Role in the Pipeline
The stable sketch is the privacy-preserving middle layer of the biological identity architecture. It receives the normalized, continuity-suitable feature stream produced by the feature extraction and normalization module, and it produces a noise-tolerant, non-invertible representation, termed a stable sketch, that enables reproducible identity verification without requiring storage of, or recovery access to, the raw biological signal data from which it was derived. The stable sketch is the bridge between the rich but privacy-sensitive biological signal stream and the cryptographic biological hash that follows it: it ensures that hash generation operates on a representation that is simultaneously stable enough for reproducible verification and abstract enough to prevent reconstruction of the underlying biological data.
The disclosed framing matters because the sketch output is not used as a template for matching. Conventional biometric systems locate identity in an enrolled reference and score each observation against it. Here the stable sketch instead feeds a temporally bound biological hash that is validated as a plausible successor within a continuity chain, with banding supporting population-scale probabilistic resolution and domain separation supporting relying-party unlinkability. The sketch is a stage in establishing behavioral continuity over time, not a stored profile.
The Multi-Stage Construction
The stable sketch is generated through a multi-stage process comprising dimensional reduction, projection, quantization, and helper data generation. The dimensional reduction stage reduces the high-dimensional normalized feature stream to a lower-dimensional representation that captures the identity-relevant variance while discarding noise-dominated dimensions. This reduction is performed through a learned projection that is specific to the modality combination in use but is not specific to any individual: the projection is a system-wide parameter mapping from the normalized feature space to the sketch space, and it does not carry individual-identifying information.
The projection stage applies a set of random but fixed projection vectors to the dimensionally reduced feature stream, producing a set of projected values that encode the individual's feature state in a representation suitable for quantization. The disclosure draws on cryptographic primitives described in the literature, including locality-sensitive hashing, secure sketches, and fuzzy extractors, and integrates them within the trust-slope continuity framework rather than treating sketch reproduction as a standalone matching operation.
Banding and Probabilistic Assignment
The quantization stage partitions the projected value space into discrete bands. Each projected value is assigned to a band based on which partition region it falls within, and the band assignments collectively constitute the stable sketch: a vector of discrete band indices that represents the individual's biological state in the projected space. The critical property of the banding scheme is that it is designed for noise tolerance through the use of overlapping or adjacent band regions with probabilistic resolution at band boundaries.
When a projected value falls near a band boundary, the individual's feature state may be assigned to either of the adjacent bands depending on measurement noise, physiological fluctuation, or environmental conditions. Rather than treating this ambiguity as an error, the architecture treats band-boundary ambiguity as an inherent and useful property: the stable sketch is understood as a probabilistic assignment rather than a deterministic one, and the trust-slope continuity validation that consumes the sketch is designed to accommodate the expected frequency of band-boundary transitions in its continuity assessment.
Helper Data Generation
Helper data is generated to enable reproducible band assignment without revealing the underlying biological feature values. The helper data is a set of values that, when combined with the correct biological signal, reproduces the same band assignments, but that cannot be used alone to recover either the biological signal or the band assignments. The helper data generation follows a secure sketch construction in which the helper data encodes the offset between the individual's projected feature values and the nearest band center, enabling subsequent capture events to correct for noise-induced deviations and reproduce the original band assignment within the noise tolerance of the scheme.
The helper data is stored alongside the biological hash chain and is updated at policy-governed intervals to accommodate gradual physiological drift. Because the architecture maintains no enrolled template, the helper data carries the burden of reproducibility: it lets a later capture snap back to the band assignments of an earlier one without the system ever holding the feature values that produced them.
Structural Non-Invertibility
The non-invertibility of the stable sketch is a structural property of the architecture, not an assumption about computational difficulty. The dimensional reduction discards information that cannot be recovered. The projection applies a many-to-one mapping that is not invertible even with knowledge of the projection vectors. The quantization discards all within-band precision. The combined effect of these three stages ensures that the stable sketch carries sufficient information for continuity validation, since the sketch of a subsequent capture from the same individual will exhibit consistent band assignments, but does not carry sufficient information to reconstruct the biological signal, the normalized feature stream, or the intermediate representations from which it was derived.
Hierarchical Band Resolutions
The banding scheme supports multiple band resolutions. A coarse banding with fewer, wider bands produces sketches that are more stable across captures, because the same individual is more likely to produce the same coarse band assignments despite measurement noise, but it provides less discrimination between different individuals. A fine banding with more, narrower bands provides greater discrimination but requires higher signal quality to achieve stable assignments. The architecture supports hierarchical banding in which both coarse and fine band assignments are computed for each capture, enabling continuity validation to operate at multiple resolutions simultaneously: coarse-band continuity for robust validation under noisy conditions, fine-band continuity for high-assurance validation when signal quality permits.
Composition with Hashing and Disambiguation
The stable sketch band assignments are one of the composite inputs to biological hash generation, combined there with a temporal binding value, a domain separation tag, and a rotating salt to produce a temporally bound, domain-scoped biological hash. Each entry in the biological trust-slope chain is such a hash, and continuity validation evaluates the sequence of hashes for trajectory coherence.
The hierarchical bands also drive population-scale disambiguation. A candidate narrowing stage operates on coarse-band stable sketch assignments to reduce the population of potential matches to a manageable candidate set at low computational cost. A fine-band comparison stage then applies the higher-resolution fine-band assignments to discriminate further among the candidates. A trust-slope reinforcement stage resolves remaining ambiguity by exploiting the temporal dimension: even if two individuals produce similar stable sketches at a single point in time, the probability that they produce similar trust-slope trajectories across multiple independent time points decreases as more time points are considered. The sketch supplies the discriminating structure; the continuity chain supplies the temporal resolution.
Disclosure Scope
The stable sketching mechanism, comprising the dimensional reduction, the projection through random but fixed projection vectors, the band-based quantization producing a vector of discrete band indices, the probabilistic treatment of band-boundary assignment, the secure-sketch helper data encoding the offset to the nearest band center, the structural non-invertibility arising from the three information-discarding stages, the hierarchical coarse and fine band resolutions, and the composition with domain-separated biological hash generation and coarse-band-to-fine-band population disambiguation, is disclosed in the cognition filing (U.S. Application No. 19/647,395 and its international counterpart) at Section 9.5. This article describes that disclosed mechanism. The scope is not limited to any specific biological modality, sensor technology, projection family, or band layout, provided the sketch remains a noise-tolerant, non-invertible representation consumed by trust-slope continuity validation rather than by template matching.
