1. The Problem: Provenance That Does Not Survive

Deepfake mitigation has hardened into a regulated obligation rather than a discretionary best practice. The EU AI Act (Regulation 2024/1689) Article 50(2) requires providers of generative systems to mark synthetic output with machine-readable signals, and Article 50(4) requires deployers of deepfake systems to disclose AI-generated content. In the United States, the FCC has declared AI-generated voices in robocalls unlawful under the TCPA, and the FTC has used its Section 5 unfair-and-deceptive-practices authority against AI-generated impersonation. The intermediary shield of Section 230 is being narrowed by a growing list of state synthetic-media statutes. Each of these instruments presumes the existence of a durable, verifiable answer to a single question: is this media what it claims to be, and where did it come from.

The two dominant technical answers both fail under adversarial pressure. Watermarking and metadata tagging embed an identity signal in the content stream or a sidecar record; watermarks are removable through transcoding, cropping, or generative reconstruction, and metadata records are decoupled from content structure and require persistent external storage. Voluntary provenance manifests from the content-authenticity standards category depend on every link of the creation-and-distribution chain preserving a signed record, an assumption that does not hold against a motivated adversary or across the installed base of capture devices that emit no manifest at all. Statistical classifiers trained on prior generative artifacts age out as generative models advance, leaving any detector structurally behind the next architecture.

2. The Inventive Root: Identity From Variance, Not From Attachment

Content Anchoring assigns each digital artifact a unique identifier (UID) derived deterministically from the artifact's internal variance and structural features rather than from its storage location, file name, cryptographic key, or transmission metadata. The extraction pipeline normalizes the artifact to a canonical scalar field and computes a nine-dimensional multi-axis variance vector organized into three structural axes: an X axis encoding cross-scale energy distribution, a Y axis encoding cross-scale frequency compaction, and a Z axis encoding structural phase persistence based on gradient orientation distribution. The vector is hashed under overlapping quantization scales into a UID that is stable under format conversion, resolution rescaling, and lossy compression within defined thresholds, while diverging predictably as variance-shifting mutations occur.

Two properties of this construction are what make it a deepfake-provenance substrate. First, because the UID encodes a position in a continuous variance space, cosine similarity between two UIDs is directly computable without decoding a fixed binary digest, so a received file can be resolved against a canonical capture within tolerance regardless of the distribution path. Second, the identity is intrinsic to the content: there is no embedded mark to strip and no sidecar to lose. Re-encoding, cropping within bounds, color-space conversion, and container reformatting are the precise transformations the variance vector is designed to survive.

3. Detection Mechanisms Grounded in the Disclosure

The disclosure supplies three structural detection signals, each computed from the content's own variance vector, that together discriminate authentic from synthetic or manipulated media.

Lineage and orphan detection. The provenance validator constructs and traverses multi-root lineage graphs by computing cosine similarity between a candidate UID and previously anchored UIDs and weighting each lineage edge proportional to that similarity. An artifact with no registered parent UID within the configured slope continuity radius is structurally unanchored. The detection signal is therefore not "this content matches a known forgery" but "this content lacks the legitimate lineage an anchor-bearing artifact would carry," a posture that does not depend on having seen the generator before.

Region-level integrity. The quadrant decomposition pipeline computes an independent variance fingerprint for four non-overlapping sub-regions of the canonical image. The comparison framework produces per-quadrant similarity scores: a derivative that modifies only one spatial region exhibits quadrant similarity near 1.0 for unchanged regions and reduced scores for the modified region, surfacing splices, insertions, and synthetically generated regions as a first-pass screen even when global similarity remains high.

Recapture and synthesis signatures. The Z-axis gradient-orientation components carry signatures that elevate under screenshot recapture and re-rendering, and a synthetic-content evaluation compares a candidate variance vector against a slope-band-indexed distribution of generative-model output. A composite risk score combining lineage absence, recapture probability, and synthesis proximity is routed to the pre-release admissibility engine rather than treated as a standalone verdict.

4. Real-Time and Streaming Embodiment

For live video broadcasts, audio streams, and continuous feeds, the UID derivation operates over a sliding window of the stream rather than a discrete file. A window of configurable duration, for example ten seconds of audio or thirty frames of video, is normalized and processed into a window-level UID, and consecutive window UIDs are compared by cosine similarity to detect structural continuity or discontinuity. When the similarity between a window UID and a registered reference UID exceeds a policy-declared threshold, the system emits a real-time match event that may block unauthorized retransmission, generate a consultation record, or invoke the admissibility engine. This is the disclosed basis for live broadcast monitoring and real-time deepfake detection without offline batch processing.

Because the variance extraction is modality-general, the same mechanism applies across the modalities deepfake campaigns exploit. Audio is normalized to a mel-spectrogram scalar field; video is handled at both the frame level and the clip level through a temporal delta vector that encodes scene transitions and motion intensity; text is mapped to a token-frequency scalar field. All paths converge on the shared variance extraction stage, so the AI-voice robocall, the face-swap clip, and the synthetic press release are anchored by one substrate rather than three single-purpose detectors.

5. Deployment Options and Embodiments

The application admits several enabling deployment configurations, none of which requires the source to have registered content in advance:

  • Passive ingestion layer. A platform or newsroom anchors every received file at intake, building an internal anchor library and a record of how anchor-based resolution aligns with verification decisions actually made, before any enforcement is bound to it.
  • Pre-publication or pre-distribution gate. The composite risk signal feeds the pre-release admissibility engine so that structurally unanchored or synthesis-proximate content is held for review at the commitment boundary rather than filtered after release.
  • Reference-corpus matching. Verified public figures, official broadcasts, and credentialed sources publish reference UIDs; incoming media is cross-resolved against them so that an impersonation of an anchored identity surfaces as a lineage mismatch.
  • Federated and edge operation. Because slope-band routing resolves a UID from its variance value without a central directory, anchors run across centralized data centers, federated institutional clusters, peer networks, and mobile edge devices, allowing capture-point anchoring on the device that recorded the original.
  • Interoperation as a credentialed observation. Where a content-authenticity manifest or preserved metadata is present, it is admitted as one authority observation within a composite verification rather than as an authoritative singleton, so the structural anchor strengthens rather than replaces existing provenance signals.

A skilled implementer can build each of these from the disclosed extraction pipeline, slope-band anchor network, lineage graph construction, and pre-release admissibility engine.

6. Compliance Mapping

The mapping to obligation is direct. EU AI Act Article 50 disclosure determinations are supported by the region-level integrity analysis and the synthesis-proximity signal, which produce a structural finding a deployer can act on. FCC AI-voice enforcement is supported by the streaming audio embodiment, which resolves a live or recorded voice stream against reference anchors in real time. FTC Section 5 impersonation matters are supported by reference-corpus matching that surfaces an unanchored impersonation of an anchored identity, together with the lineage record that documents the verification reasoning contemporaneously. Because the verification operates on the media itself rather than on the infrastructure that transmitted it, the evidentiary record is produced continuously as a by-product of normal workflow rather than assembled retrospectively.

7. Disclosure Scope

The disclosure encompasses the deepfake-provenance architecture in which a candidate artifact's multi-axis variance vector drives lineage and orphan detection that classifies an artifact without a registered parent UID within the slope continuity radius as structurally unanchored, quadrant-level integrity comparison that surfaces region-level substitution and synthetic generation, and Z-axis recapture and slope-band-indexed synthesis signatures, with a composite risk score routed to the pre-release admissibility engine. It extends to the streaming embodiment in which these signals operate over sliding-window UIDs for real-time deepfake detection across image, audio, video, and text modalities.

The disclosure does not constrain the specific policy-calibrated thresholds, the construction of the synthetic-content distribution, or the content categories to which the detectors are applied; these are deployment choices. The disclosure does constrain the structural properties: the signal must be derived from the content's own variance vector rather than from an embedded or attached mark, the orphan classification must be treated as a provenance status rather than a binary verdict, and the composite signal must route to pre-release admissibility evaluation rather than post-release filtering.

The disclosure is filed under PCT International Application No. PCT/US26/28630 and forms a structural primitive of the content anchoring system. A system that embeds a severable watermark, that evaluates authenticity only after release, or that depends on a generation-specific classifier trained against a fixed set of known forgeries is outside this disclosure.