the stores A store or district manager leaves a short, dated note for one of their stores. It is shown on the review as context, with who wrote it and when, and a planner who finds that it explains a movement records a cause from it as the planner's own statement.
- The Outlook stands apart from the review It sets what is on file for the weeks ahead beside your own plan or latest forecast, by cause, and changes no figure of the week that ended.
Five ways to take the next step.
-
Explore
See the product
The product tour shows real screens of the product on a synthetic retailer, in the order a week runs.
A private copy of a synthetic retailer, Harborview Markets. No conversation needed first.
-
Evaluate with your data
Run it on your own weeks
Eight sealed weeks of your own history, run once and scored against bars fixed before anyone sees a result, then about eight live weeks.
Create a workspace and load your own files. Sign in by an emailed link. No password to create.
-
Review the security
Check every control
Kelynto Managed on Microsoft Azure or in your own Azure account, what crosses the boundary, and every control with its status and evidence.
-
Offer it to your clients
Build it into your service or product
For planning services, retail software vendors and consultancies: the partner API, embedded views and your brand, starting with a sandbox of synthetic clients.
-
Talk to us
Twenty minutes with the founder
Your process and your questions. Or write to marc@kelynto.com.