Rights-Grade Content Generation With Provenance Tracking

Mechanism

The mechanism is an inference-time admissibility pipeline that wraps the generation loop of a model. At session initialization, a rights manifest is constructed by composing the model's training-governance ledger, which enumerates the rights envelopes under which each contribution was admitted, with a request-level rights envelope supplied by the calling application, which enumerates the rights envelopes under which the present output may be produced and consumed. The intersection of these two envelopes, computed deterministically by an admissibility resolver, defines the per-session rights manifest that governs the generation loop.

At each generation step, the candidate distribution emitted by the model is filtered by a rights-aware admissibility gate before sampling. The gate consults the rights manifest, the running output prefix, and a content-anchoring primitive that supplies, for each candidate region of the output, the set of training contributions whose influence on that region exceeds a salience threshold. Candidates that would, if emitted, produce a region whose contributing training set falls outside the rights manifest are masked from the distribution. The masking happens before the sampling step rather than after, so the model never produces an inadmissible token that must subsequently be retracted, and the entropy cost of the masking is accounted for in the same step.

Specific rights dimensions are evaluated as named sub-gates of the admissibility gate. A copyright sub-gate refuses candidates whose contributing set includes works admitted under a non-derivative license when the running output exhibits substantial-similarity indicators against any of those works, computed by a streaming similarity estimator. A trademark sub-gate refuses candidates that would complete a registered mark in a context whose rights envelope does not include the mark's owner. A persona sub-gate refuses candidates that would produce identifiable likeness or voice attributes when the contributing set includes persona-restricted contributions and the request envelope does not include a persona license. A training-data-provenance sub-gate refuses candidates whose contributing set includes contributions whose admissibility credentials have been revoked since training, which closes the loop with the deletion machinery in the training-governance ledger.

The output stream is annotated with a structural witness for each emitted region, recording the rights manifest hash, the contributing-set summary, and the sub-gate verdicts. The witnesses are aggregated into a generation-provenance document signed at output close, and the document is the artifact that downstream consumers, rights holders, and auditors verify. Because the witnesses are produced inline with generation rather than reconstructed afterward, they are causally grounded in the actual sampling decisions and are not subject to the reconstruction gap that defeats post-hoc provenance schemes.

Operating Parameters

The rights manifest is parameterized by an envelope schema that enumerates dimensions, by a per-dimension policy that specifies admit, deny, or attribute behavior, and by a salience threshold per dimension that controls how strongly a training contribution must influence a candidate region before the sub-gate is invoked. The salience estimator is a separate primitive that operates on the model's attention or activation traces and produces, per candidate, a contributing-set summary suitable for the gate. The salience threshold is tuned per dimension to balance false-deny rate against rights-leakage rate, and the chosen thresholds are recorded in the witness so that downstream auditors can reconstruct the gate's behavior.

The streaming similarity estimator used by the copyright sub-gate operates on a sliding window of the running output and a precomputed index over the contributing works, returning a per-candidate similarity score against each work in the contributing set. The index is shared across sessions and refreshed when the training-governance ledger emits a deletion event. The trademark sub-gate consults a registered-marks database that is itself signed and refreshed at a cadence appropriate to the deployment jurisdiction. The persona sub-gate consults a persona-rights registry that is consulted as a fail-closed dependency.

Operationally, the gate is implemented as a pre-sampling logits transformation that consumes the rights manifest and the contributing-set summary as auxiliary inputs and emits a masked logits tensor. The masking adds latency that is bounded by the size of the contributing set and the number of active sub-gates, and the implementation supports per-deployment caching of contributing-set summaries to amortize cost across long generations. Witnesses are emitted in canonical CBOR and signed at session close with a key bound to the model release.

The gate enforces a fail-closed posture on each of its dependencies. A salience estimator that returns a low-confidence summary causes the gate to admit only candidates whose contributing set is unambiguously inside the manifest, rather than admitting under uncertainty. A registered-marks database that fails to refresh within its freshness window causes the trademark sub-gate to revert to a conservative deny on any candidate that completes a known-prefix mark. A persona-rights registry that is unreachable causes the persona sub-gate to deny rather than to admit. Each of these fail-closed transitions is itself recorded in the witness so that the downstream provenance document distinguishes admissions made under full evidence from admissions made under degraded evidence and denials made under degraded evidence from denials made under positive evidence of inadmissibility.

Alternative Embodiments

A first alternative embodiment applies the gate at the granularity of a generation segment rather than a token, in which the model produces a candidate segment and the gate decides whether to commit, mask, or regenerate. This embodiment is appropriate for image and audio generation in which the natural unit of admissibility is a region or a frame rather than a token.

A second alternative embodiment replaces the deterministic mask with a tilt: candidates whose contributing set lies near the boundary of the rights manifest are down-weighted rather than masked, and the tilt magnitude is itself recorded in the witness. This embodiment is appropriate when strict masking would produce unacceptable degeneration and when the rights envelope admits a graded response.

A third alternative embodiment couples the rights manifest to a dynamic envelope that updates mid-session in response to user signals, so that a user who acquires a license during a session can broaden the manifest without restarting. The envelope updates are themselves signed and recorded.

A fourth alternative embodiment composes multiple rights manifests for multi-tenant generation, in which a single inference serves several rights envelopes simultaneously and the gate emits separate witnesses per envelope. The model's output stream is replicated per envelope only to the extent that the envelopes diverge, which keeps the marginal cost low for envelopes that share most dimensions.

A fifth alternative embodiment integrates the gate with a content-anchoring primitive that emits cryptographic anchors for each output region rather than only structural witnesses, which permits downstream verification without access to the witness payload.

Composition

The primitive composes upward with the training-governance primitives, which supply the per-contribution admissibility ledger that descends into the rights manifest. The composition is strict in the sense that a contribution cannot influence a generation under a rights envelope it was not admitted to support, and the structural enforcement of this property is what permits the generation to bear a meaningful rights claim at output time. The primitive composes laterally with the content-anchoring primitive, which supplies the per-region anchors that the witness references and that downstream verifiers consult.

The primitive composes laterally with a salience-estimation primitive that bridges the model's internal traces and the gate's contributing-set input. The salience primitive is replaceable: any estimator that produces a contributing-set summary in the canonical schema can be substituted without changing the gate. The primitive composes downward with a generation-provenance ledger that aggregates witnesses into signed documents and that participates in downstream rights-clearance and royalty-attribution flows.

The composition with training-governance closes a loop that is open in conventional generative pipelines: a deletion event in the training ledger, communicated to the salience index, immediately propagates into the gate's sub-gate behavior, so that contributions revoked after training cease to influence admissible generations without requiring model retraining.

Prior Art

Prior approaches to rights compliance in generative AI rely on training-data filtering, output filtering, or post-hoc indemnification, each of which addresses a fragment of the problem. Training-data filtering removes inadmissible contributions before training but provides no guarantee about the output of the trained model and no mechanism to respond to post-training rights events. Output filtering applies pattern-based or classifier-based filters to generated content but operates after sampling and produces no causal witness linking the output to the contributions that shaped it. Post-hoc indemnification shifts the rights risk onto the platform without addressing its structural source.

Watermarking and content-authenticity standards establish that a piece of content was produced by a particular model or service but do not establish that the content respects the rights of the contributions that influenced it. The disclosed primitive is distinguished by the inline, sub-gate-decomposed, contribution-aware admissibility discipline and by the composition with training-governance and content-anchoring that produces a structurally grounded rights claim rather than a post-hoc assertion.

Disclosure Scope

The disclosure covers the inference-time, contribution-aware, sub-gate-decomposed rights-grade generation primitive in any embodiment that composes a training-governance ledger with a request-level rights envelope, gates the candidate distribution against the resulting manifest before sampling, evaluates copyright, trademark, persona, and training-data-provenance sub-gates against a salience-derived contributing-set summary, and emits inline structural witnesses aggregated into a signed generation-provenance document. The disclosure extends to token-level, segment-level, image, audio, and multi-tenant embodiments and to embodiments that integrate with content-anchoring for cryptographic regional anchors. The disclosure does not extend to embodiments that omit the training-governance composition, that omit the per-candidate admissibility gate, or that omit the inline witness, since each of these elements is essential to the structural rights guarantee that distinguishes the primitive from filtering and indemnification approaches.