Core CCF Architecture
The foundational patent for Contextual Coherence Fields. Trust is earned through interaction, not assigned. The minimum gate ensures behavioral output can never exceed earned trust. Doubly stochastic mixing conserves trust energy. Four social phases emerge from coherence-tension phase space.
What it is
CCF is a behavioral architecture for autonomous systems where trust is a mathematical quantity, not a policy label. Every context the robot encounters gets a composite key derived from its sensor readings. Each key accumulates coherence through positive interactions over time—short-run reactivity and long-run familiarity tracked as dual exponential moving averages.
The architecture’s central invariant is the minimum gate: C_eff = min(C_inst, C_ctx). Effective coherence is the minimum of instantaneous environmental stability and accumulated context trust. A robot in a familiar room that suddenly goes dark gets capped by the instability. A robot in a stable but unfamiliar room gets capped by lack of history. Capability without earned trust produces zero behavioral output.
Cross-context trust transfer is governed by Sinkhorn-Knopp projection onto the Birkhoff polytope—doubly stochastic matrices where rows and columns sum to one. Trust cannot be created from nothing or amplified through transfer. Context boundaries are discovered by Stoer-Wagner min-cut on the similarity graph. Four social phases (ShyObserver, StartledRetreat, QuietlyBeloved, ProtectiveGuardian) emerge from the coherence-tension phase space with Schmitt trigger hysteresis preventing oscillation at boundaries.
Problems it solves
Social engineering resistance
The minimum gate means a robot cannot be talked into trusting a new person in a familiar room. Both instantaneous stability and accumulated history must be high. A stranger in Grandma's kitchen gets ShyObserver, not QuietlyBeloved.
Context isolation
Stoer-Wagner min-cut discovers natural boundaries in the context similarity graph. Trust earned in the living room doesn't bleed into the garage. Each context partition is mathematically isolated by cut weight.
Trust inflation prevention
Doubly stochastic mixing via Sinkhorn-Knopp ensures rows and columns sum to one. Trust is a conserved quantity. You cannot create trust from nothing or amplify it through transfer between contexts.
Behavioral oscillation
Schmitt trigger hysteresis on phase boundaries prevents the robot from flickering between shy and warm when coherence hovers near a threshold. A 0.10 deadband keeps behavior stable.
Real-world scenarios
Eldercare companion
A companion robot arrives at a care home. On day one it is ShyObserver with every resident—quiet, watchful, LED tint cool blue. Over weeks, Mrs. O’Brien visits the common room every morning. The robot’s long-run accumulator for her context rises. Eventually it reaches QuietlyBeloved: warm amber LED, gentle motor movements, responsive voice. When a new carer starts on the night shift, the robot doesn’t transfer Mrs. O’Brien’s trust to the stranger. The carer gets ShyObserver until they earn their own coherence through consistent interaction.
Warehouse AMR
An autonomous mobile robot operates across warehouse zones. In zones it knows well—consistent lighting, familiar noise patterns, stable layout—it moves at full speed with confident path planning. When rerouted to a new zone after a reorganization, the context key changes. C_ctx is near zero. The minimum gate caps effective coherence regardless of how capable the robot is. It slows down, increases sensor polling, and operates with caution until the new zone becomes familiar.
Classroom assistant
A robot helps children with science experiments. The classroom is a familiar context—high C_ctx. But when a child drops a beaker and the sound level spikes, C_inst drops. The robot doesn’t panic or freeze—it shifts to StartledRetreat (high context trust, low instantaneous stability), pulling back its arms and pausing its current task until the environment stabilises. Hysteresis prevents it from flickering between retreat and normal as the noise dies down.
What the claims cover
Claims 1, 8 -- SensorVocabulary / ContextKey
The fundamental encoding layer. Any sensor set implementing the SensorVocabulary trait produces a deterministic, hashable ContextKey. This is the identity layer that all subsequent computation keys off.
Claims 2--5 -- CoherenceAccumulator
Dual exponential moving averages per context. Short-run reactivity and long-run familiarity tracked independently, with configurable decay constants and personality-modulated scaling.
Claims 6--7, 13 -- CoherenceField
The field layer mapping ContextKey to CoherenceAccumulator. Supports positive interaction, effective coherence computation via the minimum gate, and ranked context reports for dashboards.
Claims 9--12 -- MinCutBoundary
Stoer-Wagner global min-cut on the context similarity graph. Separates contexts into natural partitions. Cross-partition influence is bounded by cut weight.
Claims 14--18 -- SocialPhase
Four discrete social phases from the coherence-tension phase space. LED tint, motor scale, and voice tone per phase. Schmitt trigger hysteresis prevents oscillation at boundaries.
Claims 19--23 -- SinkhornKnopp
Iterative row-column normalisation projecting the cross-context matrix onto the Birkhoff polytope. Trust conservation enforced. no_std compatible with fixed-size arrays.
Applications
Licensing enquiries
CCF is released under BSL-1.1 — free for evaluation and non-commercial use. Commercial licensing is available from Flout Labs.
cbyrne@floutlabs.com