Moderation

Four packages judge content: curator filters datasets, darkroom guards generation, proof scores trained LoRAs, and cortex holds the clients they share. They all lean on the same few external services. This page explains that stack as one system and links to each package's own docs for the flags and fields. The package docs are the source of truth. This page is the map.

What judges what

Judge	What it answers	Setup
NudeNet	Which body parts are exposed in an image, and how confident it is about each one.	NudeNet server
SingGuard	Whether an image or a prompt breaks a written policy. It returns a verdict and the rule that fired.	SingGuard guard server
EraX-NSFW	A second nudity detector. Curator uses it to fill in when NudeNet finds no exposed parts.	EraX-NSFW
MiVOLO	An age estimate per face and body. It feeds the under-18 rules.	MiVOLO v2

The exposed score

The suite defines nudity once, as cortex's DEFAULT_EXPOSED_LABELS: BUTTOCKS_EXPOSED, FEMALE_BREAST_EXPOSED, FEMALE_GENITALIA_EXPOSED, MALE_GENITALIA_EXPOSED and ANUS_EXPOSED. An image's exposed score is the highest NudeNet score among those five labels, or 0 when none is found. Detections under 0.2 are dropped, and a NaN or infinite score is rejected rather than read as safe. Every package that talks about "exposed" means this number, from argus_cortex.moderation (opens in new tab).

SingGuard verdicts

A SingGuard check sends the image (or just the text) with a policy: a Markdown list of rules, each a ### heading with a letter. The server answers safe, the verdict, the rule that fired, and an explanation. The policy text lives in the clients, in cortex's policies.py. The server holds only SingGuard's defaults. See the request and response fields.

EraX fills gaps

Curator runs EraX (opens in new tab) beside NudeNet on every image. It maps EraX's hits onto NudeNet's labels only when NudeNet's exposed score is 0 or the NudeNet check failed. A filled result says so (erax_filled: true).

The age estimate feeds the under-18 rules

MiVOLO estimates an age for each face and body it finds. When anyone estimated under 18 shares a frame with an exposed score of 0.2 or more, the image is minor-exposed. This is a hard exclusion in curator and a deletion in darkroom, and no request or threshold can override it.

Both packages apply the same rule, chosen by one setting (ARGUS_CURATOR_MINOR_EXPOSED, ARGUS_DARKROOM_MINOR_EXPOSED):

Mode	When it fires
`frame` (default)	Any age under 18 anywhere in the frame, plus exposure anywhere in the frame.
`overlap`	Only when an exposed region overlaps a young person's box, or overlaps nobody's box. It still fails closed: a young face whose body wasn't found, or an age with no box to place it, flags any exposure.

overlap excludes fewer images (an adult's exposure beside a fully clothed child doesn't fire), and it relies on the person detector finding every body. frame doesn't. See the under-18 rule (opens in new tab) in curator and the age check (opens in new tab) in darkroom. Darkroom gets the estimate from curator's POST /faces (ARGUS_DARKROOM_AGE_URL) or from its own [age] extra.

The age is never sent to SingGuard as a verdict input.

Where each is used

curator

Dataset filtering. Every exclusion is recorded with its reason, so a selection can be audited afterwards.

Scan filters: NudeNet (--nsfw-threshold, --nsfw-label) and SingGuard (--guard-unsafe, --guard-rules), each an opt-in exclusion. See nudity (opens in new tab) and standalone moderation (opens in new tab).
Policy check: a SingGuard verdict per image under a named policy, with optional context (age, folder, caption, scene). See policy check (opens in new tab).
Prefilter: cheap rules that drop images before the slow checks run. See prefilter (opens in new tab).
drop_nsfw (--drop-nsfw): a prefilter rule that drops an image with an exposed score of 0.2 or more. It is refused when no NudeNet is configured.

Env: ARGUS_CURATOR_NUDENET_URL, ARGUS_CURATOR_SINGGUARD_URL, ARGUS_CURATOR_ERAX_URL, ARGUS_CURATOR_MIVOLO_MODEL.

darkroom

Guarded generation. Nothing is written to disk before the verdict. See how a generation runs (opens in new tab).

Pre-flight: a text-only SingGuard check of the prompt and negative prompt, under every selected policy, before anything renders.
Per-policy image scans: each image is judged by SingGuard once per policy, on its own and with its prompt. NudeNet runs once, and MiVOLO runs when it is available.
minor-exposed deletion: as above. The record is marked critical.

Env: ARGUS_DARKROOM_SINGGUARD_URL (required), ARGUS_DARKROOM_NUDENET_URL, ARGUS_DARKROOM_AGE_URL, ARGUS_DARKROOM_THINKING.

cortex

argus_cortex.moderation (opens in new tab) holds the shared clients: NudeNetClient and NsfwResult, SingGuardClient and GuardResult, DEFAULT_EXPOSED_LABELS, THINKING_MODES, and the policy texts in policies.py. Cortex reads no environment: each package reads its own env and passes the URL in.

proof

SafetyScorer reports safety = 1 − (the highest unsafe probability) for each evaluation image, using NudeNet through cortex's client (PROOF_NUDENET_URL). Without a URL the scorer is skipped. Proof uses no SingGuard. See argus-proof (opens in new tab).

darkroom's adult policies

Darkroom runs SFW by default. Setting ARGUS_DARKROOM_ALLOW_NSFW=1 (or --allow-nsfw) lets a request name an NSFW enforcing policy. These are the adult-* family: adult, adult-granular, adult-cautious and adult-cinematic, from cortex's policies.py. With NSFW allowed, the server default is adult-granular. The variants are still being evaluated, and only one will remain at release. Until then, Studio offers them only in dev mode.

Rule letters are not the same across policies, so darkroom reads rule names from the data, never from a fixed list. adult-cinematic drops the crimes heading at B, so its letters B–G shift, and it treats depicted crime and violence as safe unless it is sexual, gory or instructional. See policies (opens in new tab).

What an NSFW enforcing policy changes: two things, and nothing else (mode (opens in new tab)):

At pre-flight, a rule that is not illegal warns instead of refusing.
An exposed score at or above the threshold blurs the image instead of omitting it.

Without ALLOW_NSFW, a request that names an NSFW enforcing policy gets HTTP 400, and a server configured with one refuses to start.

What stays illegal under every policy: each policy marks some rules illegal. A for every adult policy, and A, H and I for adult-granular. Rule A covers minors. An illegal rule firing under any policy is refused at pre-flight and deleted on output, whatever the mode. minor-exposed deletes in both modes too. See illegal rules (opens in new tab).

Why a deleted image is never shown: the image is judged in memory before anything is written. A deleted image's buffer is dropped, so the record keeps only its sha256 and pHash, never pixels, a thumbnail or a preview, because any of those would be the image again. Studio's Darkroom and Safelight follow the same rule: no image for a deleted, omitted, refused or failed record. See record (opens in new tab).

Thinking modes

SingGuard can spend more or less effort on a verdict. The client chooses with a thinking field (ARGUS_CURATOR_SINGGUARD_THINKING, ARGUS_DARKROOM_THINKING):

Mode	What it does	Cost
`fast`	Verdict and rule only.	About 1 s per image.
`fast-slow`	SingGuard's own reasoning mode: a first-line verdict, then the reasoning.	Several seconds per image.
`slow`	Reasoning first, then the verdict, so the reasoning isn't anchored to a snap judgement.	Several seconds per image.
`cascade`	`fast` on everything, then `slow` on whatever `fast` flags (or can't parse). The slow verdict stands.	One fast pass for clean content, fast plus slow for flagged content.

Curator and darkroom both default to cascade. Curator sends a frame straight to slow when its exposed score is already 0.2 or more.

Which servers support which modes:

Server	Modes	Checks
agent-composer's guard	all four	images and text-only
argus-cull's `singguard` service	all four	images only (a text-only check gets 422)
any other SingGuard deployment	`fast` and `fast-slow`, the model's own two	depends on the server

fast and fast-slow are SingGuard's own template modes. slow and cascade are added by the two Argus servers. Darkroom's pre-flight is a text-only check, so darkroom needs agent-composer's guard. When a server answers HTTP 400 to slow or cascade, curator switches to fast for the rest of the run and logs singguard_mode_unsupported.

The limits

Age estimates are automated guesses. MiVOLO can be wrong in both directions. The under-18 rule is a safety net, not proof of age. An age check that can't run never deletes on its own, but it never clears an image either.
A failed check fails closed. In curator, an image an enabled filter couldn't check is excluded (nsfw-unchecked, guard-unchecked, face-attributes-missing). In darkroom, a missing SingGuard URL stops startup, an unchecked enforcing pre-flight refuses, an unchecked enforcing scan deletes, and in SFW a NudeNet outage omits the image. Cortex raises CheckFailed rather than returning a pass. Proof is the exception: it is a scorer, not a gate, so an image it couldn't score is left out of the safety metric.
No real sexual content is in any repo, fixture or demo. Tests drive the NSFW and illegal paths with fake NudeNet and SingGuard servers that return scripted verdicts. Demo records are hand-written, and deleted records carry no image.