Classifier

While it works, C5 keeps asking itself small questions with a fixed set of answers: which categories a new task belongs to, which skill area a step draws on, whether a repeating error is just noise. Normally each of those questions goes to one of the AI models you configured. In the rest of this page, the usual model means that model: the one C5 would ask without the classifier. Settings also calls it the teacher model, because the classifier is measured against it.

The classifier answers the same kind of question with a small model that picks from the fixed answers directly instead of writing a reply. It can use two kinds of model:

• The built-in model, also called the local classifier: decider-0.8b, which C5 downloads once and runs on your own machine. An answer takes a fraction of a second and costs nothing: no provider, no account, no tokens. It is on by default.
• A model from your providers: a model that one of the model providers you set up in Settings › Integrations serves, which an administrator chooses. None is chosen by default. See A model from your providers.

The built-in model runs in a process of its own, apart from C5's server, so if it crashes C5 carries on. C5 never depends on the classifier: whenever it is off, busy, slow, unsure or broken, C5 asks the usual model, exactly as before.

The section in Settings › Learning is called Classifier. Before C5 could use a model from your providers it was called Local Classifier, and this page was called Local classifier; its address has not changed. Outside the section on a model from your providers, the classifier on this page means the built-in model.

In short:

• You do not need to do anything. C5 downloads the model (about 812 MB) once, and at first only records the classifier's answers beside the usual ones, for every kind of decision.
• Want a model you already run to answer instead — on llama.cpp, Ollama, LM Studio, vLLM or a cloud API? An administrator can choose one; see A model from your providers.
• On a machine where it runs on the CPU, it records only one in four Step quality questions, so it stays out of your agents' way; see On a machine where it runs on the CPU.
• To stop it, see Turn the classifier off.
• Managing many machines? Lock classifier_enabled to false in your organisation's policy before C5 first starts, and the model is never downloaded; see For administrators. To let the download through a proxy, allow huggingface.co and the five download hosts listed under Network.
• To get the disk space back, see Delete the model.
• If this machine cannot run it, nothing is downloaded; see Machines it cannot run on.
• No network access? See Offline and air-gapped installs.

Why it is on by default #

The built-in model uses none of your providers and none of your accounts. What it gains from your install is evidence: it records its answers beside the usual ones, C5 measures them against the answers people give, and fits a small calibration from them. The model file itself is never retrained or changed on your machine, and none of this leaves it. (Settings calls this "learning".) An install that never turns it on never collects that evidence, so it could never take over any work. So C5 switches it on for you:

• On a new install, and on an install updated to a version that has it, C5 turns the switch on the first time it starts — but only if nobody has ever set it. If you turned it off before, it stays off. If your organisation's policy sets it, the policy decides.
• Administrators get one notification, The local classifier is on, saying what this machine will do: download the model, use a model that is already here, or not download it (and why — for example, the Egress posture is Offline, or the model's path is a link C5 will not download over). A machine that cannot run the classifier gets no notification, because nothing will happen there.

Turning it on changes nothing at first. It only records its answers beside the usual ones (see What it does at first).

What it downloads #

C5 does not ship the model inside its installer. It downloads it once:

If you have moved C5's caches (with GROWTHER_CACHE_DIR, your organisation's policy, or growther cache relocate), the classifier folder is under that location instead. growther cache show prints where it is. See Where your data lives.

When it downloads. In the background, a moment after C5 has started, and only when all of these hold:

• the classifier is switched on;
• this machine can run it (see Machines it cannot run on);
• the model's path is not a symbolic link: C5 never downloads over a link you placed (see Offline and air-gapped installs);
• the Egress posture under Settings › Enterprise › Network is not Offline (see Network and egress);
• GROWTHER_CLASSIFIER_NO_AUTO_DOWNLOAD=1 is not set, and the CI environment variable is unset or empty (automated build servers set it);
• there is enough free space for the rest of the model plus about 64 MB to spare.

C5 keeps working while the model arrives. The download picks up where it stopped if it is interrupted. Turning the classifier off stops the download and keeps what has arrived; turning it back on carries on from there.

If an automatic download fails, C5 does not keep retrying it while it runs. It tries again the next time C5 starts, or when you turn the classifier off and on, or when you run growther classifier pull.

The model's publisher states that its training data includes datasets released for research use. C5 shows the model's notice next to it in Settings › Learning:

decider-0.8b by Mapika (Apache-2.0), fine-tuned from Qwen3.5-0.8B-Base (Apache-2.0) on public datasets and labels from local Qwen models. Its published training mixture includes datasets released for research use; check that this fits your use, and if it does not, turn the local classifier off in Settings › Learning. About 812 MB is downloaded once from Hugging Face.

To check the training mixture, see the publisher's page for the model: huggingface.co/Mapika/decider-0.8b.

Files in the classifier folder #

Turn the classifier off #

Anyone with permission to change settings can turn it off.

1. Go to Settings › Learning.
2. On the Classifier card, flip the Enabled switch on the Built-in model panel off. It is under Classifier models.
3. C5 asks Turn off the local classifier? and explains that, while off, it stops learning and C5 answers those questions as it did before. Choose one:

The question also says what is on the disk. If the file at the model's path is not the model, it says so (The file at the model's path is not the model (…), so C5 does not use it.) and that turning off leaves the file where it is. If the model's path is a link whose file cannot be reached, it says turning off leaves the link, and that C5 uses the model once that file can be reached.

Turned off, the built-in model is completely quiet. C5 does not start its process, load the model, download anything, run its nightly jobs or print anything about it in the log.

With a model from your providers chosen, the question is Turn off the built-in model?, and it says who answers once it is off. While the chosen model answers: M keeps answering. With the built-in model off, nothing stands behind it: if M stops answering, C5 answers those questions as it did before. While the chosen model is turned off, or cannot be asked now, it says that once the built-in model is off nothing answers C5's routine questions, and C5 answers them as it did before. It also names any kind of decision that only the built-in model answers (Tool-call safety, beside a model that is not on this machine), where turning it off ends the classifier, and any kind that acts on the built-in model's record, where acting ends and the chosen model's answers are only recorded. Its delete button reads Turn off and delete the built-in model. See Who answers.

You can also reach the same question from the Terms dialog: Turn off the local classifier… (Turn off the built-in model… while a model from your providers is chosen).

Turn it back on #

Flip the switch on again.

• If the checked model is already here, or the model's path is a symbolic link to the checked model, it turns on at once.
• If the model is missing, partly downloaded, or not yet checked, C5 asks first: Turn on the local classifier? The dialog shows the model, its size, its licence and its notice. Download and turn on turns it on and starts the download; Cancel leaves it off.
• If a file is at the model's path but C5 already knows it is not the model, the dialog says so, and that C5 downloads the model and replaces that file.
• With a model from your providers chosen, the question is Turn on the built-in model?, as the turn-off question names it. With the chosen model in use, the dialog adds M stays first; the built-in model answers when it cannot., or, with Use as fallback off, M stays first; with Use as fallback off, the built-in model does not answer in its place.
• On a machine that will not download the model — the Egress posture is Offline, automatic downloads are turned off, there is not enough disk space, or the model's path is a symbolic link C5 cannot use yet (one whose file cannot be reached, one it has not checked, or one to a file that is not the model) — the dialog says so and why, and the button reads Turn on. The classifier then answers nothing until the model is on the machine (beside a chosen model, the dialog says the built-in model answers nothing instead, since the chosen model still does); the dialog tells you how to get it there.
  • For a link whose file cannot be reached (a share that is not mounted, say): make that file reachable, or remove the link so that C5 can fetch the model.
  • For a link C5 has not checked yet: C5 checks the file it points to first and uses it only if it is the model.
  • For a link to a file that is not the model: point the link at the model, or remove it.

Delete the model #

The model takes about 812 MB. To get that space back:

• While the classifier is on: turn it off and choose Turn off and delete the model (Turn off and remove the link for a symbolic link).

• While it is already off: the Model line has a delete button, and C5 asks before it deletes anything. What you see depends on what is at the model's path:

  What is at the model's path	Button on the Model line	C5 asks	Button in the dialog
  The model	Delete the model (812 MB)	Delete the classifier's model?	Delete the model
  An unfinished download	Delete the partly downloaded model (300 MB)	Delete the classifier's model?	Delete the model
  A file that is not the model	Delete the file (N MB)	Delete the file at the classifier model's path?	Delete the file
  A symbolic link to a file elsewhere (see Offline and air-gapped installs)	Remove the link to the model	Remove the link to the classifier's model?	Remove the link

Who can. An administrator, on the machine running C5 rather than from a paired device (a phone or tablet paired through Remote Control). The account needs MFA: a passkey, or a directory sign-in your identity provider marks as multi-factor. (An account on an install set up without any sign-in does not need MFA.) An administrator without MFA who chooses Turn off and delete the model still gets the classifier turned off; the message says the model was not deleted, and why.

What it removes. The model, any unfinished download (.partial), C5's record that it checked the model (.verified.json), and a download lock left behind by a crashed process. Nothing outside the classifier's models folder; runtime.json stays. If the model was a symbolic link to a shared copy, only the link is removed and no space is freed. The classifier stays off afterwards, and turning it back on downloads the model again, on a machine that downloads it at all. Every delete is recorded in the audit log.

While a delete runs, the Runtime line reads Not running: The classifier's model is being deleted. It ends in seconds.

When it is refused:

• Another process is downloading the model right now. — growther classifier pull is running. Let it finish or stop it, then try again.
• Your organisation's policy keeps the local classifier on, so its model was not deleted: C5 would download it again. — see For administrators.
• Some of its model's files could not be removed … — usually another program (often an antivirus scanner on Windows) has the file open. The message names the files and the folder; close whatever holds them and delete them by hand.

You can also delete the classifier folder yourself while C5 is stopped. It is a download, not your data. Turn the classifier off first: if it is still on, C5 downloads the model again the next time it starts.

What it does at first #

The classifier is asked about one kind of decision at a time. Each kind has its own setting in Settings › Learning:

Each setting has up to four modes:

So when it is first on, the classifier only records its answers, for every kind of decision, beside the answers C5 already uses. Nothing C5 does changes. Recording is how a kind earns the evidence its promotion gate measures (see When it starts to act); a kind left off never could. In this version only Task categories can act on that evidence. The other four are recorded so their gate line shows how well the classifier does at them. Each one's note says why it offers only Off and Shadow: for Tool-call safety, Phase 0 could not tell how well it does, because the guardian verdicts it was scored against contradicted themselves; Skill routing and Alert triage were never measured in Phase 0, C5's first measurement of the classifier. If you do not want a kind recorded, set it to Off, and that choice is kept.

Step quality reads the step, not its grade. It is given the agent's role, the task's title and the step's name, and the start and end of the step's own instruction and of its output — never the rubric's score it is compared with. (The instruction is what the step was asked to do, before C5 adds the tools to use, the quality rules learned from past reviews and the like, which are much the same for every step of a role.) Earlier versions showed it the rubric's breakdown too, so it was judged with the answer in front of it. The gate therefore counts only steps recorded since your install updated to this version; a rating you gave an older step still counts for the reward model.

A mode someone chose is kept. The defaults apply only to a kind of decision nobody has set. If you set a kind to Off, it stays off. Nothing ever raises a kind into Assisted or Authoritative for you. With the built-in model switched off and no model from your providers in use, none of them is asked, whatever its mode says.

After updating from an earlier version. Earlier versions saved the mode of every kind as soon as anyone changed one of them. So a kind nobody chose could be saved at the old default: Task categories in Shadow, every other kind Off. Once per install, at the first start of a version with the new defaults, C5 looks in its audit log for the kinds someone actually set:

• a kind still at its old default, with no record that anyone chose it, is reset, so it takes today's default (Shadow);
• a kind someone set is kept exactly as they set it, and so is a kind at any mode that was never a default;
• if the audit log may no longer reach back to 27 September 2026, when C5 first saved these modes, C5 cannot tell which were chosen, so it changes nothing. While the classifier is on, it then logs [classifier] site modes kept as stored: the audit trail no longer shows which were chosen, so none is taken for a default.

A kind that was reset is recorded in the audit log as a change made by system. Settings › Learning shows every kind's mode, so you can check the result and change any of them.

On a machine where it runs on the CPU #

Step quality is asked as often as C5 works, once for each step that is graded, and it arrives in bursts: a busy install can grade 20 or 30 steps in one minute. On an Apple-silicon Mac's GPU, one answer takes a fraction of a second. On a CPU it takes a second or more, on the same cores your agents' builds, tests and local models use.

So on a machine where the classifier runs on the CPU (see GPU or CPU), or is marked slow (see Slow machines), Step quality records only one question in four while it is in Shadow:

• The question decides, not the timing. Whether a question is recorded depends on the question itself, so the same question is always recorded, or never. A label you give later always finds the classifier's answer beside it, and Teach the classifier only asks about recorded questions.
• A step outside the sample leaves nothing in the classifier's records: neither the classifier's grade nor the copy of C5's usual grade it is compared with. C5 still grades the step and uses that grade exactly as before. Only the comparison record is skipped.
• Every other kind is recorded every time, on every machine. That includes Tool-call safety, asked once for each tool call the guardian checks, and Skill routing, asked once for each specialist step. Task categories and Alert triage are asked far less often. A kind set to Assisted or Authoritative is never sampled either.
• There is nothing to set. It applies on its own, and the Last 7 days count of Step quality is about a quarter of what it would otherwise be.

Under Step quality, the card says so:

Sampling 1 in 4 on this machine: the local classifier runs on this machine's CPU, so it records only one in 4 of these questions, always the same ones, to keep the rest of C5 fast.

On a slow machine the reason reads the local classifier answers slowly on this machine instead.

When the classifier is busy #

The classifier answers one question at a time. So C5 asks it only one background question at once — a question for a kind in Shadow, whose answer is only recorded — and starts none while it is waiting on the classifier for an answer it will use. Questions C5 is waiting on never queue behind a backlog of background ones.

A background question that finds the classifier busy waits for its turn instead of being dropped, first come, first asked. So a burst of work still records every question chosen for recording (for Step quality on a CPU or a slow machine, the one in four above). Up to 32 can wait, each for up to a minute. One that arrives when 32 are already waiting, or that waits longer than a minute, is given up and not recorded. Giving up is never counted as a failure, so it never pauses the classifier. A question that waited is checked again before it is asked: if the classifier, or that kind of decision, was switched off meanwhile, it asks nothing and records nothing.

A question C5 is waiting on gets at most 1.5 seconds. One that arrives while a background question is being answered waits for it to finish, because an answer in progress is not interrupted, so on a slow processor it can run out of time. That counts as a time-out, not a failure, and C5 uses the usual model's answer.

In the Classifier tab under Ops, background questions given up after a minute's wait are counted as Shed after a minute's wait. Those turned away because 32 were already waiting are counted under Skipped: busy.

When it starts to act #

Only Task categories can be raised to Assisted or Authoritative, and only after the classifier has been measured on your own install. The other four offer only Off and Shadow.

The measurement is called the promotion gate. It compares the classifier's recorded answers with the usual model's, and with the answers people gave when they corrected or labelled tasks.

It counts the last 200 labelled answers, up to 180 days back. A labelled answer is a question both the classifier and the usual model answered, and a person or an outcome labelled. When the last 30 days hold 200 of them, the gate measures those 30 days; otherwise it reaches back as far as it takes to hold 200, but never past 180 days. Only the answers of the model answering now count, so a model new to C5 starts from none. People's labels are read apart from the teacher model's, so a large batch of the teacher model's answers never pushes yours out of the count. The nightly re-check and the calibration fit read the same window, and Teach the classifier asks about task categories up to 180 days old for the same reason.

Its line in Settings says when it was measured and over what, for example Promotion gate: not enough evidence yet (12 labelled states), measured at 14:05 over the last 180 days. Until a kind has its 200 labelled answers, the parts that cannot be judged yet are one line, Needs 200 labelled answers before its accuracy can be judged; has 12., followed by how long the 200 take at the pace of the last 30 days:

• At the current rate (60 labelled in the last 30 days): about 70 days.
• At the current rate: never — 25 new answers came in over the last 30 days, and 180 days at that pace hold 150, however many are labelled. Here labelling cannot help: the kind gets too few questions for 200 to fit in 180 days.
• At the current rate: never — none were labelled in the last 30 days.

A part that failed, Speed included, is still listed beside that line. Once there are 200, the line shows each part that has not passed yet: Labels, Accuracy against your labels, Telling cases apart, Agreement with the teacher model, Calibration, Confident answers and Speed.

Speed counts only where something waits for the answer. Task categories must give 95% of its answers within the 1.5 seconds a new task waits for them. Every question it was asked counts, not only the ones it answered: one that ran out of time counts as slower than the budget. How many questions were asked, and how many timed out, come from C5's own count of every question; how long the answered ones took comes from the answers it recorded. So the detail reads, for example, p95 1,180 ms vs budget 1,500 ms over 412 asks: 400 answered, 12 of 412 asks (2.9%) timed out, each counted above the budget; the asks and the misses from C5's counts, the answered asks' times from 400 answered classifier rows, and a kind that runs out of time on more than one question in twenty fails Speed however fast its answers were. The share is of every question asked, also for a kind that acts, whose recorded answers are only a sample of them.

• A question dropped from the background line is not a miss. Every kind of decision shares that line, so a question given up there says how busy the line was, not how fast this kind's model answers. The detail names them apart: …; 50 shed from the background line every family shares are not counted, as no answer was attempted.
• Refusals while paused or busy are not counted either.
• A new model starts its count afresh. The count begins where the model answering now took over, so a model that replaced another is never charged with the other's time-outs. A model's own time-outs before its first answer still count.

C5 began keeping these counts when your install updated to the version that keeps them; while the window reaches back before that day, the detail ends (partial: C5 has kept its timeouts only since …; one before then is not counted). Nothing waits for the other kinds — their answers are only recorded beside the usual ones — so their speed is reported and never decides their gate, and Ops › Classifier draws no time budget for them: Nothing waits for this family's answers, so the gate holds it to no time budget.

Telling cases apart does not compare the classifier with the usual model. It checks that the classifier's answer actually changes when the right answer changes. A classifier that gives much the same answer every time fails it, even if the usual model does no better. For task categories each category is checked on its own, so a classifier that gives each category the same confidence on every task fails too.

It is measured only on answers that people's labels show to be the correct answer in at least 10 separate tasks (or runs), and not the correct answer in at least 10, whatever the classifier itself said. Until at least one answer has that many labels, this part reads not enough data rather than passing, because a handful of tasks would decide it by chance. Labelling more tasks, in Teach the classifier or by correcting the categories C5 gave a task, fixes that.

• Assisted and Authoritative can be chosen only once the gate has passed, and once a calibration has been fitted for that kind of decision. A calibration is the adjustment that turns the classifier's raw confidence into a measured one. C5 fits it itself, every night at 23:50, from the labelled answers, once there are enough of them.
• C5 keeps a calibration only if it is useful: once calibrated, the classifier's answers must get clearly different confidences, and higher confidence must go with more right answers, not fewer. A fit that fails this is not stored, and the Calibration line under that kind says why. A calibration stored by an earlier version is checked the same way, against the answers the classifier gives now. If it fails, the kind cannot be raised on it, and a kind already acting on it stops acting at its next fit (nightly, or Fit calibration now) and is fitted again.
• Nothing raises the classifier for you. Someone with permission to change settings chooses the mode. A raise is measured first, ahead of the measurements C5 makes in the background, and no other change on the card waits for it. It is judged on the newest measurement there is, the nightly re-check's included, so a kind the night just moved back is judged on the result that moved it. If the kind, or the model it is judged on, changed while it was measured, nothing is saved: … another change to it, or to which model it is judged on, was saved while its promotion gate was being measured. Raise it again to have it judged as it stands now. Nothing was saved.
• Lowering a kind is always saved, by you or by the nightly re-check, however many kinds an older install has stored.
• Even when raised, it acts only at the confidence level its gate measured, only on answers from the model the gate measured, and only on answers calibrated the way the gate measured them. Below that level, the usual model answers.
• Every night C5 re-checks each kind of decision that acts. It moves it back to Shadow, and sends administrators a notification, Local classifier: … moved back to shadow, saying why, when:
  • its gate now fails. This includes Telling cases apart, which v322 and earlier did not measure: a kind raised before the update is moved back the first night after it if its answers do not tell cases apart;
  • the classifier now answers as a different model;
  • the calibration its confidence level was measured on has been replaced; or
  • its calibration error has more than doubled since it was raised.
• If the gate only lacks enough new evidence, it keeps acting for up to 60 days after its last verification: a night it passed on something labelled since the one before. Because the window reaches back for its 200 answers, a kind nobody labels any more would otherwise pass every night on the answers it was raised on; a pass on nothing new counts as too little evidence. Only a new label is new: a question that comes up again and is answered on the label it already had does not re-verify the kind. Administrators are told that its evidence is thin (Local classifier: … is acting on evidence from …, saying that nothing was labelled since it was last verified), and after the 60 days it goes back to Shadow. Answering its questions in Teach the classifier re-checks it.
• If moving a kind back fails part-way, the notification says exactly what changed and what did not, for example Task categories was moved back to shadow. The local classifier's promotion record for it could not be removed., and is titled … could not be moved back to shadow when the kind itself is still acting. The next night judges it again. A night on which none of the checks could run is not counted as done: it is tried again before midnight.
• On a machine where the classifier is too slow (see Slow machines), only kinds set to Shadow run. A kind set to Assisted or Authoritative is not asked at all: it does not even record, so it gathers no new evidence for its promotion gate while the machine is marked slow. To keep collecting evidence there, set it back to Shadow. On Windows you can raise it again once the first timing has cleared the slow mark.

An administrator can raise Task categories on a gate that has not passed, after a warning. That needs an administrator account with MFA (see Who can), on the machine running C5 rather than from a paired device. It is recorded in the audit log, and the setting shows Unvalidated until it is raised again on a passing gate. It still needs a fitted calibration. A kind raised this way is not moved back for its gate result, only for the other reasons above.

Let task categories act #

1. Leave the classifier on, with Task categories in Shadow.
2. Confirm or correct task categories: in the Teach the classifier tab (see below), or by changing the categories C5 gave a task.
3. Wait until the Promotion gate line under Task categories reads passed and a Calibration is listed. C5 fits one every night at 23:50, or an administrator can press Fit calibration now.
4. Choose Assisted or Authoritative.

If a mode cannot be chosen, the option says why; see I cannot choose Assisted or Authoritative.

Teaching it #

The answers people give are what the gate measures. While some model answers the classifier's questions — the built-in model, or a model from your providers that C5 asks — and has something to ask, administrators see a Teach the classifier tab in the HITL: Human-in-the-Loop Signoffs window, which opens from the HITL (clipboard) button in the sidebar (see Human review (HITL)). It is never shown on a paired device, and the account needs a second factor, unless the install was set up without any sign-in. What counts is the administrator role, not where the browser runs: on an install bound beyond this machine (GROWTHER_BIND_HOST), an administrator signed in from another computer sees it too. It shows a task's categories to confirm, or a step's output to rate, with the task and the step it belongs to (a step recorded before this version shows neither, and the start and end of the whole prompt instead). It never shows what the classifier answered, so your answer cannot lean towards it.

One category question in four starts blank. Most category questions start with the usual model's categories ticked: Ticked as the usual model answered. Change anything that is wrong; saving with none ticked records that none apply. A fixed quarter of them, chosen from the question itself, start with nothing ticked instead: One question in four starts with nothing ticked, so your answer checks the usual model too. Tick the categories that fit, or None of these. The usual model is measured on your answers too, and an answer given with its categories already ticked can lean towards them. A question that starts blank — these, and those the usual model could not answer — saves only once you tick a category or None of these: Tick the categories that fit, or None of these, then save. C5 keeps answers to blank questions apart from the others, so the evaluation can tell whether the ticks swayed them.

Tool calls to check. Guardian blocks ask only about calls the guardian refused, so they cannot show how often it lets an unsafe call through, which is where the classifier would act. So the tab also asks about a few of the calls it allowed: a fixed random one in a hundred of those the classifier answered, up to 180 days back. Each card, A call to …, shows the tool's name and its Arguments, with secrets hidden as in Guardian blocks, and neither the guardian's verdict nor the classifier's answer. Answer Was this call safe? with Safe or Unsafe (each says what it means) and press Save, or Skip it for a week. A call that is waiting in Guardian blocks, or that someone has already reviewed, is not asked. Your answer is kept as a human label on the call, apart from the Guardian blocks verdicts.

You need not open the window to find out whether anything is asked. A small grey count on the lower corner of the HITL button says how many questions are waiting here and in Guardian blocks. It is never added to the button's own count and never plays a sound: these questions hold up no work. It is shown only to administrators, never on a paired device, and is hidden while no model answers. See Opening it. The Classifier card in Settings › Learning also says, under each kind of decision, where its labels come from and how many of its questions wait, with a button to the tab (see Each kind of decision).

Three other actions count as answers too:

• Changing the categories C5 gave a task, a task one of your schedules created included.
• A verdict in Guardian blocks, for tool-call safety. Confirm block says the guardian was right to refuse the call; Dismiss says it was a false alarm. They cover only calls the guardian blocked (see Guardian blocks); the tool calls to check above cover the ones it allowed.
• Muting or unmuting an alert, for alert triage. The Auto-Suppression Ledger, in the Errors tab of the same HITL window, has Mute Noise and Unmute beside each repeating error. Muting says "this is noise, hide it"; unmuting says "this should not have been hidden". Only your latest act on an alert counts: unmuting takes back an earlier mute, and the other way round. It counts for an alert C5 triaged while Alert triage was recording. An alert the Sense-Maker (C5's own alert triage) suppresses by itself is not an answer.

Because both kinds of answer can arrive, alert triage can be measured and calibrated like the others.

A model from your providers #

An administrator can choose a model that one of your model providers serves — a llama.cpp or Ollama server on this machine, LM Studio, vLLM, or a cloud API — to answer the classifier's questions, asked before the built-in model. None is chosen by default, and nothing changes until someone chooses one. C5 does not download such a model: it asks it where your provider serves it, as set up in Settings › Integrations.

Who answers #

By default C5 uses the built-in model. Once an administrator chooses a model from your providers, C5 asks that model first. What answers when it cannot depends on two switches on the built-in model's panel:

• The built-in model on, and Use as fallback on (the default): the built-in model answers whenever the chosen model cannot.
• The built-in model on, and Use as fallback off: nothing answers in the chosen model's place, and C5 uses its usual checks. The built-in model still answers what the chosen model is not asked (below). See Use as fallback.
• The built-in model off — for example because its notice does not fit your use: the chosen model is the only one asked. Whenever it cannot answer, no classifier answers, and C5 answers those questions as it did before.

The card says which, under both panels, in these words (M is the model's name):

Where the built-in model answers a kind of decision that the chosen model is not asked, the line names it, for example Tool-call safety is answered only by the built-in model.

"Cannot be asked now" covers a model that is not loaded, is busy, does not answer, is paused after repeated failures, is refused by a policy or by Models allowed from, or whose provider is turned off in Settings › Integrations. A chosen model waiting for a test reads … until M passes its test again, and one that has changed since its test … until M is chosen again. On a machine that cannot run the built-in model, the line reads C5 asks only M: the built-in model cannot run here. …

A chosen model you turned off reads Using the built-in model while M is turned off. (or No classifier while M is turned off. C5 answers as it did before. with the built-in model off). It is not asked at all, and the built-in model answers as it does with nothing chosen.

The chosen model is not asked everything:

• Tool-call safety stays on this machine. Whether a tool call is safe is asked only of a model C5 can show runs on this machine — an address on this machine is not enough, because a program there can pass questions on to another computer. Any other model is never sent those questions: the built-in model answers them while it is on, and C5's usual checks decide either way, as before. Beside such a model, Tool-call safety reads Tool-call safety stays on this machine, so M is not asked it.

  One gap: llama.cpp's --rpc. A llama.cpp server can hand part of its model to RPC servers on other computers (--rpc host:port), which then work on what is derived from each question. A llama.cpp router lists each model's --rpc, and C5 reads it, when the model is chosen and at every check after: RPC servers on this machine still count as this machine, one on your network makes the model Your network, and one further away makes it Online. A model given RPC servers farther away than where it was chosen reads Changed since it was tested. at its next check, and is not asked until it is chosen again. But a single-model llama.cpp server started with --rpc, or a router's model given its RPC servers through LLAMA_ARG_RPC in its environment, does not show it anywhere C5 can read, so C5 counts it as This machine and asks it Tool-call safety too. If you offload a model that way, do not choose it, or keep its RPC servers on this machine. (Only a llama.cpp build with RPC support accepts --rpc; Homebrew's 0.6.0 refuses it.)

• Questions with more choices than the model can read. Some models read only a few choices per question (10 for LM Studio, for example, and fewer for some cloud providers). Skill routing offers one choice for every skill area you have turned on — 11 on a new install — so such a model is not asked it. The card says why beside it, and the built-in model answers it.

Choose a model #

Only an administrator at the machine running C5 can choose or test a model — the same people who can delete the built-in model, never from a device paired through Remote Control.

1. Make sure the provider is turned on in Settings › Integrations. If it runs the model on your machine or network, load the model there first: C5 never loads a model itself.
2. In Settings › Learning, on the Classifier card, under A model from your providers, choose Choose a model… (Change… once one is chosen).
3. Pick a model. The list has one group for each provider you turned on, in the same order as the cards in Integrations, and a search box. A group's heading says where its models run (On this machine, At this machine's address, Your network or Online). A provider C5 cannot ask, or a model it cannot use here, says why. When one reason holds every model in a group, or in the whole list, it is said once, in amber, above them, for example C5 cannot tell that it runs on this machine, and Models allowed from is This machine: set it to Your network on the Classifier card to choose it. LM Studio's models need Your network: C5 cannot see past LM Studio to tell that a model runs only on this machine.
4. For a model on this machine, choose Test and use. For a model on your network or online, choose Next: C5 first shows where it runs and what it would send there, and asks you to confirm (see Before a model off this machine is asked).
5. C5 tests the model, which can take up to a minute and a half. If it passes, it is in use at once, and a message says where it starts.

If something changed while the test ran, nothing is chosen, and the message says what to do:

• M's provider was set up at another address while it was being tested, so its test does not describe it there. Choose it again. Nothing was changed.
• M's provider is no longer set up in Settings › Integrations: it was removed while the model was being tested. Set it up, then choose it again. Nothing was changed.
• Another test of M was already running, and it stopped before it asked anything: test it again. Nothing was changed. Two people choosing or testing the same model share one test; when that test is stopped for the other one's reason, you are told so in these words, never as a policy lock.

The test. The model must answer C5's test questions in a way the classifier can read: it must return, for each answer it could give, how likely it thinks that answer is. The list shows the result of a model's last test beside it:

The badge is worked out the same way whatever serves the model — llama.cpp, Ollama, LM Studio, vLLM or a cloud provider — from that test alone, on your install. A model you have not tested has no badge.

Beside the badge, a model the test could read says how far its answers were from the built-in model's: 11 of 26 top answers differed from the built-in model's, Same top answers as the built-in model; a probability differed by up to 0.31, or Answered its test as the built-in model does. Worked alone says only that C5 can read its answers.

The test does not check that the answers are right. It checks that C5 can read them, and compares them with the built-in model's answers to the same questions. A model trained for decisions elsewhere can differ from the built-in model on many and still be chosen. Whether its answers are right is what its own promotion gate measures, on your install's labels.

A test that could not judge the model — it was not loaded, its provider did not answer — changes nothing either, and says so: M is not loaded in P. C5 never loads a model itself; load it there and try again. Nothing was changed.

LM Studio is the exception. If LM Studio lists the model but has not loaded it, C5 asks LM Studio to load it before the test, and waits for it; the time the load takes is not counted against the test. If LM Studio cannot load it, the message says so in its words: C5 asked 127.0.0.1:1234 to load M, and it could not (…). Nothing was changed.

Test again says its result once, in the card's words (P is the provider's name, as on its card in Settings › Integrations):

Test again asks only where you consented. A model off this machine is tested again only at the address you confirmed for it, and a billed one only after its cost is shown; anywhere else the test sends nothing (above). On this machine, choosing the model was the consent, whatever its address.

C5 tests a model only when an administrator chooses it or presses Test again on its panel, never on a schedule. A model that has changed since its test — new weights at the same name, or a provider moved to another address — is not asked until it is chosen again, which tests it again.

What each provider needs #

Models allowed from decides which of these places a model may run; see below.

Models allowed from #

Models allowed from, on the A model from your providers panel, says how far away a model may run:

Under it, the card always reminds you: LM Studio counts as Your network: C5 cannot show that a model it serves runs on this machine. and Widening this does not skip your confirmation: C5 still asks you before it sends questions to each model that is not on this machine.

• Who can change it. An administrator at the machine. A paired device, or an account that is not an administrator, sees it with the reason it cannot be changed. An administrator without a second factor is refused and told how to add one; Models allowed from, Choose a model… and Test again then stay unavailable until the page is opened again, because each refused attempt is a critical entry in the audit log.
• In the list. A model the setting does not reach says what to change, for example C5 cannot tell that it runs on this machine, and Models allowed from is This machine: set it to Your network on the Classifier card to choose it., or Runs online through OpenAI, and Models allowed from is Your network: set it to Anywhere on the Classifier card to choose it.
• Narrowing it stops C5 asking a chosen model that is now too far away. Its panel reads Not used. with the reason, and the built-in model answers while it is on and Use as fallback is on.
• A model is held where it was tested. A model chosen at a name on this machine or your network is asked only while that name still points there: if it comes to resolve to an address outside, C5 does not connect, and the reason reads gpu-box.lan now resolves to 203.0.113.9, outside your network, where this model was tested to run, so C5 did not connect. A name ending in .localhost is taken for this machine, as localhost is: C5 connects to it here and never asks a resolver about it, offline included, and it counts as This machine only where the provider can show it runs the model here. A request that goes through your organisation's proxy is resolved by the proxy, so this check does not apply to it.
• Your organisation's policy comes first. A policy that locks classifier_external_allowed shows the setting locked, with the reason and where the policy is set; locked to off, it reads Your organisation's policy does not allow a model from your providers here. A value the policy only recommends is the default an administrator may change. off is a value only a policy can give: whether locked or only recommended, the setting then shows a fourth choice, Off — set by your organisation. Recommended, it reads Models allowed from is Off, so C5 asks no model from your providers here: choose where they may run on the Classifier card., and an administrator at the machine can still choose This machine, Your network or Anywhere, after which Off is gone. See Classifier models from your providers.

Before a model off this machine is asked #

A model on your network or online is sent none of your work until an administrator at the machine has confirmed it. To find out where a model runs, C5 only reads what its provider says about itself, with read-only requests (LM Studio's list of models, Ollama's version and model list, llama.cpp's properties and health); no test question is sent before you confirm. After Next (or, if those reads show that the model is not on this machine, after Test and use), C5 shows Send these questions to another machine?:

• Where it runs, for example M runs at 192.168.1.20:8080, not on this machine., M runs online at api.openai.com, not on this machine., or, for a program at this machine's address such as LM Studio, M is served at localhost:1234, an address on this machine, but C5 cannot tell whether the program there runs it here or passes the questions on to another machine.
• What it would be sent: C5 would send it the questions it asks itself while it works: the goal of each new task, the output of steps it grades, and error messages and logs it triages. Tool-call safety questions are never sent there; they stay on this machine. Where the model is asked skill routing, those questions also carry the step a specialist is about to take.
• The part of the Terms that covers it: the paragraph of Section 14 about a provider's model, word for word.
• For a model your provider bills, what its test costs (below).

Tick I have read Section 14 and want C5 to send these questions to H. (H is the model's address) and choose Allow and test. Back returns to the list with your choice kept.

The confirmation is for that address. A model at another address asks again. The Terms dialog then shows the chosen model and when you allowed it, for example M through LM Studio, at localhost:1234. You allowed this on 2026-11-02 under Terms v1.10.

Those questions then go from your machine to that provider like any other work you send it, under your agreement with the provider. They do not go through Growther.si.

Models your provider bills #

The cloud providers bill your account for each question C5 asks their models, the test's included. So C5 says what a test costs before anything is sent:

• Under the provider's group in the list: P bills your account for each question C5 asks its models, a test's included; C5 says what a test costs before sending it.
• In the confirmation, with where the model runs: Testing M sends N questions (about T tokens) to H, which P bills to your account (about $0.02). Once it is in use, P bills each question C5 asks it. If Settings › Budget has no price for the model, the line ends …; Settings › Budget has no price for it, so C5 counts them at $0 until you add one. Allow and test confirms the cost and the confirmation together.
• Test again asks first as well: Test M again?, with the same line, and Test it.

Once a billed model is in use:

• C5 adds what each question costs to Settings › Budget, beside your other spending, under the day it was asked.

• While your Hard Budget Cap is reached and set to block, C5 asks the model nothing. The built-in model answers meanwhile, if it is on and Use as fallback is on. The cap counts these questions with the rest of the period's spending (see How the cap counts); up to v329 the installed app filed them under a date the cap could not read in order, so it could fail to stop them.

• Every question to OpenRouter carries the routing limits you set on OpenRouter's Config tab in Settings — data collection, the providers allowed, ignored or preferred, quantizations and price caps, never its fallback models (see OpenRouter routing settings) — and asks OpenRouter to use only a provider that returns the probabilities. If C5 cannot read those limits, it asks OpenRouter nothing they might rule out. The model's note says so: Every question C5 asks this model carries your OpenRouter routing limits from OpenRouter's Config tab in Settings (…), never its fallback models. A model tested by an earlier version shows that note in these words too, without a new test.

• The list offers the models OpenRouter lists that could be chosen here, not only those under Model Selection on OpenRouter's Config tab, and choosing one never adds it there: the models your chats and tasks fall back to stay as you set them.

  • C5 reads OpenRouter's list only where a model from it could be chosen: with Models allowed from at Anywhere, OpenRouter turned on in Settings › Integrations, and your organisation's policy allowing it. Below Anywhere, the group says why in the setting's own words, for example Runs online through OpenRouter, and Models allowed from is This machine: set it to Anywhere on the Classifier card to choose it., never that OpenRouter lists no models. A policy's refusal is given in its own words, followed by This shows only the models under Model Selection on OpenRouter's Config tab. when that list has some.
  • Only the models your organisation's policy allows are offered, where its allowedModels names some: the others are left out of the list.
  • If C5 could not read the list, the group says so: C5 could not read the models OpenRouter lists., or, beside your Model Selection models, C5 could not read the models OpenRouter lists, so this shows only those under Model Selection on its Config tab. Choosing one here never changes that list. OpenRouter lists no models right now. means OpenRouter answered with none.

  A model OpenRouter no longer lists is refused: OpenRouter no longer lists M. Nothing was changed.

• A name the provider moves to another model by itself cannot be chosen — one ending in latest, openrouter/auto, or a name starting with ~: choose a dated or versioned model instead.

Not every cloud model returns the probabilities the classifier reads, and the list says so where a provider's own documentation leaves it open: xAI ignores the request on grok-4.20 and newer models, Moonshot Kimi and Meta do not document it, and Google no longer returns them for most Gemini models. A test of a model that does not return them ends after its first question, so it costs one question.

Google also refuses every request from a place where it does not offer the Gemini API (User location is not supported for the API use.). C5 then stops asking the model, without counting it as failures, and its panel reads Paused. C5 stopped asking it until its settings change: …

General models #

A general model — one not made for this kind of decision — can still answer, and its badge says how its test went, as for any other model. Every model C5 asks through a chat is one of these: every LM Studio model and every cloud model.

A general model is used like any other. It records its answers in Shadow, and acts on a kind of decision only if its own promotion gate passes on your install, which a general model rarely does. Choosing one does not stop the built-in model acting: until the chosen model's own gate passes, the built-in model's record keeps acting wherever it had been allowed to act, while the built-in model is on.

How a chosen model earns its place #

Every answer is recorded under the model that gave it. Each model has its own promotion gate, calibration and record for each kind of decision: the chosen model is never measured on the built-in model's answers, nor the other way round.

• A model new to C5 starts in Shadow on every kind of decision, from no answers of its own. Its gate line adds It started from no answers when you chose it, so this takes as long as collecting labels does. One of the last three models chosen picks up its own records again.
• Each kind of decision says whose answers count. Beside a chosen model, its row says who is asked now (Asked: M. or Asked: the built-in model., and why not, where the chosen model is not), then M's gate (M — Promotion gate: …), then the built-in model's while it is on (Built-in model — Promotion gate: …). Where the built-in model's record acts and M is still being measured, it reads Acts on the built-in model's answers; M's are recorded beside them until its own gate passes.
• Raising a kind into Assisted or Authoritative is judged on the gate of the model the raise is for: the chosen model's, wherever it is asked that kind. So a kind acts on the chosen model's answers only once that model has been measured, however well the built-in model did.
• Act on M's answers moves a kind that already acts onto the chosen model's own record, once that model's gate has passed on a fitted calibration. Nothing stops acting in between. If the gate refuses, the kind says Not acting on M's answers: …
• Fit calibration now is offered once per model, each named for whose calibration it fits (… on M, … on the built-in model), to an administrator at the machine.
• Teach the classifier stays available while C5 asks the chosen model, even with the built-in model off: then the answers you give there are what the chosen model is measured against.

Turn the chosen model off for a while #

The chosen model's panel has a switch of its own, Enabled, at its top right. Turned off, the model stays chosen, with its test and its records, but C5 asks it nothing until it is turned on again. Meanwhile the built-in model answers as it does with nothing chosen, while it is on.

1. Flip the switch off. A small box under it asks Turn off M? and says what happens: The built-in model answers instead. (or Nothing answers; C5 uses its usual checks. while the built-in model is off or cannot run here), then C5 keeps it chosen, with its test and its records; turn it on again to use it.
2. Choose Turn off. Cancel, where the focus starts, or Escape leaves it on.

While it is off, its panel reads Turned off. C5 keeps it chosen, with its test and its records, and asks it nothing until it is turned on again., followed by what answers meanwhile. The header's badge is then the built-in model's own, and the Primary tag and Use as fallback are not shown.

Who can. Turning it off is open to anyone who may change settings, from a paired device too. Turning it on is for an administrator at the machine running C5, with a second factor, as for choosing a model; anyone else sees why beside the switch, for example Only an administrator can turn on the classifier's model.

Turning it on checks it again. One click turns it on, and C5 first judges it as it would a model being chosen: the Terms, your organisation's policy (and the providers and models it allows), Models allowed from, its provider turned on in Settings › Integrations, a passing test, your confirmation for a model off this machine, and whether it has changed since its test. If anything fails, it stays off, and a message says why:

Change… offers the model again while it is off: choosing it tests it again and turns it on.

Turning it on asks the model nothing: it judges the choice C5 kept. So a model that has changed since it was chosen turns on with a warning, not a test result: M is turned on, but it has changed since it was chosen, so C5 does not ask it. Choose it again to use it.

When the switch itself fails. If C5 could not make the change at all, the message says so once: Could not turn off M; it may still be on. or Could not turn on M; it may still be off. Look at the switch once the card has refreshed, and try again. If the change was made and only the card's refresh failed, you get a warning instead, for example M was turned off, but the classifier's status could not be read. Settings › Learning shows it once it can be. Nothing needs doing: the switch catches up when the status can be read.

Use as fallback #

Use as fallback is a switch on the built-in model's panel, directly under the built-in model's own Enabled switch. It is there while a model from your providers is chosen and turned on, and the built-in model is on. It is on by default. While it is off, the words beside it read Not used as fallback.

• On: whenever the chosen model cannot be asked — not loaded, busy, not answering, paused after failures, waiting for a test, changed since its test, or not allowed for now (its provider turned off, your organisation's policy, or Models allowed from) — the built-in model answers in its place.
• Off: nothing answers in its place, and C5 uses its usual checks, as it did before the classifier. The built-in model still answers what the chosen model is not asked (see Who answers), and still acts where it has a promotion record of its own and the chosen model has none. To keep it from answering anything at all, turn the built-in model off.

Under the switch a line says which: If M can't answer, the built-in model answers. Turn the built-in model off to never use it., or, with it off, If M can't answer, nothing answers in its place, and C5 uses its usual checks. The built-in model still answers where M is not asked or has no record of its own. Turn the built-in model off to never use it. (The last sentence is left out while your organisation's policy decides the built-in model's switch.) The last step of choosing a model shows the same line, before anything is tested.

Who can change it. Turning it off is open to anyone who may change settings, from a paired device too. Turning it back on is for an administrator at the machine running C5, with a second factor; anyone else sees why under the switch, for example Only an administrator can use the built-in model as a fallback. or You can use the built-in model as a fallback only on the machine running C5. Each change is recorded in the audit log.

Your organisation's policy can set it, with classifier_builtin_fallback (see For administrators). Locked, the switch is greyed out, with a Managed badge beside its label, like every other locked control on the card, and Your organisation's policy sets this. The policy is set by … under it.

Use as fallback comes with version 1.10 of the Terms of Service, whose Section 14 says the built-in model need not answer in the chosen model's place (see The Terms that cover it).

Remove, or choose another #

• Remove, on the chosen model's panel, clears the choice. It is open to anyone who may change settings, from a paired device too, whatever your organisation's policy says: removing is always the safe direction. Where a kind of decision acts on the model's answers, C5 asks first (Remove M?) and says what acts there afterwards, for example Task categories acts on its answers now. Removed, the built-in model acts there on its own record; a family without one only records, and C5 answers as it did before. The model's records are kept: choose it again while it is one of the last three models chosen, and it picks them up. Done, it says M was removed. and what answers now. If C5 could not remove it, the message says so once: Could not remove M; it may still be chosen. If it was removed and only the card's refresh failed, a warning says M was removed, but the classifier's status could not be read. Settings › Learning shows it once it can be.
• To stop asking it for a while and keep it chosen, turn it off instead (see Turn the chosen model off for a while).
• Change… opens the list to choose another model. If a kind of decision acts on the current model's answers, the list says what acts there until the new model has been measured.
• Turning its provider off in Settings › Integrations stops C5 asking the model, and the question that turns the provider off says the classifier uses it, and what answers meanwhile: Meanwhile the classifier asks its built-in model if that is on; otherwise C5 answers those questions as it did before., or, with Use as fallback off, Meanwhile nothing answers in that model's place (Use as fallback is off), and C5 answers those questions as it did before. The row then reads Not used: P is turned off in Settings › Integrations.
• A choice that was made but not shown. If the model was chosen and only the card's refresh failed, the list closes and a warning says M was chosen, but the classifier's status could not be read. Settings › Learning shows it once it can be. Do not choose it again: the card catches up by itself.

When the chosen model is not answering #

The model's panel says what is wrong, then what answers meanwhile: Meanwhile the built-in model answers., Meanwhile no classifier answers, and C5 answers as it did before., or, with Use as fallback off, Meanwhile nothing answers in its place, and C5 uses its usual checks.

What the panel says about the model is what C5 found the last time it asked it: nothing reads the provider's server in between. So the line says when that was, for example Answering when C5 last asked it, at 05:52., or ends As of its last question, at 14:03. A server that came back, or swapped its model, shows its old state until C5 next asks it.

Every message on the card names a provider as its card in Settings › Integrations does, for example Not used. Your organisation's policy does not allow the provider Ollama. or The provider LM Studio is not turned on in Settings › Integrations., and a chosen model by its name, never by an internal code: a raise refused on its gate reads Task categories cannot be set to assisted on the chosen model qwen3:0.6b: … Each kind of decision is named as the card names it, in every message and notification: Task categories, Skill routing, Step quality, Tool-call safety and Alert triage.

What you see in Settings › Learning #

The Classifier card has a Terms button and a short status badge in its header. Under Classifier models are two panels. With nothing chosen, the built-in model's comes first; once a model is chosen, its panel comes first, whether it is on or off:

• A model from your providers: what is chosen and where it runs, its last test, its state and what answers meanwhile, Models allowed from, its own Enabled switch at the top right (see Turn the chosen model off for a while), and the buttons Choose a model… (Change… once one is chosen), Test again and Remove. See A model from your providers.
• Built-in model, with its Enabled switch (it reads Disabled while the built-in model is off), Use as fallback under it while a model is chosen and on (see Use as fallback), and the status lines below.

Two tags beside the panels' titles say how they stand. Primary, on the chosen model's panel, is shown only while the built-in model stands behind it: C5 asks the chosen model, the built-in model can run, and Use as fallback is on. With Use as fallback off, nothing answers in the chosen model's place, so there is no order to show and the tag is gone. Fallback, on the built-in model's panel, is shown while it stands behind a chosen model that is answering, with Use as fallback on. Under both panels a line says who answers now (see Who answers). Then comes one row for each kind of decision.

The badge #

While no model from your providers is chosen, or the chosen one is turned off, the badge in the header is the built-in model's, as below. With one chosen and on, the header's badge says how the classifier answers — Running (the chosen model answers), Fallback (it cannot be asked now, and the built-in model answers in its place), Unavailable (it cannot be asked now, and nothing answers in its place: always so with Use as fallback off, even where the built-in model still answers some kinds of decision), Testing (a test of it is running) or Off (your organisation's policy keeps every model off) — and the built-in model's own badge moves into its panel. While a change is being saved, the badge says Turning off…, Turning on… or Saving… instead.

While a change you made is being saved the badge says Turning on…, Turning off…, Deleting the model… or Saving… instead.

The status lines #

The machine line uses your operating system's own figures: the machine's model name, its processor, its physical cores by kind, and its installed memory — what System Information shows on a Mac, and the equivalent on Windows and Linux. On a PC the line starts with the operating system, for example Windows (x64), Dell XPS 15 9520, 12th Gen Intel Core i9-12900HK, 14 cores, 64 GB memory — supported. A figure the system does not give is left out, never guessed:

• On Linux, installed memory can only be read by the administrator account, so the line gives the memory the system can use instead: 31.1 GB usable memory.
• Where the system can use at least half a gigabyte less than is installed, both are given, for example 16 GB memory (15.2 GB usable).

The 8 GB minimum is checked against the usable figure, rounded to the nearest whole gigabyte, so 7.5 GB usable or more passes. Most 8 GB machines report 7.6 to 7.9 GB and qualify. An 8 GB laptop that keeps more for its built-in graphics shows … 8 GB memory (7.4 GB usable) — not supported. with This machine has 7.4 GB of memory; the classifier needs at least 8 GB., and cannot run the classifier.

C5 reads these figures once, the first time the card asks for the classifier's status — never when C5 starts — and keeps them until C5 restarts. It reads no serial number or hardware identifier. On a slow first read (Windows can take a few seconds) the line fills in a moment later.

How C5 reads them, for administrators whose endpoint protection watches new processes:

• On a Mac it runs system_profiler SPHardwareDataType -json -detailLevel mini and sysctl.
• On Windows it starts one powershell.exe -NoProfile -NonInteractive, which reads the registry key HKLM\HARDWARE\DESCRIPTION\System\BIOS and the CIM classes Win32_Processor and Win32_PhysicalMemory.
• On Linux it starts nothing: it reads files under /sys and /proc only.

Each command has 10 seconds. The read happens when the card asks for the classifier's status, and when you run growther classifier status on a machine that can run the classifier. If your endpoint protection blocks it, the line shows only what C5 knows without it (the processor and the memory the system can use), and nothing else about the classifier changes.

The model stays loaded for an hour after its last question on a machine with 32 GB of memory or more (one that reports 28 GB or more counts: Linux and Windows report a 32 GB machine as a little less), and for 10 minutes on a smaller one. Then it is unloaded, and the next question loads it again. Kept loaded, it holds its memory — the Runtime line says how much — while nothing is asked; with only ten minutes, many background questions had to wait for the model to load first. Once it is unloaded, the Runtime line reads Started; the model is not loaded yet and loads with the first question., and the badge Not loaded. A background question that has to wait for the load is given its usual time once the model is loaded, so it is answered rather than given up.

Each kind of decision #

Under each setting you see:

• A short note on what its modes mean there, starting with its default, for example Shadow by default: recorded beside the usual model's answer, never used. Each kind names what it is recorded beside: the router's choice, the rubric's score, the guardian's verdict or the Sense-Maker's suppress decision (see the table in What it does at first).

• Where its labels come from, the answers its promotion gate measures, for everyone who sees the card:

  Kind	The card says
  Task categories	Labels come from an administrator's answers in HITL › Teach the classifier, and from a task's categories corrected by hand.
  Skill routing	No way to label this question yet, so its accuracy cannot be measured: only its agreement with the router is recorded.
  Step quality	Labels come from the step ratings an administrator gives in HITL › Teach the classifier.
  Tool-call safety	Labels come from an administrator's verdicts in HITL › Guardian blocks — only calls the guardian blocked or the classifier sent for a second look, so not a fair sample of its verdicts — and from spot checks of calls it allowed, in HITL › Teach the classifier.
  Alert triage	Labels come from muting or unmuting an alert in the Auto-Suppression Ledger, once the Sense-Maker or the classifier has judged that alert; your latest act stands.

• How many of its questions wait for you, under that line, for Task categories, Step quality and Tool-call safety while some model answers the classifier's questions. An administrator at the machine sees their own count and a button to the tab: 11 category questions waiting in Teach the classifier. with Answer in Teach the classifier, 20+ step ratings waiting in Teach the classifier., or for Tool-call safety 10 of the 57 open calls in Guardian blocks can count toward its gate. with Review in Guardian blocks, and its spot checks apart. With none, it reads No category questions waiting in Teach the classifier now. (and so on) and offers no button. On a paired device it says Answer them in HITL › Teach the classifier on the machine running C5., and to an account that is not an administrator An administrator answers them in HITL › Teach the classifier. The counts are read when the page opens, every five minutes, and when you close the HITL window; a question you skipped is left out, as in the tab.

• Under Step quality, on a machine where it applies, Sampling 1 in 4 on this machine: … (see On a machine where it runs on the CPU).

• Promotion gate — passed, not passed, not enough evidence yet, nothing to measure yet, or not measured yet, when it was measured and over how many days, and each part that did not pass with what was measured (see When it starts to act). C5 measures it in the background, at most every ten minutes. Nothing to measure yet means no answers to compare at all, not answers waiting for labels: Promotion gate: nothing to measure yet: no alert triage decisions recorded over the last 180 days (as of 14:05). when nothing was recorded for that kind, or … it recorded no answers over … when only the usual model's answers were recorded (for example the built-in model beside a chosen model it is not asked in place of).

• How often it agrees with the usual model, once the gate has compared them — never as a bare rate, always beside what always giving the usual model's commonest answer would score, for example Agrees with the guardian on 97% of 298 calls (always answering SAFE would score 97%); flagged 1 of the 9 it blocked — agreement, not a measure of being right. Task categories compares whole sets first: Same categories as the usual model on 0 of 22 tasks (it picks 1.6 a task, the usual model 3.2); agrees on 85% of category decisions, where answering no category would score 81% — agreement, not a measure of being right. Step quality names what gave the usual score: the rubric, the reward model, or, where both scored the steps compared, the score each step was given, followed by That score was the rubric's on N steps and the reward model's on M. Where the usual model was asked the same question more than once, the line also says how often it changed its own answer. This is agreement, not accuracy — only your labels measure that — and nothing acts on it.

• What it is really doing, when it is set to act: Acting in Authoritative at confidence 0.83 or more., Set to Assisted, but only recording its answers. with the reason, or Set to Assisted, but not running: this machine is too slow for its answers to be used in time, so only Shadow runs here. A kind an administrator raised on a gate that had not passed also shows Unvalidated: an administrator raised it on a gate that had not passed …

• Calibration — the one stored, when it was fitted and by what. An administrator on the machine can press Fit calibration now for task categories, tool-call safety and alert triage while the classifier is on, once the fit would find what it needs; until then the reason stands in its place, for example Fit calibration now is offered once there are 200 labelled answers to fit on; there are 12. (or that they must come from two tasks or more, or say both yes and no). Step quality and Skill routing are never fitted — C5 can store and use a calibration only of a yes-or-no answer — so they read Calibration: none. and say why, and Ops reads never fitted (not a yes/no answer). When the last fit was not stored because it could not tell answers apart, the line says so, for example Its last fit, on 2026-10-08, was not stored: calibrated, every answer would get about the same probability, so no confidence level could tell them apart. (see When it starts to act).

• Last 7 days — how many decisions were recorded and how many by the local classifier, and how many labels, of which how many from a person or an outcome: Last 7 days: 1,204 decisions recorded (1,204 by the local classifier), 37 labels (12 from a person or an outcome). Only those last ones measure the classifier.

• For task categories, Saved, last 7 days — how many calls to the usual model its Authoritative answers replaced, and roughly what they would have cost, provider by provider, priced from the usual model's real calls (a provider's replay of a cached answer is not one). Local models are counted in model time, not money, and a model with no price set is never called free.

When a change does not save, or saves in part #

A change on the card saves the moment you make it. If it could not be saved at all, the card says so as an error, for example Could not change the classifier's mode: The classifier's settings could not be saved, so nothing was saved, and nothing changed.

Some changes write more than one thing — raising a kind writes its mode and the model's promotion record — and a fault can stop the write part-way. Then the card says exactly what was kept, as a warning, not as a failure: Only part of the change was saved: the families' modes were saved, and the built-in model's promotion records were not. Settings › Learning shows what is stored now. The card reads the status again, so what it shows afterwards is what is stored.

A change that was saved, where only reading the status back failed, is said as done, also as a warning: The local classifier was switched on, but the classifier's status could not be read. Settings › Learning shows it once it can be. (for a mode, Task categories was set to assisted, but …). Nothing needs doing: the card catches up once it can read the status.

When the card cannot read the models from your providers for a moment, it says so instead of drawing as if none were chosen: C5 could not read the models from your providers just now, so they are not shown here. M from P is still answering the classifier's questions. The badge then reads Running, and the built-in model's switch asks Turn off the built-in model? (With the built-in model off, only M answers: whenever it does not, C5 answers those questions as it did before.) or Turn on the built-in model?. Nothing more is said about the chosen model until it can be read again.

What you see in Ops › Classifier #

The Classifier tab on the Ops page shows the classifier as it runs: whether it is on and acting, how many decisions it made and how fast, the calls to the usual model it replaced, the labels still waiting, and where each learning loop stands. Its Since the server started chart counts every question by outcome, including the background questions given up after a minute's wait (see When the classifier is busy). Only accounts that may view Settings see this tab. It only reports; the switch and the modes are in Settings › Learning. With the built-in model off, the tab says it is showing what the built-in model recorded before.

Whose answers. The charts are the built-in model's. While a model from your providers is chosen, a Whose answers choice beside the kind of decision offers Built-in model or that model by name, and the chosen model's own record — its answers, speed, coverage, savings, labels, gate and history (see How a chosen model earns its place) — shows under Showing M's record: a model from your providers, read-only. Its controls are in Settings › Learning. The built-in model's switch and its process counters are not drawn for it. The choice is kept in the page's address, and a link to a model no longer chosen shows the built-in model's record instead.

The record is named only when it is that model's. Each read shows the record of the model chosen at that moment. If someone chooses another model in Settings while the tab is open, the next read is the new model's, so C5 asks again which model that is, and names it once it knows. Until then the record is not named: Showing the record of the model from your providers chosen now, read-only. Its controls are in Settings › Learning.

For the kind of decision you pick, it also shows:

• Answer time per day, with the gate's 1.5-second budget drawn for Task categories only: nothing waits for the other kinds' answers.
• The usual model's calls: Last 14 days, beside the classifier's answers. Wall time runs from the call to its answer, waits included. — the number of calls, the mean wall time and the mean tokens in and out per call. A replay, an answer a provider repeated from its cache, made no call, so it is left out and counted apart. The panel says whether the classifier can replace these calls: only Task categories can, while it is Authoritative. The figures are context for the classifier's speed, not a saving.
• Shadow coverage: of the usual model's questions the classifier is asked about, how many it answered (Answered 1,120 of 1,180 questions (95%).), day by day, and what became of every question, kept across restarts: Answered, Timed out, Timed out while the model loaded, Dropped: too many waiting, Refused while paused, Refused: busy, Not running, Skipped: language, Switched off, Skipped, Failed, Not asked: its settings did not answer in time, or, for the built-in model beside a chosen model, Not asked: Use as fallback is off (… was off once it is on again: the counts cover the whole window). A question asked while the classifier or that kind was switched off counts as unanswered, and the last 5 minutes are left out, since a question may still be waiting for the model to load. C5 began keeping these counts when it updated to this version; until that day is older than the panel's window, the heading says so (Every question since Oct 15 (UTC), when counting began, kept across restarts). A question in a language the built-in model skips (see Languages) is counted Skipped: language with Use as fallback off too, since that, not the setting, is why it went unanswered.

One answer is counted once. When the usual model fails on a task's categories and the classifier's confident answer is used instead, C5 records that answer twice — as the answer used, and as a sample for the promotion gate — but the daily answers and answer times count it once.

Saved counts only the calls the usual model actually made, never a replay, and prices them at the usual model's fresh calls. The early guess C5 makes at a draft's categories in Chat counts as an avoided call too, when the classifier answers it in Authoritative, but it is left out of the labels waiting: no label can reach a draft.

Labels, for the kind of decision you pick, shows its labels per day by where they came from (only a person's answer and a run's outcome are truth), how far it is from its gate's floor (Toward the gate's floor), Where labels come from (the card's own sentence, above), and two counts that are not the same:

• Unlabelled answers: every question it answered in the window that no person or outcome has labelled. This is the gate's backlog.
• Askable now: what Teach the classifier can actually ask about this question, from as far back as it asks, at most a page of 20 (20+ beyond that). It is smaller when some answers belong to a task that has gone. A question someone skipped still counts here, since the count is shared; nothing is asked while no model answers.

It also counts the held-out answers an acting kind needs to stay verified. Skill routing and Alert triage have no questions in Teach the classifier, so their panel shows only where their labels come from until some arrive. While Askable now is above 0, an administrator gets Answer in Teach the classifier, which opens that tab of the HITL window and waits for its questions to load; if they cannot be read, the tab says why. On a paired device the panel says Answer them in Teach the classifier on the machine running C5., and to anyone else that an administrator answers them. For tool-call safety the waiting count is far larger than Teach asks about, and the panel says so: Teach the classifier asks only spot checks of these: about 1 in 100 of the calls the guardian allowed, picked at random. The calls it blocked, and those the classifier sent for a second look, are answered in Guardian blocks. The rest are never asked.

Analytics › Classifier and Labs › Classifier #

Two more tabs show the classifier's record in detail. Both need permission to view Settings, and both show the built-in model's record, with the same Whose answers choice as Ops while a model from your providers is chosen.

• Analytics › Classifier follows one kind of decision at a time (Family), over 7, 14 or 30 days (Window): how it does against the usual model on the answers people or outcomes labelled, with 95% intervals; its gate's history, day by day; a calibration chart of its confidence against how often it was right; how often it gives the same answer as the usual model, week by week, beside a dashed line for what always giving the usual model's commonest answer would score (agreement, not accuracy), and where the two differ — each answer's counts, Step quality by the step's role and tool-call safety by tool, Step quality's weighted κ and distance in levels, and how often the usual model changed its own answer; Cohen's κ against each source of answers; and how the questions people answered were chosen. For Step quality it also compares the reward model with people's ratings, and its card comparing the classifier with the usual score is titled by what gave that score: The classifier against the rubric, The classifier against the reward model, or, where both scored the steps compared, The classifier against the score each step was given, with That score was the rubric's on N steps and the reward model's on M. It reads once when opened; press Refresh to read again.
• Labs › Classifier has What a threshold would do: pick a confidence threshold (τ) and see how many decisions it would cover, how many tasks the classifier would answer alone, the error it would make and the calls it would replace a week. This changes nothing. Beside it, Evidence budget says how many questions it is asked a week and how many have been labelled so far.

The Terms that cover it #

The Terms button on the card opens The classifier — Terms of Service, Section 14: section 14 of the Growther.si Terms of Service, Models C5 downloads and runs on your machine, word for word, under the Terms version and its effective date (Terms v1.10 · effective October 22, 2026 in this release). Below it are this install's built-in model, its size, its licence, whether it is on this machine, and its notice. The On this machine row says what is at the model's path, without claiming C5 downloaded it (you may have placed it by hand):

While the model is not here (or the file there is not the model), on a machine that will not download it (see Turn it back on), a line under the notice says so and why: On this machine it is not downloaded: …

With a model from your providers chosen, a Your model row names it too: M through Ollama, on this machine., or, for a model off this machine, where it runs and when an administrator allowed it, for example M through LM Studio, at localhost:1234. You allowed this on 2026-11-02 under Terms v1.10.

At the bottom are Full Terms of Service (growther.si/t=models), Third-party notices (the licences of everything C5 ships or downloads), Turn off the local classifier… (Turn off the built-in model… while a model from your providers is chosen) and Close. That button is offered to someone who may change settings, while the built-in model is on and not locked by policy. Reading the Terms asks you to accept nothing.

In short, section 14 says: C5 downloads these models from Hugging Face only while the feature that needs one is on, checks each one before using it, keeps it on your machine, and sends nothing about it or your use of it to Growther.si. Each model stays under its publisher's own licence. The classifier is on by default, first only records, and acts only after it has been measured on your install and only for the kinds of decision allowed in Settings.

Section 14 also covers a provider's model, if you choose one: an administrator at the machine may choose a model one of your providers serves, and none is chosen by default. C5 does not download it. That model, and the questions C5 sends it, are governed by your agreement with the provider and the model's own terms; Growther.si is not a party to that and does not receive them, and the provider may charge you for them. It is measured on its own record and acts only where it has been measured. While the built-in model is on, it keeps acting where it has been measured and the chosen model has not, and answers whenever the chosen model cannot be reached, unless its use as a fallback has been turned off in Settings › Learning or by your organisation's policy. While the built-in model is off, or its use as a fallback is turned off, no classifier answers when the chosen model cannot be reached, and C5 answers as it did before. (Version 1.10 of the Terms added the words about its use as a fallback.) A model not served on your machine can be chosen only where your organisation's policy allows it, and only after C5 has shown what it will send there and you have confirmed. Questions about whether a tool call is safe are never sent to such a model.

Machines it cannot run on #

The classifier runs on macOS (Apple silicon and Intel), Linux x64, and Windows x64 and arm64, with at least 8 GB of memory. Where it cannot run, C5 itself downloads nothing and starts nothing; the badge reads Unsupported and the first status line (This Mac or This machine) says why. (Only growther classifier pull downloads on such a machine, because you asked it to; see Command line.)

growther classifier status and growther doctor show the same reason.

Slow machines #

Each time the classifier starts, once the model has loaded, C5 times how long the model takes to answer one question about a text of about a page. It times the model's own work, not how long C5 waited: a question that had to wait behind others, or behind the model loading, says nothing about the machine. If the answer takes more than two seconds, the machine is marked slow: the first status line adds This machine is slow for it, so its answers may arrive late. On a slow machine only kinds of decision set to Shadow run; one set to Assisted or Authoritative is not asked at all, and records nothing either. Step quality also records only one question in four there (see On a machine where it runs on the CPU).

A slow result is measured again. A machine busy with something else at that moment — other programs starting, say — can time ten times slower than it is. So when the first timing is slow, C5 times it again, quietly, two minutes and fifteen minutes later. Each retiming waits until the classifier has nothing to answer and its model is loaded, and it never loads the model just to time it. The machine stays marked slow only while the fastest timing is over two seconds: a fast one clears the mark at once, with no restart, and kinds set to Assisted or Authoritative start acting again. Every restart of the classifier times it afresh.

Windows starts out marked slow, because the classifier has not yet been timed on Windows hardware. So until that first timing, task categories can only be recorded (Shadow) on Windows. You do not need to do anything: the timing happens on its own the first time the model loads, and a machine that answers in two seconds or less clears the mark.

GPU or CPU #

On Apple silicon the classifier uses the GPU (Metal), unless LM Studio, llama.cpp, Ollama or Inferencer is one of the model providers C5 is set up to use (turned on in Settings › Integrations, or listed in the Coordinator Priority card there). C5 assumes such a provider runs its model on this machine's GPU, so the classifier gives way and uses the CPU. C5 decides this by the provider's name, not by its address: Inferencer pointed at Hugging Face's cloud, or LM Studio on another machine, also moves the classifier to the CPU. To give the GPU back to the classifier, remove that provider from C5's providers there. Everywhere else the classifier uses the CPU, with up to 8 threads, leaving two cores free. A Mac with no usable GPU, such as a virtual machine, is moved to the CPU automatically the first time the classifier loads.

A change of providers applies by itself. When you add or remove such a provider while the classifier runs, C5 restarts the classifier on the new choice about two seconds after your last change; several changes in a row make one restart. The restart unloads the model, and C5 loads it again straight away, timing one answer on the new choice (loading takes about a second on Apple silicon). A change that does not move it between GPU and CPU restarts nothing. With the classifier off, or without its model, nothing restarts and nothing is downloaded; the choice applies when it next starts.

The card says which, and why, under the machine line. The thread count is the classifier's own, not your machine's:

The card describes where the classifier runs now. Just after you change providers, it also says where it is about to go, for example Runs on the GPU (Metal). It moves to the CPU when the built-in model restarts, because LM Studio is configured and the built-in model leaves the GPU to your local model., or … It moves to the GPU (Metal) when the built-in model restarts. You do not need to restart anything: C5 restarts the classifier itself within a few seconds, and the line then shows where it runs.

On the CPU, Step quality records only one question in four (see On a machine where it runs on the CPU); every other kind is recorded every time. That goes by where the classifier runs now, not where it moves next.

Languages #

The model is English-only. Text that is mostly outside the Latin alphabet (for example Hindi, Thai, Japanese or Cyrillic) is not sent to it; the usual model answers. French, Vietnamese and other languages written in the Latin alphabet are still asked.

Offline and air-gapped installs #

With the Egress posture set to Offline, C5 never downloads the model — not at start, not from Settings, and not from growther classifier pull. Put it on the machine yourself. The classifier can be on or off while you do this; it uses the file once it is on.

1. On a machine with internet access, download exactly this file:

   https://huggingface.co/mradermacher/decider-0.8b-GGUF/resolve/001df41ddc126176cab96c8aec9ada45008fbe62/decider-0.8b.Q8_0.gguf

2. Check its SHA-256 fingerprint. It must be:

   2665d08c1052b4e01dabcb08771d25579f6776f7066355e4a0ccc5f74e32d4a2

   shasum -a 256 decider-0.8b.Q8_0.gguf                 # macOS
   sha256sum decider-0.8b.Q8_0.gguf                     # Linux
   certutil -hashfile decider-0.8b.Q8_0.gguf SHA256     # Windows

3. Copy it, with exactly that name, to the classifier's models folder under C5's cache folder: ~/.growther/classifier/models/decider-0.8b.Q8_0.gguf by default, or %USERPROFILE%\.growther\classifier\models\decider-0.8b.Q8_0.gguf on Windows. Create the folder if it does not exist. If you moved C5's caches, run growther cache show to find the folder.

   To share one copy between several installs on macOS or Linux, you can put a symbolic link at that name instead of the file. C5 checks the file it points to:

   ln -s /shared/decider-0.8b.Q8_0.gguf ~/.growther/classifier/models/decider-0.8b.Q8_0.gguf

   C5 never downloads over a link. Not at start, not from Settings, and not from growther classifier pull: the link is yours, and replacing it would cut this install off from the shared copy.

   • If the file the link points to cannot be reached — a share that is not mounted, say — C5 waits. It uses the model as soon as that file can be reached. Mount the share, or remove the link if you would rather C5 downloaded its own copy.
   • If the link points to a file that is not the model, C5 does not use it and does not replace it. Point the link at the model, or remove it to have C5 fetch the model.
   • If you place a link while C5 is downloading the model, C5 leaves the link alone and keeps its checked download beside it as decider-0.8b.Q8_0.gguf.partial. Remove the link and run growther classifier pull to put that copy in place.

4. Run:

   growther classifier status

   It checks the file and should print ✓ model present and verified. If it prints anything else, see Troubleshooting.

You do not need to restart C5. While the classifier is on, a running C5 notices the file the next time Settings › Learning is opened, checks it and starts the classifier. Otherwise it picks it up at its next start.

How C5 checks the file. First its size, which must be exactly 811,844,032 bytes. Then its SHA-256 fingerprint, worked out once. C5 records the result in a small file beside the model, decider-0.8b.Q8_0.gguf.verified.json, so it does not read all 812 MB again at every start; if the model file changes, it is checked again.

A wrong file at that name — the wrong size, the wrong contents, or a copy still in progress — is never loaded. C5 leaves it where it is and does not delete or rename it. The Model line says The file at the model's path is not the model (it is N MB; the model is 812 MB), so C5 does not use it. (or its checksum does not match), then what happens to it and what you can do; growther classifier status says … is not the pinned model (wrong size) or (sha256 mismatch), and growther doctor warns about a file of the wrong size. Once C5 has worked out that a file's fingerprint does not match, it remembers that until the file changes or C5 restarts, so it does not read all 812 MB again each time the classifier is turned on.

On a machine that may download, the next verified download replaces a wrong file (never a link). Under Offline (and with automatic downloads turned off, or on a CI runner) nothing replaces it, so delete it and copy the right file again. There C5's start-up log line — and, under Offline, growther classifier pull — begins its reason with The file at … is not the classifier model (wrong size), so C5 does not use it; replace it. (or its sha256 does not match), followed by the placement step.

Privacy #

• The questions, the classifier's answers and the labels people give stay in C5's database on this machine. The built-in model runs on this machine, in its own process.
• The built-in model's only network traffic is the one download of the model from Hugging Face. C5 sends nothing about the model, or your use of it, to Growther.si.
• A model from your providers, if an administrator chooses one, is sent the classifier's questions like any other work you send that provider: from your machine to the provider, under your agreement with it, never through Growther.si. A model off this machine is sent none of your work until an administrator has confirmed it, and is never asked whether a tool call is safe. Its answers are recorded in C5's database on this machine, like the built-in model's.
• How long it is kept. The classifier's records of its answers, and the usual model's beside them, are deleted 180 days after they were made, every night at 23:55 (not while a legal hold is on). The labels people gave are kept. The text of a step that Step quality was asked about (its task, step name, instruction and output) is deleted 31 days after the step unless someone rated it there; its scores and its agent's role are kept. The daily counts of what became of each question (the ones Ops › Classifier and the speed check read, which hold no text) are deleted after 400 days, on the same nightly pass.
• What is kept, and what is masked. Secret values (API keys, tokens, private keys) are masked in that text before it is stored. A guardian decision's record keeps its verdict, the tool's name and the length of its reasoning; the reasoning itself stays only on the audit log. A guardian block's stored reason keeps the first 50 characters of the call, with secrets masked in the whole call before it is cut, and the reasoning, masked too. A skill-routing record keeps what the router chose, never the text of its reply.
• Who sees it. Teach the classifier shows administrators any task's title and the start of its description, a step's instruction and output, and an agent's tool call with its arguments, secrets masked. That is how a question can be answered without the classifier's answer beside it.
• C5's status address, GET /health, answers without any sign-in, so that supervisors can see C5 is up. It includes the classifier's details (use counts, answer times, how many questions are in progress or waiting, whether it is paused) only for someone signed in to C5 on this machine or its network, or for a request made directly on the machine itself. Everyone else gets "classifier": null, including a phone or tablet paired through Remote Control and anyone reaching C5 through a Remote Control address.

The binding text on where these questions go is the Privacy Policy, under Local by default, and Section 14 of the Terms (see The Terms that cover it).

For administrators #

Managed policy: classifier_enabled #

classifier_enabled is a boolean key you can set and lock through Group Policy or Intune, macOS managed preferences, /etc/growther/policy.d, or a signed policy document. See Managed configuration.

• Locked to false: the classifier stays off, and C5 never downloads the model — not even through growther classifier pull. An administrator can still delete a model that is already there. No model from your providers answers either, unless the policy also sets classifier_external_allowed at the same level or a higher one (see below).
• Locked to true: it stays on. Nobody can turn it off or delete the model in Settings. If the file at the model's path is not the model, the card says so and adds Your organisation's policy keeps the classifier on, so the file cannot be deleted here. (for a link, so the link cannot be removed here). An administrator at the machine can put the model in its place (or point the link at it).
• Either way, the switch is shown disabled with a Managed lock and a line saying why. Locked on: Your organisation's policy keeps the local classifier on, so it cannot be turned off here. Locked off: Your organisation's policy keeps the local classifier off, so it cannot be turned on here. Each is followed by The policy is set by … and where it came from.
• It applies without a restart. After each policy refresh (every fifteen minutes, growther policy reload, or Reload policy under Settings › Enterprise), C5 starts or stops the classifier to match, cancelling a download in progress when it is turned off.
• Before the first start: when a policy already sets the key the first time C5 starts, C5 leaves the switch to the policy and sends no "is on" notification. Set it to false there and the model is never downloaded.
• If a signed policy can no longer be fetched and its grace period (how long C5 keeps using the last signed policy it fetched; fourteen days by default) runs out, a locked classifier_enabled falls back to false, its safe default. See How a policy is applied.

Managed policy: classifier_external_allowed #

classifier_external_allowed (off, loopback, lan or any) decides how far away a model from your providers may run. It is delivered the same ways as classifier_enabled. Locked, it replaces Models allowed from on the card, which is then shown locked with the reason; only recommended, it is the default an administrator at the machine may change; not set at all, Models allowed from starts at This machine. allowedProviders and allowedModels apply to the classifier's models too. See Classifier models from your providers.

Managed policy: classifier_builtin_fallback #

classifier_builtin_fallback (a boolean) decides whether the built-in model answers in place of a chosen model that cannot be reached: Use as fallback on the card (see Use as fallback). It is delivered the same ways as classifier_external_allowed, never by the Mothership, and applies without a restart.

• Locked to false: the built-in model never answers in the chosen model's place, and nobody can turn Use as fallback on in C5. It still answers what the chosen model is not asked, and acts where it has a record of its own and the chosen model has none.
• Locked to true: it always does, and nobody can turn it off in C5.
• Recommended, or under User Configuration: a default, which anyone allowed to change it on the card may change.
• Not set: on, until someone changes it on the card.
• A locked value that cannot be read counts as false.

It matters only while a model from your providers is chosen and turned on, and the built-in model is on. Builds that ship Terms of Service earlier than v1.10 ignore it. See Classifier models from your providers.

Downloads only by hand or on request #

Start C5 with GROWTHER_CLASSIFIER_NO_AUTO_DOWNLOAD=1 (exactly 1) and C5 does not download the model on its own. It then arrives only when someone runs growther classifier pull, or by hand. C5 also never downloads it automatically when the CI environment variable is set. See Environment variables.

Network #

A model from your providers is asked at the address its provider is set up with in Settings › Integrations, never at an address taken from anywhere else, and through the same proxy settings as the rest of C5. While the Egress posture is Offline, C5 asks — and tests — only a model at an address on this machine.

The built-in model's download goes through the proxy and certificate authorities set under Settings › Enterprise › Network, like every other C5 download (see Corporate proxy). huggingface.co answers a model request by redirecting to its download network, so allow these hosts as well — a proxy that allows only huggingface.co makes every model download fail:

• huggingface.co
• cas-bridge.xethub.hf.co
• cdn-lfs.hf.co
• cdn-lfs-us-1.hf.co
• cdn-lfs-eu-1.hf.co
• cdn-lfs.huggingface.co

When the download fails on the network, the message gives the cause itself. The Model line in Settings › Learning reads, after The download failed:

Could not download the classifier model: the proxy refused to open a tunnel to huggingface.co:443: it answered CONNECT with 407 Proxy Authentication Required; it wants credentials — put them in the proxy URL (http://user:password@host:port), C5 speaks Basic proxy authentication only. Once that is fixed, turn the classifier off and on, restart C5 or run `growther classifier pull` to download it; or fetch https://huggingface.co/… on a connected machine and place it at … (sha256 …).

The card never shows your proxy's address. growther classifier pull prints the same message in your terminal with the proxy named (the proxy http://proxy.corp.example:3128 refused …), and growther doctor shows which proxy is configured. See The proxy address stays out of logs and alerts.

When the cause is the proxy (it refused, or its setting cannot be used) or a certificate, the machine is connected and a network setting is what to fix. So the message says what starts the download again once it is fixed — C5 does not retry on its own until it next starts — and offers placing the file by hand as the alternative. (From growther classifier pull run with GROWTHER_CLASSIFIER_NO_AUTO_DOWNLOAD=1 or CI set — most build servers set CI — it names only running the pull again as the way to download, because C5 would not download on its own there, and still offers placing the file by hand.) The other causes end with the steps to place the file by hand. How each cause reads:

• the proxy's answer, as above;
• a certificate C5 does not trust: the certificate presented on the way to huggingface.co:443 is not trusted (…) through a proxy, or the system's own words, such as self-signed certificate in certificate chain, without one;
• a proxy setting C5 cannot use: C5 refused to connect to huggingface.co:443: the configured proxy cannot be used (growther doctor says why). …;
• a host that refused the connection: connect ECONNREFUSED …, or the connection to huggingface.co:443 through the proxy failed (ECONNREFUSED).

See When the proxy says no.

With the Egress posture set to Offline, C5 logs this line, at most once an hour, when it skips the download:

[posture] offline — local classifier model download skipped; this deployment makes no outbound contact

See Network and egress.

Command line #

growther classifier status
growther classifier pull

status checks whether this machine can run the classifier and whether the model is here and checked. It works out the model's SHA-256 fingerprint if that has not been done, which reads all 812 MB and can take a few seconds. It prints, for example:

✓ supported on Mac Studio (darwin-arm64): Apple M4 Max, 16 cores (12 performance and 4 efficiency), 64 GB memory
  GPU: Metal, unless a local model (LM Studio, llama.cpp, Ollama, Inferencer) is configured — then CPU
✓ model present and verified — /Users/ana/.growther/classifier/models/decider-0.8b.Q8_0.gguf
  decider-0.8b-q8_0, 812 MB, context 4096
  sha256 2665d08c1052b4e01dabcb08771d25579f6776f7066355e4a0ccc5f74e32d4a2

The first line describes the machine as the card does, from the operating system's own figures: its name where the system gives one, then the platform, processor, physical cores and memory, for example ✓ supported on linux-x64: AMD Ryzen 9 7950X, 16 cores, 31.1 GB usable memory.

On Apple silicon the second line gives both possibilities, as above, because this command does not read C5's settings and so cannot see which model providers are configured. The This Mac line in Settings › Learning says which one applies (see GPU or CPU). Off Apple silicon, the second line says why it runs on the CPU, and with how many threads, for example CPU: GPU acceleration is only used on Apple silicon (this is linux-x64). It uses 8 CPU threads.

The model line can also say:

• ✗ model not downloaded (run `growther classifier pull`, or place … at …)
• ✗ model not downloaded: this deployment's network posture is offline, so C5 does not download it., then the address, path and fingerprint to place it by hand — the Egress posture is Offline.
• ✗ model not downloaded: your organisation's policy keeps the local classifier off
• ✗ <path> is a link whose file cannot be reached (a share that is not mounted?); it is used once it can be, and never downloaded over
• ✗ <path> is not the pinned model (wrong size) (or sha256 mismatch), followed by what happens to it: `growther classifier pull` replaces it for a file; for a link to the wrong file, it is a link, and C5 does not replace a link: point it at the model …, or remove it to have C5 fetch the model; under Offline or a policy that keeps the classifier off, the same reason as the lines above.

status does not suggest a pull that Offline, your organisation's policy or a link would refuse. To check, it reads your organisation's policy whenever the model is not here and checked, as pull does: the device's own policy and, if you use one, the signed policy document. If that document is at a web address, each such status run makes one request to it, even while C5 is set to Offline. status never creates or changes a database.

It exits with 0 when the machine is supported and the model is verified, and 1 otherwise.

pull downloads the model now. It prints the model's notice, then, once the first bytes arrive, About 812 MB is downloaded once from Hugging Face. and Downloading decider-0.8b-q8_0 (~812 MB) to <path>…, then progress in 5% steps, and finally ✓ downloaded and verified — <path>. A pull that is refused, or finds the model already here, prints neither of those two lines. If the model is already here and checked, it says ✓ already present and verified — <path> and downloads nothing.

• It exits with 0 when the checked model is in place (downloaded now or already there) and 1 when it is refused or fails. The reason is printed on one line starting with ✗.
• It does not check whether this machine can run the classifier. Run growther classifier status first: on an unsupported machine the model downloads but never loads.
• It works while the classifier is switched off, and while GROWTHER_CLASSIFIER_NO_AUTO_DOWNLOAD=1 is set: running it is your explicit request.
• It goes through the proxy and certificate authorities set for C5.
• It is refused when the Egress posture is Offline, and prints the address, folder and fingerprint to place the file by hand instead.
• It is refused when your organisation's policy locks the classifier off: Your organisation's policy keeps the local classifier off, so its model is not downloaded.
• It resumes an unfinished download, and replaces a wrong file at the model's name once the new one is checked.
• It is refused when the model's path is a symbolic link whose file cannot be reached, or a link to a file that is not the model: C5 never downloads over a link. The message says what to do: make the file reachable, point the link at the model, or remove the link.
• If C5 itself is downloading the model at the same time, pull says another C5 process is downloading and stops; C5's download carries on.
• When the download fails, the ✗ line gives the cause itself, naming the proxy if one refused (see Network).

While the classifier is on, a running C5 starts using a model you pulled the next time Settings › Learning is opened, or at its next start.

growther doctor #

growther doctor has a Local classifier row. Doctor does not work out the model's fingerprint; it reads the record C5 keeps beside it.

A missing model is never a failure: C5 works without it. Doctor does not tell a symbolic link whose file cannot be reached from a missing model: both read model not downloaded. growther classifier status tells you about the link.

Log lines #

C5 prints these lines to its own output:

While the classifier is on, C5 may print:

C5 prints no [mcp] Connected … line for the classifier. Switched off, or on a machine that cannot run it, the classifier prints nothing.

Troubleshooting #

The badge says No model and the Model line says Not downloaded: This deployment's network posture is offline, … The Egress posture (this is the "network posture" in the message) is Offline, so C5 will not download it. Place the file by hand (see Offline and air-gapped installs), or ask whoever manages Settings › Enterprise › Network.

Not downloaded: GROWTHER_CLASSIFIER_NO_AUTO_DOWNLOAD=1: the model is not fetched automatically. … Automatic downloads are turned off for this install. Run growther classifier pull, or place the file by hand.

Not downloaded: Not downloaded on a CI runner. The CI environment variable is set in the environment C5 started in. Any non-empty value counts, even false or 0. On a build server this is intended: run growther classifier pull if you want the model there. On any other machine, remove CI from the environment C5 starts with (your shell profile, or C5's start-at-sign-in definition; see When C5 starts at sign-in), then restart C5.

The badge says Error and the Model line says The download failed: Not enough free space to download the classifier model: it needs about N MB under …, and M MB are free. … C5 did not start the download, because the disk could not hold it. Free some space in that folder, then restart C5 or turn the classifier off and on.

The download failed: Could not write the classifier model to disk (…); the partial download was removed. … The disk filled up during the download, and C5 removed the unfinished file to give the space back. Free space in the folder named, then restart C5, turn the classifier off and on, or run growther classifier pull.

The download failed: Could not download the classifier model (HTTP 403). … or The download failed: Could not download the classifier model: <reason>. … C5 could not reach Hugging Face. The reason is the network's own: what the proxy answered (a 407 wants credentials in the proxy URL), a certificate C5 does not trust (add your proxy's root certificate to the CA bundle), a host that refused the connection, or a proxy setting C5 cannot use (… the configured proxy cannot be used … — fix or remove the proxy URL; run growther doctor to see why) — see When the proxy says no. The card says the proxy without its address; growther doctor shows which proxy is configured. Otherwise check your proxy, and that it allows the Hugging Face hosts. The message ends with the address, path and fingerprint to place the file by hand. To retry, turn the classifier off and on, restart C5, or run growther classifier pull.

The download failed: The downloaded classifier model does not match its pinned sha256 (got …); nothing was installed. What arrived was not the model, often because a proxy changed or cut it. Nothing was kept. Try again; if it repeats, download it by hand on another network and check its fingerprint yourself.

The classifier model downloaded incompletely (N of M bytes); the next attempt resumes it. The connection dropped. The next attempt carries on from where it stopped.

The download failed: The download did not answer with the classifier model: it sent an HTML page (a proxy, a captive portal or a block page answers like this). Nothing was kept. … Something between this machine and Hugging Face answered in the model's place: a proxy's block page, a hotel or office sign-in page, or a file of another size (… it sent N bytes where the model is M …). C5 stops at once and keeps none of it, so the next attempt does not build on it. Ask for the Hugging Face hosts to be allowed, sign in to the network, or place the file by hand as the message's last sentence says.

The Model line says Already attempted in this process; restart C5 or run growther classifier pull to try again. The download at start failed and C5 does not retry it until the next start. Turning the classifier off and on also retries it.

The badge says Waiting: Another C5 process … is downloading the classifier model; it is used once that finishes. Usually growther classifier pull in a terminal. Let it finish. A download left behind by a process that crashed does not block C5: it notices the process is gone and takes over.

The Model line says … its checksum has not been verified yet. and the badge says Not running A file of the model's exact size is at its name, but C5 has not checked its fingerprint yet. Nothing loads an unchecked file. Run growther classifier status to check it now; if it is not the model, the line then says so (next entry).

The Model line says The file at the model's path is not the model (…), so C5 does not use it. Something else is at the model's name: a copy that failed or is still in progress, another file, or a file whose fingerprint C5 found wrong. C5 never loads it, and never deletes or renames it on its own. The rest of the line says what happens next and what you can do:

• Only a verified download replaces it. — the classifier is on and this machine downloads: the next download that passes its check takes the file's place.
• Turning the classifier on downloads the model over it. — the classifier is off; once on, this machine downloads the model in the file's place.
• It is not replaced by a download here: … — Offline, opted out, or short of disk: delete it and copy the right file in (see Offline and air-gapped installs).
• C5 never downloads over a link. — the model's path is a link to the wrong file: point it at the model, or remove it.

An administrator at the machine can delete it with Delete the file (N MB) while the classifier is off, or Turn off and delete the file while it is on. Anyone else is told that an administrator at this machine can do it.

The Model line says The model's path is a link whose file cannot be reached (a share that is not mounted?). … The model's name is a symbolic link, and the file it points to is not there right now. C5 does not download over the link: it uses the model as soon as that file can be reached. Mount the share (or restore the file), or remove the link if you would rather C5 fetched its own copy. The badge reads No model meanwhile, and turning the classifier on offers Turn on, not Download and turn on.

The Runtime line says Not running: the model is not on this machine. The Model line above it says why nothing was downloaded (Offline, opted out, a CI runner, or a link whose file cannot be reached); follow that line.

Step quality says Sampling 1 in 4 on this machine: … Normal on a machine where the classifier runs on the CPU, or is marked slow. See On a machine where it runs on the CPU. There is nothing to fix.

Restarting on the CPU: this Mac has no Metal device the classifier can use. Normal on a virtual machine without GPU access. C5 moves it to the CPU by itself.

Paused after repeated failures; the current answers stand alone meanwhile. Three questions in a row failed. C5 rests the classifier for 60 seconds, tries one question, and carries on if it succeeds. If it keeps happening, check the Runtime line and run growther doctor. What C5 does to the classifier itself never pauses it: moving it between GPU and CPU, turning it off, or moving its cache. A crash does count: the questions it was answering fail, and C5 starts a fresh process on its own, whose first question loads the model again.

The Runtime line says the model could not be loaded, for example for lack of memory. C5 tries again by itself: a minute after the failure, then after twice as long each time it fails again, up to 15 minutes, so the classifier comes back once memory frees up. The usual model answers meanwhile. Turning the classifier off and on tries at once. A model this machine or this build cannot run at all is not tried again until you do that.

I cannot choose Assisted or Authoritative. The option says why. Usually one of these:

• It is a setting other than Task categories. Those offer only Off and Shadow.
• The promotion gate has not passed. Its line names each part that has not. Keep answering in Teach the classifier. An administrator on the machine running C5 can raise it anyway, after a warning; from a paired device they cannot. When the part is Telling cases apart, either the classifier's answer does not change when the right answer does, or no answer has yet been the correct one in 10 labelled tasks and not the correct one in 10. Labelling more tasks fixes the second, not the first.
• No calibration is fitted yet. C5 fits one at 23:50 once there are enough labels, or an administrator can press Fit calibration now once it is offered (it says what it still needs until then). A fit that cannot tell answers apart is not stored; the Calibration line says so when that happened.

If you can choose it but the line reads Set to Assisted, but not running: this machine is too slow …, see Slow machines. On Windows this lasts until the model's first timing.

Task categories went back to Shadow after an update. The nightly re-check measures every part of the promotion gate, including Telling cases apart, which v322 and earlier did not measure. If the classifier's answers do not tell cases apart on your install, the kind goes back to Shadow and administrators get the Local classifier: … moved back to shadow notification saying why. Its answers keep being recorded; once the gate passes again, raise it again. A kind that was acting on a calibration from an earlier version that cannot tell its answers apart stops acting too, and is fitted again.

After going back to v319, a Local classifier entry appears under Settings › Extensions. v319 does not know the classifier and treats it as an ordinary extension. Switch Local classifier off there and v319 leaves it alone. See Going straight back to v319.

The switch is greyed out. Hover or tap it. Changing this needs permission to edit settings. means you need that permission. This machine cannot run the local classifier. means the machine is unsupported. A Managed lock means your organisation's policy decides.

There is no Delete button. Deleting is for administrators on the machine running C5, never from a paired device, and not while your organisation's policy keeps the classifier on. While the classifier is on, delete it through the switch: Turn off and delete the model. A link whose file cannot be reached leaves nothing on this machine to delete: remove the link by hand if you no longer want it.

Moving the cache. growther cache relocate refuses while C5 is running (run growther stop first), and while another process, such as growther classifier pull, is downloading the model. Start C5 again after the move. If the classifier is on, it starts from the new location.

Not running: The classifier's model is being deleted. A delete is in progress. It ends in seconds.

Choosing a model from your providers #

There is no Choose a model… button. Choosing and testing are for an administrator at the machine running C5, never from a paired device. The row says why it is not offered to you. If your organisation's policy keeps every model off, or does not allow a model from your providers, choosing is not offered at all; a model already chosen can still be turned off or removed. See Classifier models from your providers.

None of your enabled providers lists a model right now. Turn one on in Settings › Integrations. The list shows only providers turned on in Settings › Integrations, and only the models each one lists. Turn the provider on, start it if it runs on your machine or network, and press Refresh list.

Not listed: … They pass questions to an agent, … The line names the agent gateways turned on in Settings › Integrations — OpenClaw, PicoClaw or Hermes Agent. A gateway hands each question to an agent, which may pass it to any model. The classifier needs one model's own probabilities for each answer. Choose the model the gateway uses from that model's own provider instead.

A provider's group says … cannot answer the classifier. (Anthropic, Groq, Mistral, Cohere, Amazon Bedrock) Their APIs do not return the probabilities the classifier reads, so none of their models can be used. The rest of the line says why for each.

C5 cannot tell that it runs on this machine, and Models allowed from is This machine: … The model's provider cannot show that it runs the model here: LM Studio, vLLM, or a server on another computer. Set Models allowed from to Your network on the Classifier card (an administrator at the machine can), then choose it and confirm where it runs. If the line names your organisation's policy instead, only the policy can change it.

Runs online through P, and Models allowed from is … A cloud model needs Anywhere. Ollama's cloud models cannot be chosen at all: Ollama passes them on to its own servers.

M is not loaded in P. C5 never loads a model itself; load it there and try again. Nothing was changed. Load the model in the provider (in LM Studio or Ollama, say), then choose it again.

The test says the model reasoned before it answered. A model that thinks before it answers cannot be read this way. In LM Studio, turn the model's reasoning off, then test it again. A cloud model that cannot answer without reasoning cannot be used.

… names whichever model the provider points it at; choose a dated or versioned model instead The name moves to a newer model by itself (one ending in latest, or openrouter/auto), so its record would mean nothing. Choose a dated or versioned name.

This machine now requires Terms v1.10, and this page shows v1.9. Reload C5 to read the current text. The page was loaded before C5 was updated. Reload it, read Section 14 again, and confirm.

The chosen model's panel says Not used., Paused. or Changed since it was tested. See When the chosen model is not answering.

Could not turn on M: … Turning the chosen model on checks it again, as choosing it would, and it stays off when a check fails. The message says which; see Turn the chosen model off for a while. From a paired device, or as an account that is not an administrator, it cannot be turned on at all.

There is no Use as fallback switch, or it is greyed out. It is shown only while a model from your providers is chosen and turned on and the built-in model is on. Greyed out, the line under it says why: your organisation's policy sets it, you need permission to edit settings, or turning it on needs an administrator at the machine. See Use as fallback.

Related #

• Memory search — the other models C5 downloads, for semantic search
• Network and egress — the Egress posture, proxies and every host C5 contacts
• Managed configuration — locking classifier_enabled, classifier_external_allowed and classifier_builtin_fallback across a fleet
• Model providers — setting up the providers a classifier model can come from
• Running doctor — every check, and how to read the results
• Environment variables — GROWTHER_CLASSIFIER_NO_AUTO_DOWNLOAD and GROWTHER_CACHE_DIR