FAQStraight answers, limits included.

What the figures are counted over

• Every figure on a console screen is counted over what that screen has in front of it: the page you are viewing, a fixed span of days, or the computers reporting in. It is never a total for your organization, even on the last page of your selection, and a page with more behind it is a floor. Filters work in your browser on the rows on the page and move no figure above them.
• Findings are counted as matches, not kinds of finding and never people; AI is counted as installations, never people. A summary can count one incident on several rows, and nothing is added across panels, divided into another figure or joined by a matching name: they count different things over different spans.
• The coverage percentage is over computers enrolled with the agent: a machine with no agent sits outside it rather than counting as not reporting, and a retired enrollment is in neither figure. Each AI panel is over its own group of computers and a span of days the service sets, looking back only as far as the check-in history kept.
• Events over time are counted across your whole organization over a fixed range of whole days or weeks, leaving out the period in progress and everything before it, grouped by when RedactWall received each record rather than when the device said it happened, and never finer than a day. Scan usage comes from the billing meter for the current month and holds counts and sizes only, so nothing about one scan can be answered from it.

What a screen deliberately does not total or show

• No screen totals your organization’s events, findings or held incidents, ranks policies, detectors or users, or scores an incident for risk, severity, percentile or confidence: no identity is attached to an event, and a decision carries two yes-or-no facts and no scale. There is no investigation, no grouping of held operations into one incident, no policy name, no PDF export, no scheduled email and no logo.
• No screen shows where an event was going or where in the text a finding sat, and there is nothing to search. Fields held elsewhere, such as which computer and account an event came from, a reviewer’s note, a held original or a full address, are never sent to a screen, and RedactWall’s own administration console cannot read your evidence or anything an employee typed.
• A separate employee activity store can retain encrypted content after mandatory hard-stop values are removed. Content access is role controlled and reads are audited. Operational retention defaults to 30 days, with a maximum of 90 days.
• The training policy requires explicit opt-in before future customer data is collected for training. Legacy samples are excluded from training until their rights are reviewed; they remain subject to their ordinary retention period of 365 days.
• The opt-in policy is merged to the product main branch and shipped in release 0.1.12, which was installed on production on 2026-09-24. The live collection path has not been independently exercised, so this page does not claim that collection is already disabled in the live service. Do not use customer data until applicable terms have been reviewed and agreed.
• The console reads and records; it puts nothing into force. Putting a policy, a hierarchy place, a fleet operation, an entitlement or branding into force is an operator’s job done elsewhere, only an owner or a security administrator may declare policy, and nothing runs by itself. RedactWall cannot run a command, install software, execute a script or reboot a machine; withholding a module refuses new sensors and revokes nothing already enrolled; and a commercial hold never turns off inspection or the hard stops.
• A policy declared or composed on a screen changes nothing any machine enforces: a declaration is saved, not activated; a detector or starting policy is a file the console sends nowhere and does not check, so a screen refusing nothing is no verdict that it will be accepted; a policy screen shows what was recorded at activation; and per-category rules cannot be composed on a screen yet.
• Some lists are deliberately shorter than what exists, such as people or routing groups, so a missing entry is not evidence that it does not exist. Beyond the gaps under the AI tool count, the inventory cannot see servers declared by apps it does not read, calls to servers with no RedactWall guard, or AI use that is not a tool call, so an empty table is never evidence that an estate is quiet.

What a check proves, and what it does not

• A check-in is a periodic sample, and a state records what it could observe: seen running is execution, not a person; installed, never seen running means no check-in caught it; declared means an app was given access, not that anything ran; and a tool ran is the only positive evidence of use. A computer reports the version of its AI-tool list, not its agent program, so the screen finds out-of-date agents and cannot certify up-to-date ones.
• Raw content permitted counts decisions, not what a device did afterwards; refused with nothing read says nothing about the content; and a false positive is a reviewer’s classification, not a measurement of the detectors. Settled incidents include those that expired undecided, so the hours figure is time to resolution, and no figure names a reviewer.
• Routing an incident moves it into a queue and grants no authority. A rationale is a choice from a fixed list; a bulk decision is one judgment over the rows selected, with no check that each incident is still the one you looked at; a settled incident accepts no change; and an incident’s summary recommends neither outcome.
• The on-screen evidence check gives the same four answers as the check your own team runs, examines at most the number of entries on its label, and refuses a longer record rather than partly verifying it; its counts are over the entries examined, and nothing after the last entry recomputed is vouched for. The overview’s evidence counts come from that check, on request, and count records in the history, not things that exist now.
• The category list beside a policy form is what this version of RedactWall detects plus the mandatory floor; any well-formed name is accepted, so a missing category is not evidence that nothing produces it, one no detector produces never matches, and a custom detector can produce categories the console cannot show. A rule can raise an action, never lower it; a refusal quotes none of your text; look-alike letters slip past every detector; and an approved detector runs only on your own workstations.

What an absence means

• A zero, a blank, an empty list or a word such as no run yet, disabled or no decision recorded is shown only when that was measured or recorded, and means only that: a share or an average is never zero when there is nothing to divide, a module with no decision is on because nobody decided, and an empty list is read against the coverage figures before it is taken as an estate with no AI.
• When the service did not answer, or your role may not see a figure, no figure and no zero is shown, and only that figure is affected. The service gives one answer to several situations, such as a record that is not there or one you may not read, so nothing says which it was; and a record that is not on the page loaded is not a record that does not exist, because the console cannot look one up on its own.
• Never reported, could not tell, partial, present with nothing to look for, let through with outcome unknown and not shown all mean RedactWall does not know, never that nothing is there: a newly enrolled computer that has not checked in looks exactly like one whose agent never ran, and a field held elsewhere is listed as not shown so the gap is visible. A field listed as received and not displayed is a code computed with a secret key: the code is withheld, not the fact that it arrived.
• An empty page, an empty timeline lane, a queue with no states chosen, or a read your role may not make, that the service limited, that did not finish or that could not be reached shows what that screen read, which in each of those cases is nothing beyond itself: it says nothing about whether your organization has any incidents, holds or events, whether a particular hold or event exists, which AI is installed on any computer, the state of any endpoint, which policy any computer is enforcing, how a hold is routed now, which build any agent is running, or what is scheduled. The console renders only what the service sent, so an address that does not name an evidence identifier is not repeated on the screen.
• One answer also covers a session with no workspace chosen, a session with no access to the institution and an account whose membership was withdrawn, and a refused re-authentication covers a wrong password, a disabled credential and a session too old to re-authenticate; the console does not invent a difference it was not given. A request refused because it did not come from the address the service expects is not an answer about the password.