When a figure is wrong

They are not presentation preferences. They are what makes a wrong figure findable rather than invisible, and they are enforced in the engine rather than left to a style guide.

1. A figure we could not read renders a dash and a reason. Never a zero, never a filled-in estimate. A zero is a measurement, and printing one where we measured nothing is the single most expensive thing a figure can do: a watched line reading "at or above 25" against a coerced zero would send a breach notice about a company that never breached.
2. A figure read from a filing names it, and one that was not says so. Where a statement stands behind it: the statement, the fiscal year, the date it was filed, the accession number and a link to the document itself. Where none does, and a price, a figure computed from the rows above it and a line standardised by a data provider are the three cases, the trail says which of them it is instead of leaving the space blank.
3. A figure that is not meaningful for the filer is refused, not computed. Return on invested capital for a bank is the standing example. The arithmetic would produce a number; the number would mean nothing, so the surface says which test does not apply and why.

The methodology page carries the engine-level version of all three, with the thresholds and the refusal rules as the code holds them.

Three things, in descending order of how much they matter, and the order is not rhetorical: it is how a report is prioritised when two arrive on the same day.

1. A notice that should have reached you and did not, or reached you wrong. This is the service failing rather than a figure being off, and it is treated as the most serious class of fault there is.
2. A figure that is wrong on a surface somebody acts on: a watched assumption's reading, a provenance trail, an evidence record.
3. A discrepancy that is visible but that nobody is acting on inside a day: a coverage count, a percentile, a sector label, a company name.

A feature request, a figure no filer reports, or a disagreement about method is none of the three. Each still gets a real answer naming which it is, because a silence reads as a queue.

We assign the severity when we acknowledge a report, not the person reporting it. Writing "urgent" does not make something critical, and writing nothing does not stop it being critical.

Critical

A notice about a watched line that did not arrive, arrived wrong, or arrived repeatedly. The overnight pass not running counts, and so does it running and reading nothing.

High

A figure a reader reasons with: a watched assumption that has not broken, a provenance trail citing the wrong statement, a delivery channel failing for some readers and not others.

Normal

Something visible that nobody is acting on inside a day: a count, a percentile, a label, or a figure that is right but reads as though it were not.

Not a fault

A request, or data we do not hold. Classified and answered, with no implementation implied by the answer.

There is no response time published on this page, and that absence is deliberate rather than an omission. One person answers these. A promise of a fixed window is a promise about the day they are unreachable, and we would rather tell you what we will do than quote you an hour we might not keep.

Four routes, and a report through any of them is in scope. The first is the one to use if you are looking at the figure right now.

• The control beside the figure. A figure on a company page opens into its provenance trail for an account in its trial or on a subscription, and the trail ends with a way to report it. It sends the company, the metric, the figure as you saw it and the filing the page cited, so the report arrives with everything needed to check it.
• The message bubble, which shows only while its relay to a phone is set up, so a message reaches a phone rather than a queue.
• Email, at hello@valuescreener.ai.
• A claim under the alerting commitment in clause 9.2 of the Terms, which is adjudicated from our own records rather than from memory.

Reports that nobody files are in scope too. A failed overnight pass, a blocked notification path and a job that did not register raise themselves, on the same footing as a report from a reader, and the log below says which of the two found each correction.

When a figure was wrong, the reply names the source statement, the fiscal year, the filing date and the accession; it states the window the wrong value was served for; and it re-tests any watched line that used the figure against the corrected value, so you learn whether a notice was owed and never sent. If the correction affects anyone else, it is published below before it is filed away.

Where a fix is not possible, the answer says so and says why. A figure nobody filed cannot be computed, and a vendor series that ends where the plan ends does not get extended by inventing the rest.

These are stated because they are easy to blur and the blurred version flatters us.

• The check running is not the message arriving. How many overnight passes completed and how many notices were accepted are two separate counts here, and nothing adds them.
• Accepted by the provider is not delivered to your mailbox. A message can be accepted and refused an hour later.
• Confirmed delivery is not measured today. It needs a webhook from our mail provider that is not yet configured on this installation, so every surface reports delivery as not available and none of them prints a zero. The strongest claim we can honestly make right now is that the provider accepted the message. Making this measurable is the top item on the operations list, and it is a configuration change rather than a build.

Everything we have published, newest first. Each entry says what it read, what it reads now, the window it was wrong for, the filing it is now read from, the cause, and what changed so it cannot happen the same way again.

No corrections have been published yet. This counts rows in the record rather than describing it, so the day there is one, it appears here.