Content Anchoring: Computable Identity for Media That Changes

1. Problem and Premise: Why Byte Equality Is the Wrong Question

The dominant model for content identity is byte equality. Two artifacts are the same when their byte sequences hash to the same digest; otherwise, they are different. This formulation served a generation of systems in which content was authored, stored, and retrieved without intermediate transformation. It does not survive the modern media lifecycle, in which a single image traverses a recompression pipeline at upload, a thumbnail generator, a content-delivery transcoder, a screenshot reposted on another platform, and a generative editing pass before anyone attempts to evaluate its provenance. At each step the bytes change, and at each step every byte-equality identifier becomes useless.

The consequences are operational and adversarial. Newsrooms cannot reliably correlate altered versions of source imagery. Platforms cannot cluster near-duplicate posts into a single provenance event. Investigators lose attribution when a video is re-encoded for distribution. AI developers cannot audit dataset reuse or contamination because preprocessing pipelines mutate every artifact before storage. Watermarking schemes try to repair this by embedding a synchronizing signal into the content, but watermarks degrade under cropping, heavy compression, and adversarial removal, and they require near-universal adoption to function as a provenance layer. Perceptual hashing schemes try to repair it by computing similarity-tolerant fingerprints, but most are domain-specific, brittle to geometric edits, and lack a governance model for resolving disputes.

The premise of content anchoring is that the right question is not "are these bytes equal" but "do these artifacts share the same underlying structure within expected mutation envelopes, and if so, what governed lineage connects them." Identity is no longer a label attached to bytes. It is a computed property of the entropy structure of the artifact itself, evaluated against drift envelopes for the operations the artifact has plausibly undergone. Where lineage exists, it is recorded positively as a governed sequence of structural transitions; where lineage is absent for an artifact whose origin is claimed, the absence itself constitutes evidence that the claim is false.

2. Core Primitive: Structural Identity from Entropy

The central construct is the anchor-bound content identifier. The identifier is computed from intrinsic structural measurements of the artifact, normalized to a canonical representation, and quantized through a stability-preserving transformation that maps bounded mutation to the same identifier while diverging on meaningful change.

For raster imagery, the structural measurements include multi-scale local entropy distributions (variance, gradient magnitude, and second-order statistics evaluated at grid sizes typically ranging from 4x4 to 64x64), coarse orientation histograms over the dominant gradient field (8 to 32 angular bins), inter-scale ratios that capture how detail compacts or spreads under scale change, edge-density to global-variance ratios, and color-distribution moments computed in a perceptually uniform space. For audio, the measurements include log-mel band energy ratios over short and long windows, spectral flux trajectories, harmonicity envelopes, and tempo-coupled rhythm structure. For video, the measurements compose per-frame image structure with inter-frame motion-magnitude statistics and shot-boundary structure. For text, the measurements include token-distribution entropies at multiple n-gram scales, sentence-length distributions, and structural punctuation rhythms.

Each measurement family is selected for two properties. First, stability under benign transformation: the measurement should produce the same value, within bounded variation, when the artifact is recompressed, resized, transcoded, or re-encoded within a documented mutation envelope. Second, sensitivity to meaningful change: the measurement should diverge when the artifact is cropped beyond a defined extent, composited with unrelated content, or generatively transformed.

The measurements are combined and quantized into the identifier through a stability-preserving quantization that rounds bounded variation to canonical bins. The output is not a unique fingerprint of every possible artifact in isolation; it is a structural address that locates the artifact within a neighborhood of the global identity space. Two artifacts that share an anchor are, by construction, neighbors in entropy structure under the system's drift model. Two artifacts that do not share an anchor have either diverged structurally or never were related.

3. Mechanism: Lineage Chains and Derivative-Class Detection

A single anchor identifies one structural neighborhood. A sequence of anchors, connected by recorded transformations, identifies a lineage. The lineage chain is the durable record of how an artifact has evolved.

A lineage entry binds three things: a prior anchor, a successor anchor, and a transformation class with its parameter envelope. Transformation classes include resize (with scale-factor envelope), recompression (with codec and quality envelope), format conversion (with codec pair), color regrade (with bounded chromatic shift), crop (with bounded fractional retention), and composite (with bounded source-region count). Each class has a documented effect on the structural measurements and a corresponding admissibility predicate: a candidate successor anchor is admissible as a derivative of the prior anchor under the class if and only if the measured structural change lies within the class's envelope.

Derivative-class detection is the operation that asks, given two anchors and no prior lineage record, whether one is a plausible derivative of the other and under which class. The system computes the structural delta and tests it against each class's envelope. A successful match yields a candidate lineage entry; multiple matches yield a candidate set ordered by likelihood. Failure to match any class indicates the two artifacts are not in a derivative relationship, which may be because they are unrelated, because the transformation lay outside any documented class (heavy cropping, generative substitution), or because an attempt was made to break structural continuity deliberately.

The lineage chain is therefore not a passive log. It is a structurally validated record in which each link is admissible only under a documented transformation class, and each link is independently re-verifiable from the anchors alone. An attacker cannot fabricate a lineage entry without producing two artifacts whose structural relationship matches the claimed class envelope, which requires actually performing the corresponding transformation on the actual prior artifact. The lineage is therefore self-witnessing: it cannot be forged without performing the work it claims to record.

4. Mechanism: No-Watermark, No-Metadata-Required Operation

Content anchoring operates on the artifact itself, without modification and without dependence on accompanying metadata. This property is structural rather than incidental.

Because the anchor is computed from intrinsic structure, it is recoverable from any copy of the artifact regardless of carrier, transport, or container. An image extracted from a screenshot, a re-encoded video downloaded from a third-party mirror, an audio clip captured by ambient recording all yield the same anchor (within drift envelopes) as the originals from which they are derived. There is no signal to strip and no field to remove, because the identity is the structure.

Metadata, when present, can be cross-validated against the structural anchor, but it is not required for resolution. A C2PA manifest or an EXIF block claiming a particular origin can be tested against the artifact's structural anchor and lineage chain: if the claim is consistent with the structural record, it is corroborated; if it is inconsistent, it is at minimum suspect. Metadata thus becomes an optional witness rather than a load-bearing carrier of identity.

Watermarks, similarly, are not used and are not required. Watermarking schemes fail under heavy compression, cropping, format conversion, and adversarial removal, all of which leave the structural anchor intact. The architecture deliberately avoids any mechanism that requires modification of the artifact, because such mechanisms create a class of attacks (strip the watermark, claim no provenance) that the present architecture does not admit (there is no watermark to strip).

The absence-as-evidence property follows. When an artifact is presented and its structural anchor does not appear in any governed lineage chain, the system can report this fact positively. For high-stakes content domains in which authoritative sources commit lineage to anchored registries at the moment of authoring, the absence of any lineage for an artifact claimed to be from such a source is itself probative evidence that the artifact is not what it claims to be. This converts forgery detection from a problem of detecting added artifacts (watermarks, signatures) to a problem of detecting missing positive lineage, which adversaries cannot manufacture without producing the corresponding governed structural transitions.

5. Mechanism: Adaptive Index Resolution Without Global Registry

Identity is only useful when it can be resolved at scale, and resolution must not require a single global registry that all participants consult. Content anchoring resolves identity through an adaptive index whose structure mirrors the entropy space of the anchors themselves.

The index is organized hierarchically by structural neighborhood. Coarse partitions correspond to broad regions of the structural space (orientation-dominant content, high-entropy textural content, low-entropy graphical content); finer partitions correspond to progressively narrower neighborhoods. A presented anchor is resolved by descending the index along the partition path that admits it, terminating at the leaf neighborhood whose drift envelope contains the anchor. Lineage candidates within that neighborhood are returned for further evaluation.

Critically, the index is segmentable. Different governance domains can maintain index partitions for the structural neighborhoods they govern, exchanging anchors and lineage entries through bridges that operate at neighborhood granularity rather than at full-content granularity. A news organization can govern lineage for its own published imagery without participating in a global content registry; a platform can govern lineage for content uploaded through its pipelines without owning the global structural space. Cross-domain resolution, when needed, proceeds through anchored neighborhood bridges that carry only the structural address and the governed lineage, not the underlying content.

The index supports three resolution modes. Direct resolution returns the lineage for an exact anchor match. Neighborhood resolution returns candidate lineages for anchors within bounded structural distance. Lineage-traversal resolution walks an anchor backward through derivative-class edges to identify a presumptive origin. None of these modes requires a global table of all content; all of them operate on locally governed partitions of the structural space.

6. Operating Parameters

Anchor sizes range from 64 bits for low-assurance neighborhood addressing to 2048 bits for high-precision lineage anchoring, with 256 to 512 bits being typical for general media. Quantization grain is selected per measurement family to balance stability against discrimination; typical grains correspond to 1 to 5 percent variation tolerance per scalar measurement, with multivariate envelopes calibrated against measured drift on representative transformation pipelines.

Stability targets under common transformations are configurable but typically include: invariance to JPEG recompression at quality 50 and above, invariance to resize within a factor of 2 in each dimension, invariance to format conversion within standard codec families, invariance to color regrade within bounded chromatic shifts, and bounded structural drift under cropping up to 20 percent of frame area. Beyond these envelopes the anchor is permitted to diverge, and a candidate lineage entry must be supplied to maintain continuity.

Index neighborhood sizes are tunable per governance domain. A typical configuration maintains leaf neighborhoods of 10^3 to 10^5 anchors, with coarser partitions aggregating up to 10^9 anchors per domain. Resolution latency targets fall between 1 and 50 milliseconds for direct and neighborhood resolution under nominal load; lineage-traversal resolution is bounded by the depth of the lineage chain and typically completes within 100 milliseconds for chains of fewer than 32 hops.

Transformation-class envelopes are documented per class and are calibrated against representative pipelines. Recompression envelopes are typically expressed as bounded percent-change in spectral energy ratios; resize envelopes as bounded inter-scale measurement deltas; crop envelopes as bounded fractional retention with corresponding measurement contraction; composite envelopes as bounded source-region count with per-region structural admissibility.

7. Alternative Embodiments

The architecture admits multiple embodiments differentiated by media class, governance topology, and integration depth.

In an authoring-side embodiment, anchors and lineage entries are committed at the moment of content creation by a governed authoring tool. The tool computes the anchor for the canonical artifact and records it to a domain-bound index along with its origin attestation. Subsequent legitimate edits committed through the same or federated tools extend the lineage chain. Artifacts presented later can be resolved against this commitment record, and absence of a record is interpreted under the absence-as-evidence property.

In a platform-pipeline embodiment, anchors are computed at ingestion and at every transformation step within a content-handling pipeline. The pipeline maintains an internal lineage chain that records the transformation each artifact has undergone. Cross-platform federation occurs through bridges between platform-bound indices.

In a forensic embodiment, anchors are computed against artifacts of unknown origin and resolved against any accessible governance domain's index. Direct hits identify the artifact's lineage; neighborhood hits identify candidate originals; lineage-traversal walks back through derivative-class edges to identify presumptive sources. The forensic embodiment does not require the original authoring system to have used the architecture; it only requires that the artifact retain enough structure to anchor.

In an AI-dataset-governance embodiment, anchors are computed for every training artifact and recorded to a dataset-bound index. Preprocessing transformations are recorded as lineage entries. Subsequent model outputs can be tested against the dataset index to identify training-set membership and reuse, supporting licensing audit, contamination detection, and dataset-version tracking without requiring per-artifact metadata propagation.

Each embodiment uses the same underlying primitive; they differ in where the anchor computation occurs, where lineage entries are committed, and how the index is partitioned.

8. Composition with Adjacent Primitives

Content anchoring composes with keyless device pseudonymity and with continuity-based biological identity to produce capabilities none of the three can produce alone.

Composition with keyless device pseudonymity binds an anchor to the device that produced or processed it without exposing a transferable cross-domain identifier. The lineage entry can record the device pseudonym under which a transformation was performed, allowing later resolution to distinguish, for example, transformations performed by a governed pipeline from transformations performed by an arbitrary downstream party. The pseudonym is non-transferable across domains; the anchor is universally computable; the lineage entry binds them at the moment of transformation.

Composition with continuity-based biological identity binds an anchor to the validated continuous human under whose engagement the artifact was authored. The lineage entry records the continuity-witness assurance level at the moment of authoring. For artifacts whose authorship is later disputed, the anchor's lineage chain provides a structurally validated record of which continuous human, under what assurance posture, produced or transformed the artifact. The biological witness is non-transferable; the anchor is forgery-resistant; the composition is verifiable authorship without watermarks, signatures, or transferable credentials.

Composition with governance bridges allows anchored lineages to be exchanged across domains under policy, with each domain governing the structural neighborhoods relevant to its operations and federation occurring only at neighborhood boundaries. This composition supports global-scale provenance without global consensus and without any single domain holding authority over the entire structural space.

9. Distinctions from Prior Art

Content anchoring is structurally distinct from prior content-identity primitives in ways that are not bridgeable by parameter tuning.

Cryptographic hashing (MD5, SHA-2, SHA-3) computes byte-equality fingerprints. Any byte change produces a different hash. Hashing cannot survive the routine transformations that media undergoes in production systems. Content anchoring differs by computing identity from structure rather than bytes; the same artifact in a different container, codec, or resolution yields the same anchor within drift envelopes.

Digital watermarking embeds a synchronizing signal into the artifact. Watermarks fail under heavy compression, cropping, generative editing, and adversarial removal, and they require the embedding step to occur before distribution. Content anchoring is intrinsic and computed from existing structure; nothing is embedded, nothing can be stripped, and the anchor is recoverable from any sufficient copy of the artifact.

Perceptual hashing computes similarity-tolerant fingerprints. Most perceptual hashing schemes are narrow (one media type, one transformation regime), brittle under adversarial edits, and lack a governance layer for resolving disputes. Content anchoring differs by combining structural measurements drawn from multiple stability families, by maintaining a lineage layer that records derivative-class transitions positively, and by operating within an adaptive index that supports navigational resolution rather than nearest-neighbor search over a global table.

C2PA and related manifest-based provenance schemes attach signed metadata to artifacts. Manifests are stripped when artifacts are re-encoded, screenshotted, or processed by non-participating tools, after which provenance is lost. Content anchoring requires no metadata; identity is recoverable from the artifact itself, and a manifest, when present, becomes an optional cross-validating witness rather than a load-bearing carrier.

Blockchain-based content provenance commits content hashes or descriptors to a global ledger. The model inherits the byte-equality problem of cryptographic hashing (any transformation breaks the commitment) and requires global consensus operations for routine commitment. Content anchoring resolves identity through domain-partitioned adaptive indices without global consensus and survives transformation through the structural drift model.

None of these prior systems supports the absence-as-evidence property. Watermarks can be absent because they were never added; manifests can be absent because they were stripped; hashes are always different after any change. Only structural anchoring with positive governed lineage produces a system in which the absence of a lineage record for a claimed origin is structurally probative.

10. Disclosure Scope and Limitations

This disclosure specifies the conditions under which media identity can be computed from intrinsic entropy structure, maintained across legitimate transformation as a positively recorded lineage chain, resolved through adaptive index partitions without global consensus, and validated such that absence of governed lineage for a claimed origin is itself probative evidence of forgery. The disclosure encompasses the anchor-bound identifier construct, structural measurement and stability-preserving quantization, lineage chains under documented transformation classes, derivative-class detection, the no-watermark and no-metadata-required architecture, adaptive-index resolution, and composition with adjacent primitives.

The disclosure does not assert universal robustness against all adversarial transformations. Operations that exceed the documented mutation envelope, including heavy cropping, generative substitution, adversarial restructuring, or compositing of unrelated artifacts, can produce a new anchor with no intrinsic structural link to the prior. This is expected behavior, not a failure mode: in such cases the architecture supports best-effort discovery and post-hoc analysis where residual structure remains, and rights-grade provenance is supplied by governed execution surfaces in which artifacts are anchored before mutation, not by retroactive attribution of arbitrary outputs.

The disclosure does not describe a system for interpreting content for meaning. The structural measurements are deterministic and model-free; they do not require, and the architecture does not depend on, semantic inference about what the artifact depicts. Governance scopes are bounded by structural neighborhoods, not by content categories.

Within the operating envelope, content anchoring offers a structural alternative to byte-equality hashing, watermarking, manifest-based provenance, and blockchain-based content registries that is, by construction, more resilient to routine media transformation, more resistant to adversarial stripping, and more compatible with decentralized governance than the systems it is intended to displace. References to scalable provenance, AI-dataset governance, and forgery detection describe properties of the primitive within its operating envelope and do not assert deployment completeness or outcome guarantees in any particular regulatory or operational context.