01
Content risk
Safety concerns can appear for only a few seconds. Analysis is being designed to locate the relevant moment while preserving whether a video promotes, documents, condemns, or responds to it.
Swivver Signal · Human-impact intelligence
The human-impact layer for short-form video.
Short video can teach, move, entertain, and connect at extraordinary speed. It can also scale deception, unsafe material, dishonest advertising, and design patterns that compete aggressively for attention.
Swivver Signal is being built to help platforms, publishers, advertisers, creators, AI systems, and researchers evaluate those risks before reach turns them into scale.
Pre-release. This page describes intended scope under evaluation; general production access and performance guarantees are not available.
Why Signal
Attention should be earned, not extracted.
The problem is not brevity. It is scale without inspection. Signal is being built so media systems can ask whether a video is safe, authentic, honest, useful, accessible, and proportionate in how it captures attention—not only whether it keeps someone watching.
02
Integrity risk
Synthetic or altered media can be legitimate. The useful questions are what changed, whether it was disclosed and authorized, whom it represents, and whether it is used deceptively.
03
Feed and session risk
A clip can have a different impact inside autoplay, repeated exposure, and personalized sequencing. Signal separates observable design patterns from separately governed session context; it does not diagnose addiction.
The Signal Profile
One score can rank. A profile can explain.
A universal ‘good video’ score would hide too much. Signal’s intended primary output is a multidimensional profile that keeps distinct risks, useful qualities, evidence, and uncertainty visible.
Core dimensions
01Safety and age suitability
02Identity and impersonation
03Synthetic-media disclosure
04Advertising integrity
05Information quality
06Originality and rights
07Accessibility
08Technical quality
09Attention design
1
Confidence and unknowns
Results are designed to show when evidence is strong, weak, conflicting, unavailable, or outside tested scope. Unknown is a valid state.
2
Timed evidence
Signals are intended to point to the video, audio, claim, or provenance segment that contributed to a result.
3
Decision context
Audience, language, policy, threshold, review state, and model version stay attached so a result can be interpreted and challenged.
Assessment scope
What Swivver Signal is being built to assess
Each capability is designed to return useful evidence while keeping the limits of that evidence visible. The intended output is context for proportionate decisions—not an automatic verdict.
Scroll horizontally to compare all three columns.
| Capability | What Signal is designed to return | Boundary kept visible |
|---|---|---|
| 01Safety and age suitability | Segment-level signals for explicit sexual material, graphic content, violence, exploitation, illegal material, and other age-sensitive content. Results can support warnings, age gates, review, distribution limits, or removal under an explicit policy. | Context matters. News reporting, medical education, documentary evidence, and artistic work may depict harmful material without promoting it. A model result is not automatically a legal determination. |
| 02Harassment, hateful conduct, self-harm, and dangerous challenges | Signals for targeted abuse, threats, hateful conduct, encouragement of self-harm, dangerous instructions, coercion, and harmful challenges. Analysis should consider who is targeted and whether the video promotes, condemns, documents, or responds to the behavior. | Sarcasm, counterspeech, reclaimed language, crisis support, satire, and quoted material can change meaning. Ambiguous or severe cases require contextual review. |
| 03Synthetic media and AI disclosure | Checks for available provenance credentials, metadata, watermarks, visual or audio forensic signals, lip-sync inconsistencies, synthetic speech indicators, and visible disclosure. The result can support labeling, provenance warnings, or further review. | No single detector proves that media is AI-generated. Missing credentials mean provenance is unknown—not that the media is fake. Valid provenance also does not prove that the message itself is true. |
| 04Impersonation and deceptive identity use | Comparison of claimed identity with available face, voice, account, provenance, and enrolled reference information. The system can support identity-owner alerts, labeling, review, and escalation. | Identity matching is strongest when a person or organization has lawfully enrolled reference material. Parody, dubbing, authorized avatars, lookalikes, reenactments, and fan works require context. Biometric data requires strict purpose and privacy controls. |
| 05Advertising integrity and scams | Detection of promotional intent, missing or unclear sponsorship disclosure, artificial urgency, suspicious scarcity, price or discount claims, product-result claims, fabricated endorsements, landing-page mismatches, and recognizable scam patterns. | An unsupported or unverifiable claim is not automatically an intentional lie. Intent should not be inferred from a mismatch alone. Medical, financial, political, and other regulated claims should route to specialist review. |
| 06Factual and information integrity | Extraction of objective claims, entities, quantities, dates, sources, and evidence relationships. Results can distinguish supported, contradicted, developing, materially incomplete, outdated, and currently unverifiable claims. | Signal cannot settle every disputed, emerging, or specialist question. Its result depends on evidence access, source quality, freshness, language coverage, and domain expertise. |
| 07Editorial and informational value | Genre-aware assessment of coherence, title-to-content alignment, completeness, source density, useful contribution, explanatory value, and relevance to the intended audience. | Editorial value is not universal taste. Comedy, art, news, education, commentary, and product demonstrations need different criteria. These signals should primarily support ranking and creator feedback—not removal. |
| 08Originality, provenance, rights, and reuse | Perceptual matching for duplicates and reuploads, near-duplicate detection, transformation analysis, provenance checks, attribution signals, and available rights information. | Originality is limited by the reference corpus. Remix, quotation, licensed reuse, trends, parody, and fair-use questions require legal and contextual analysis. Metadata alone cannot prove ownership. |
| 09Technical and accessibility quality | Signals for audio intelligibility, caption accuracy and synchronization, unreadable text, blur, compression, clipping, excessive loudness, broken framing, aspect-ratio problems, visual flashing, and other accessibility concerns. | Production polish is not the same as human value. Signal should not systematically disadvantage low-budget reporting, minority production styles, disability-related speech, or unconventional aesthetics. |
| 10Attention-capture and compulsive-design patterns | Measurement of cut rate, saliency turnover, caption replacement rate, seamless looping, repeated engagement prompts, artificial urgency, withheld payoff, continuous novelty, intense audio transitions, visual pulsation, and related design features. | These are design-risk proxies—not proof of ‘hypnosis,’ addiction, psychological harm, or creator intent. Repetition, fast editing, ASMR, music, gaming, and experimental art are not automatically harmful. |
| 11Feed and session well-being | Where privacy-governed telemetry is available, Signal can combine clip-level indicators with repetition, autoplay, session duration, rapid reopening, time of use, feed diversity, and cumulative exposure. This can support stopping points, diversity controls, repetition limits, notification changes, and user-set goals. | Session risk cannot be determined from a video alone. It requires additional context, data minimization, clear controls, and separate governance. It should never become a covert psychological profile of an individual. |
NIST synthetic-content guidanceExplains why provenance, labeling, detection, testing, and oversight work best together.C2PA Content CredentialsRecords media origin and editing history; it is provenance, not a truth certificate.WCAG 2.2 flashing guidanceDefines measurable thresholds for flashes and one category of visual safety.
Inside Swivver
Signal extends the Swivve Object. It does not replace it.
Article and source context remain the foundation. Signal adds inspectable media- and distribution-level context, while the Trust Ledger preserves how profiles, reviews, and policy actions change over time.
- 01
Short video + context
A Swivve keeps the clip connected to its written article, sources, claims, provenance, and rights.
- 02
Signal Profile
The profile adds inspectable media- and attention-design signals, confidence, and timed evidence.
- 03
Trust Ledger
Analysis, policy versions, review actions, appeals, and reversals can remain visible as state changes.
- 04
Proportional decision
An explicit policy may recommend or apply feedback, context, a label, resequencing, review, or another bounded action.
This connection matters because a profile is not static: evidence, confidence, policy, review, appeal, and reversal state may change.
Explore the Trust LedgerBeyond Swivver
Integrity risks do not stop at one platform.
Signal is being built as a service for partner-supplied video because shared ecosystem risks require testing and interoperable evidence beyond Swivver’s own media objects.
Shared risks cross boundaries
The same clip, claim, identity misuse, or commercial deception can move between publishers, platforms, ad systems, creator tools, and AI products.
External evaluation improves the work
Partner testing can expose weak assumptions across languages, genres, audiences, and policies while improving interoperability and review design.
Profiles travel with the media
Partner-supplied media can be evaluated inside Swivver, where machine-readable Signal Profiles, evidence, policy actions, and review state stay connected.
Swivver partner workflow
Partner-supplied video
Swivver AI + policy automation
Evidence + Signal Profile + Trust Ledger
Signal partners
Building the integrity layer with a focused network.
Signal connects Swivver with partners across aligned AI, open infrastructure, evaluation, and policy-aware automation.
Become a Signal partner
We welcome partners prepared to test difficult cases, contribute governed evaluation material, and measure real outcomes.
Planned system architecture
Understand the whole media object, not just a sampled frame.
Signal is designed as a sequence of inspectable steps rather than one general-purpose model acting as a universal judge.
- 01
Preserve origin and context
Begin with original media, available metadata, provenance credentials, creator declarations, account context, and applicable rights information.
- 02
Read the whole media object
Consider video, audio, language, captions, on-screen text, claims, speakers, editing patterns, and timing together—not just sampled frames.
- 03
Run specialist analyses
Use separate capability paths for safety, identity, synthetic media, advertising, information, originality, accessibility, and attention design.
- 04
Fuse evidence through time
Keep segment-level evidence and apply a versioned policy for the audience, language, region, content category, and risk tolerance.
- 05
Route uncertainty responsibly
Send ambiguous, severe, specialist, identity-related, appealed, or otherwise high-impact cases to accountable human review.
Proportional intervention
A signal is not a sentence.
The same evidence may justify different actions for a child audience, an adult archive, a satire feed, a marketplace, or a newsroom. Signal is intended to support explicit policy—not replace accountable judgment.
A proportional action ladder
- 01Allow
- 02Improve
- 03Inform or label
- 04Age-gate or resequence
- 05Human review
- 06Remove under explicit policy
Low production value, unconventional style, minority expression, or model uncertainty should not be sufficient grounds for removal.
No universal oracle
One model cannot objectively define cultural, artistic, journalistic, or human value.
Unknown is valid
Missing or conflicting evidence should produce abstention—not manufactured certainty.
No diagnosis
Attention signals are observable design proxies, not a diagnosis of addiction, illness, intent, or self-control.
Escalation for high-impact decisions
Automation can apply configured actions; severe enforcement, appeals, and specialist judgments route to accountable review.
Narrow biometric use
Face or voice references should be lawful, consented, purpose-specific, access-controlled, and deletable.
Founding cohort
We are inviting the people who can make this credible.
We are inviting a focused cohort to challenge the taxonomy, evaluation methods, privacy architecture, review workflows, integration model, and release criteria.
Who should join
Product and platform partners
Short-video platforms, publishers, creator tools, marketplaces, ad systems, and media infrastructure teams.
Research and evaluation partners
Behavioral, HCI, media-forensics, accessibility, digital-well-being, and recommendation-system researchers.
Safety, policy, and identity partners
Trust-and-safety teams, privacy and child-safety experts, civil society, creators, rights holders, and affected communities.
What collaboration may include
- Policy and taxonomy mapping
- Review of representative and difficult examples
- Error analysis across languages, genres, audiences, and policy categories
- Signal Profile and automation schema review
- Human-review, privacy, and appeals workflow design
- A limited pilot proposal when evaluation criteria are met
A useful introduction includes a defined use case, lawful access to evaluation material, willingness to measure errors and reversals, and an owner for policy or governance decisions.
Request collaborationOpens your email client. Requests are reviewed for product fit, data rights, privacy, safety, and research value.
Signal FAQ
Questions, answered with the boundaries intact.
Founding collaboration
Better video should win for better reasons.
Help test difficult cases, measure real outcomes, challenge assumptions, and shape more accountable infrastructure for short-form video.