Cross-Band Resolution Pathfinding: Traversal Between Variance Bands Under Mutation

Mechanism

The cross-band resolver maintains, for each anchored work, a band-indexed manifest of structural-variance signatures derived at registration time. The visual band signature is computed over a normalized perceptual representation of the work, stable under common visual mutations such as re-encoding, cropping within disclosed bounds, color-space conversion, and modest geometric warping. The audio band signature is computed over a normalized spectro-temporal representation, stable under sample-rate conversion, low-bitrate transcoding, additive noise within disclosed bounds, and short silences. The RF watermark band signature, when present, is a deliberately injected low-amplitude pattern recoverable from the carrier signal under typical broadcast and re-broadcast conditions.

When a candidate fragment is presented for resolution, the resolver extracts whichever band signatures the fragment will support. A fragment that arrives as a video file yields all three bands; a fragment that arrives as a transcript yields none of the three but yields a derived textual signature that the resolver treats as an additional band; a fragment that arrives as a screen-capture yields visual and audio but not RF. For each yielded band, the resolver computes a similarity score against the manifest and produces a per-band candidate set ranked by similarity.

The cross-band step is the joining of these per-band candidate sets. Rather than choosing the top candidate from any single band, the resolver constructs a graph in which nodes are works in the manifest and edges are weighted by the per-band similarity scores. A pathfinder traversal across this graph identifies the work whose joint similarity, summed and normalized across the available bands, exceeds a configured admission threshold by the largest margin. The traversal explicitly accounts for content that drifts between bands under mutation: a work that originated as audio but has been re-rendered with visual accompaniment will appear strongly in the audio band of the manifest and weakly in the visual band, and the traversal weights the available evidence accordingly.

The output of the resolver is not a single answer but a tuple of the matched work identifier, the per-band similarity scores that contributed, the joint confidence, and the residual unexplained signal. The residual is an explicit budget of signal in the candidate fragment that the resolver could not attribute to the matched work, expressed as a fraction of the input signal energy in each band. A resolution with low residual indicates a complete match; a resolution with high residual indicates either a derivative or a composite work, and the resolver flags the fragment for further analysis rather than declaring a single match.

Drift between bands under mutation is the central problem this mechanism is built to handle. A work registered as a video may, by the time it is presented for resolution, have been reduced to its audio track alone, summarized as a transcript, or re-rendered with new visuals over the original audio. Each of these transformations strips signal from one band while preserving signal in another. A single-band resolver, configured to prefer the stripped band, will fail; a single-band resolver configured to prefer the surviving band may succeed, but only by accident of configuration. The cross-band resolver succeeds in all of these cases by construction, because the joint similarity surface is shaped to weight whichever band's signal remains. The pathfinder traversal is deliberately willing to follow strong evidence in a minority of bands when the majority of bands have been stripped, provided the surviving evidence exceeds the band-cardinality-adjusted admission threshold.

Joint resolution is strictly stronger than any single band because the joint similarity surface admits fewer false positives. A visual collision is unlikely to coincide with an audio collision and an RF collision simultaneously; an adversary attempting to spoof identity must spoof in every band concurrently, which is materially harder than spoofing in any one. Joint resolution also degrades gracefully: when a band is missing, the resolver simply omits it from the joint score and adjusts the admission threshold to the disclosed schedule for that band cardinality.

Operating Parameters

The resolver is parameterized by per-band similarity metrics, per-band admission contributions, joint admission threshold, and traversal depth. Per-band similarity metrics are the established structural-variance distances disclosed in the parent application; the resolver does not impose a particular metric but does require that each band's metric be calibrated such that a unit of similarity in one band represents the same evidentiary weight as a unit of similarity in another. Calibration is performed at registration time across a reference corpus and is committed to the manifest so that later resolutions reproduce the same calibration.

The joint admission threshold is expressed as a function of band cardinality. With three bands available, the threshold is set such that joint similarity must exceed the sum of any two single-band thresholds. With two bands, the threshold is the sum of the per-band thresholds plus a disclosed margin. With one band, the resolver falls back to single-band resolution but flags the result with reduced confidence. The threshold schedule is committed to the manifest and is not adjustable per query.

Traversal depth governs how many candidate works are joined across bands. Shallow traversal joins the top-N candidates from each band; deeper traversal expands neighbors of high-scoring candidates to discover near-matches that would be missed by a strict top-N intersection. The disclosed range is N from sixteen to two hundred fifty-six, with the upper end appropriate to corpora of order ten million works. Traversal beyond the disclosed range produces diminishing returns and is not contemplated.

Residual budget thresholds govern when a resolution is declared a match versus flagged as derivative or composite. The disclosed budget is up to fifteen percent of input signal energy in any single band for a clean match, with totals up to twenty-five percent across bands; above these levels, the fragment is treated as derivative and is reported to the caller as such rather than as a clean identification.

Calibration drift is itself a parameterized concern. Over the lifetime of a corpus, the underlying distribution of similarity scores in any band may drift as new content is registered, as encoders and codecs evolve, and as adversaries probe the system. The resolver tracks calibration drift by periodically computing similarity statistics over a held-out validation subset and recording them in the manifest. When drift exceeds a disclosed bound, the resolver flags the manifest for re-calibration; pending re-calibration, it widens its admission thresholds defensively to avoid false positives that would result from outdated calibration. The drift bound and the defensive widening are themselves committed parameters and are visible to any party verifying a resolution.

Alternative Embodiments

In a first embodiment, the resolver runs as a centralized service against a single coherent manifest. This embodiment is appropriate to platform operators who hold a closed corpus and who can perform calibration once at corpus boundary changes. In a second embodiment, the resolver is federated across multiple manifests held by independent operators, with cross-manifest resolution performed by exchanging per-band candidate sets rather than raw signatures. The federated embodiment preserves manifest privacy and is suited to multi-platform deployments.

A third embodiment generalizes the band set to include additional modalities such as derived textual signatures, motion-vector signatures for video, and acoustic-event signatures for environmental audio. The mechanism is band-cardinality agnostic: any band whose similarity metric can be calibrated against the reference may be added without changing the traversal logic. A fourth embodiment runs the resolver on edge devices against a locally cached subset of the manifest, with cache misses resolved against a remote service; this embodiment is suited to provenance checks performed close to the point of consumption.

A fifth embodiment integrates the resolver with a streaming ingest pipeline so that resolution is performed continuously over windowed segments rather than once per fragment. Streaming resolution allows the system to detect content boundaries within a stream and to identify works that begin partway through a session, at the cost of additional state maintained per stream.

A sixth embodiment is adversarial-aware: it accepts as input a model of expected mutations between registration and resolution and pre-computes per-band signature variants under those mutations. The variants are stored alongside the canonical signatures and are searched in parallel, with the resolver preferring the canonical variant when both score highly and falling back to the mutation-aware variant otherwise.

Composition With Surrounding Primitives

The cross-band resolver composes with the structural-variance registration pipeline, consuming the manifests it produces and respecting the calibration committed at registration. It composes with the lineage primitives, recording every resolution decision into a tamper-evident log so that later parties can audit not only what was matched but what evidence supported the match. It composes with the policy and access-control primitives, deferring to upstream policy on whether a resolved identity may be disclosed to the caller, summarized, or returned only as an opaque token.

This compositional posture means that the resolver is not an island. A resolution it produces can be reconciled by any later party against the same manifest, the same calibration, and the same residual budgets, and that party will reach the same conclusion. This determinism is what makes the cross-band design useful in provenance and licensing contexts: the parties to a dispute can independently verify the resolution rather than trusting the resolver's verdict on faith.

Prior-Art Distinction

Single-band content fingerprinting is well-established. Visual perceptual hashes, audio fingerprints in the style of Shazam-class systems, and digital watermarks each have substantial prior art. What is absent from that art is the joint resolution across heterogeneous bands with calibrated similarity metrics, an explicit residual budget, and a band-cardinality-aware admission schedule. Prior multi-modal systems generally select a single band per query rather than joining bands, or they fuse bands with ad hoc weights that are not committed to a manifest and are not reproducible across implementations.

The disclosed mechanism is also distinguished from generic ensemble classifiers, which combine model outputs but do not preserve the per-band evidentiary trail required for audit. Here, every band's contribution to the joint score is recoverable, and the residual budget makes explicit what fraction of the input the resolver could not explain. This evidentiary structure is, to the inventor's knowledge, not present in the prior art.

Disclosure Scope

This disclosure encompasses the cross-band resolution pathfinder as a structural component of the content anchoring system, including the band-indexed manifest, the per-band similarity calibration, the joint admission schedule, the residual budget mechanism, the traversal procedure, and the parameter ranges identified above. It encompasses centralized, federated, edge-cached, streaming, and adversarial-aware embodiments, and any combination thereof. The claim scope contemplated is independent of any particular band set, similarity metric, or substrate, and is defined by the structural relationship between bands, manifest, and traversal rather than by any specific implementation choice.