Support for the people who protect children online.
A connected workspace for reviewing concerns, assessing conversations, and preserving the record. Designed to take repetitive work off an operator’s shoulders while keeping consequential decisions in human hands.
The existing review interface, populated with sample cases. View full size ↗
Support the operator.Less repetitive searching and assembly.
Keep the context.Review the conversation and its history.
Preserve the record.Connect evidence with a traceable workflow.
Why this work matters
The work is demanding. The tools should help.
Behind every review queue is a person making difficult judgments. Reading concerning conversations, comparing records, and assembling documentation demands sustained attention. Repeating the same manual steps adds to that burden.
Custos brings those tasks into one workspace. Machines can help organize material, surface patterns, and prepare records for review. The operator can then spend more attention on context, uncertainty, and the next responsible decision.
Reducing avoidable workload is a design goal. Effects on fatigue, accuracy, and time saved still need to be measured with operators.
The core workflow
From a concern to an accountable report.
Four connected steps: identify a concern, validate it through human review, prepare a secure evidence package, and report through the appropriate national channel.
01 / CONCERN
Identify what needs attention
Detection brings a conversation or pattern into the operator’s review queue, together with its source context.
Automation assists
02 / VALIDATION
A person checks the evidence
The operator reviews context and supporting indicators, records uncertainty, and decides whether to advance, hold, or dismiss the concern.
Human decision
03 / SECURE PACKAGING
Build a traceable package
Assemble records and metadata, verify file hashes, and seal a manifest with linked custody events. Prepare the report for review.
Integrity checks and review
04 / REPORTING
NCMEC CyberTipline
The approved workflow sends the report and supporting material to the National Center for Missing & Exploited Children and records the returned report ID.
Authorized reporting
Not ready to advance? Hold for more context or dismiss with a recorded reason. A flag alone does not establish wrongdoing.
After reporting: NCMEC reviews tips and makes information available to the appropriate law-enforcement agency for possible investigation. How the CyberTipline works ↗
Custos includes report preparation and submission code; the documented default is a mock environment. Live API reporting requires NCMEC-issued access and a verified production setup. NCMEC is a private nonprofit operating the national CyberTipline, not a federal agency. No affiliation or endorsement is implied.
Inside the working demonstration
Context before conclusions.
The case workspace puts the conversation, assessment, evidence, and notes within reach. Operators can inspect why a concern was raised and return to the source material instead of relying on a score alone.
The assessment approach draws on published grooming research, including the Sexual Grooming Model, the Online Grooming Discourse Model, and O’Connell’s work on online grooming.
Research-informed, with a clear distinction. The Sexual Grooming Model has published content-validation research. That does not independently validate Custos’s AI implementation, accuracy, or risk scores. Those require their own evaluation.
Case workspace · Seeded example
Actual application screenshot. All people, records, scores, and activity shown are demonstration content. View full size ↗
Evidence handling
A record that can be traced and checked.
A useful concern must be backed by a clear record. Custos’s evidence workflow associates files with cases, records handling events, and uses file fingerprints and linked custody records to support integrity checks.
Case-package seals and session manifests help document what was included at a particular point. The purpose is to make review and handoff more accountable, with less manual reconstruction.
These are integrity mechanisms, not a guarantee of legal admissibility. Operational controls, collection authority, and independent validation remain essential.
Evidence workspace · Sample records
The existing evidence view connects individual artifacts with their records. View full size ↗
Supporting the operator
Let automation carry the repetitive work.
Research with digital-forensics analysts describes the emotional toll of repeated exposure to child-abuse material, difficulty switching off, and pressure from backlogs. A separate study of commercial content moderators also found substantial psychological distress. Analyst research · Moderator research
That evidence informs the design goal: reduce avoidable searching, repeated review, and record assembly while preserving careful human judgment.
01 / ORGANIZE
Start with context.
Bring flagged concerns and their records into one queue and case workspace. The aim is less switching between tools and less rebuilding the same picture from scratch.
02 / EXPLAIN
Make the reasoning inspectable.
The assessment code requests research-linked indicators, short supporting excerpts, and a rationale. Operators can check those suggestions against the source conversation instead of accepting an unexplained score.
03 / PREPARE
Automate the paperwork.
Use file checks, manifests, and draft reports to reduce repetitive handling and transcription. Human review remains essential before a concern becomes an external report.
Built on expert research.
Winters, Jeglic, and Kaylor’s Sexual Grooming Model underwent content validation with 18 experts. Lorenzo-Dus, Izura, and Pérez-Tattam’s Online Grooming Discourse Model describes patterns in online interactions. Custos’s assessment rubric adapts these published frameworks, alongside O’Connell’s work, into a structured review aid. SGM study · Online discourse research
These studies establish research foundations, not Custos’s accuracy or a proven reduction in fatigue. The SGM study concerns in-person grooming; its adaptation to online AI assessment needs separate validation. Fewer manual steps and avoidable repeat exposures are intended benefits to test, alongside errors and missed concerns. Software supports, rather than replaces, manageable workloads, breaks, supervision, and access to qualified support.
Where Custos fits
The workflow. Feature by feature.
These offerings serve different roles. Sentinel is a detection library that Custos also uses; Thorn offers detection and review products; Cinder is an investigation platform. The comparison below separates confirmed overlap from features their public materials do not establish.
Related capabilityExpert-informed training; mapping not public
Not publicNo SGM/OGDM mapping found
Case activity and review history
✓ Demo
Outside scopeStandalone library
Not publicDetailed audit history not established
✓ DocumentedDecisions and reasoning logged
File hashes linked to custody events
◇ In code
Outside scopeStandalone library
Related capabilityHash matching documented; custody chain not public
Not publicAudit trail is not proof of a custody hash chain
Sealed case packages and session manifests
◇ In code
Outside scopeStandalone library
Related capabilityPreservation documented; package seals not public
Not publicSeals and manifests not established
NCMEC reporting workflow
◇ In codeMock default
Outside scopeStandalone library
✓ DocumentedEnterprise reporting
✓ DocumentedNCMEC integration
✓ Demo / ◇ In code Custos evidence, not independent validation✓ Documented Described by the providerRelated capability Overlap, with a narrower or different scopeOutside scope Separate system needed beyond the documented libraryNot public Not established in the sources reviewed
*The Thorn column covers separately described Predict and Enterprise offerings; it does not imply all capabilities are included in one product. “Not public” is not a finding that a feature is absent. No vendor access or independent product testing was used.
What the deeper review established
Sentinel: documents detection scores and explanations. Case management, evidence custody, package sealing, and submission need a surrounding application. That scope assessment concerns the published library, not Roblox’s internal safety systems. Repository and documentation
Thorn Safer: Enterprise includes a review interface, reporting, and preservation; Predict includes a grooming label. Hash matching for detection does not, by itself, establish evidence custody or sealed manifests. Current product update · Enterprise architecture
Cinder: documents unified cases, logged decisions, and NCMEC integration. Its public descriptions do not establish the specific research-model mapping or cryptographic custody and manifest mechanisms compared here. Case investigation · Platform documentation
Reviewer protection is not unique to Custos: Thorn describes wellness features, and Cinder describes blur, grayscale, and redaction controls. Feature availability does not itself prove reduced trauma or fatigue. Thorn solutions · Cinder human review
Custos’s focus: connect research-informed assessment, human validation, traceable evidence packaging, and an accountable reporting route around the operator. The opportunity is the fit of that workflow, not a claim that other tools offer no overlapping capabilities.
Research and reporting references
Public sources supporting the research distinction and descriptions above. These references do not imply a partnership or endorsement.
Project presentation reviewed November 11, 2026. Screenshots are from the repository’s existing seeded demonstration and show an earlier interface version. They illustrate the product, not live investigations or measured performance.
Help shape the next stage.
We welcome conversations about workflow needs, evaluation, and responsible project partnerships.
