6 US Provisionals Filed · 146 Claims · Patent Pending

CCF Patent Portfolio

Six US provisional patent applications filed between February and April 2026 by Colm Patrick Byrne / Flout Labs, Galway, Ireland. 146 claims covering core CCF architecture, extensions, trust-constituted action, fleet analytics, self-model and meta-awareness, and enablement specifications.

Filing portfolio

Priority date: 23 February 2026. Non-provisional deadline: 23 February 2027. All filings by Colm Patrick Byrne, micro entity.

US 63/988,438
February 23, 2026 · 34 claims

Context-keyed coherence accumulation, minimum gate, doubly stochastic mixing via Sinkhorn-Knopp, Stoer-Wagner min-cut boundary, four-quadrant social phases, personality modulation.

US 63/994,113
March 2, 2026 · 28 claims

Hierarchical block-diagonal mixing, environmental identity formation with irreversibility proofs, privacy-responsive trust accumulation, emergent home context discovery.

US 64/037,374
April 13, 2026 · 25 claims

Trust-constituted action spaces for autonomous intelligence. Behavioral gating where unsafe actions are geometrically absent, not policy-prohibited. LLM coherence verification.

US 64/039,626
April 15, 2026 · 13 claims

Implementation-grade parameter defaults, Sinkhorn-Knopp convergence bounds, corrected simulation evidence, self-contained claim rewrites for Prov 1 and Prov 2.

US 64/039,623
April 15, 2026 · 30 claims

Privacy-preserving operational identity fingerprinting for autonomous agent fleet monitoring. Environmental classification, relocation detection, deployment verification, drift monitoring — all from fewer than 20 scalar values without raw sensor data.

Read-only relational self-model with hardware-enforced trust isolation. Evidence-preserving accumulator merge/split with reversible lineage. Behavioral causation trail for privacy-preserving explainability. Sponsored relational bridging with temporary trust scaffolds for socially introduced transitions.

Core claim map — Prov 1 (US 63/988,438)

34 claims across 6 modules. Each implemented in ccf-core on crates.io. The claim numbers below correspond to the core patent application.

Claims 1, 8

SensorVocabulary / ContextKey

The fundamental encoding: any sensor set implementing SensorVocabulary<N> produces a deterministic, hashable ContextKey. This is the identity layer — every subsequent computation is keyed to it.

#1
A method for encoding multi-dimensional sensor state into a deterministic context identity using a SensorVocabulary trait with const-generic feature vector length N.
#8
A ContextKey structure combining a vocabulary instance with a pre-computed FNV-1a hash and normalised feature vector, enabling O(1) context lookup and cosine similarity comparison.
Claims 2–5

CoherenceAccumulator

The accumulation layer: dual exponential moving averages (short-run α_s and long-run α_l) per context key. Positive and negative interactions advance or decay each EMA independently.

#2
A CoherenceAccumulator maintaining two EMAs per context with configurable decay constants λ_s and λ_l.
#3
A method for computing instantaneous coherence from α_s scaled by Personality.reactivity.
#4
A method for computing context coherence from α_l representing accumulated long-run evidence.
#5
Decay-on-tick: both EMAs decay toward zero on every tick regardless of interaction event, providing natural forgetting.
Claims 6–7, 13

CoherenceField

The field layer: a map from ContextKey to CoherenceAccumulator. Supports positive_interaction(), effective_coherence(), and report_context_with_key() across all known contexts.

#6
A CoherenceField maintaining a bounded map of CoherenceAccumulators keyed by context hash, with O(1) lookup and insert.
#7
effective_coherence(): combines instantaneous and context coherence via the minimum gate condition.
#13
report_context_with_key(): produces a ranked context report across all known keys, enabling dashboard and audit output.
Claims 9–12

MinCutBoundary

The partition layer: Stoer–Wagner global min-cut on the context similarity graph. Separates the context space into two partitions; cross-partition influence is bounded by cut weight.

#9
A method for constructing a weighted context graph where edge weights are cosine similarities of context feature vectors.
#10
The Stoer–Wagner algorithm applied to the context graph to find the global minimum cut in O(|V|^3).
#11
A MinCutResult structure reporting partition_s, partition_complement, and min_cut_value.
#12
Cross-partition coherence blending proportional to min_cut_value, bounded to [0, 1].
Claims 14–18

SocialPhase

The output layer: four discrete social phases (ShyObserver, StartledRetreat, QuietlyBeloved, ProtectiveGuardian) determined by the minimum gate. Each phase maps to LED tint, motor scale, and voice tone.

#14
Four SocialPhase variants derived from the Cartesian product of (α_s threshold) × (α_l threshold).
#15
led_tint(): a method returning a 24-bit RGB colour per phase as a visual output signal.
#16
expression_scale(): a method returning a f32 motor/movement scale per phase.
#17
The minimum gate: phase is ShyObserver unless both α_s > θ_instant AND α_l > θ_context.
#18
Personality struct with bounded warmth, reactivity, caution fields modulating all thresholds.
Claims 19–23

SinkhornKnopp

The manifold layer: iterative row-column normalisation that projects the cross-context coherence matrix onto the doubly-stochastic manifold, preserving the conservation of social budget.

#19
A SinkhornKnopp projector taking an N×N coherence matrix and returning a doubly-stochastic matrix.
#20
Iterative alternating normalisation: D_r · A · D_c converging to A*.
#21
Convergence criterion: max absolute row/column sum deviation < ε.
#22
Integration with CoherenceField: SinkhornKnopp applied after cross-partition blending.
#23
no_std compatibility: projector uses fixed-size arrays, no heap allocation.

Prov 2 — Architectural Extensions (US 63/994,113)

28 claims across 4 extensions. Hierarchical mixing, identity formation, privacy trust, emergent home.

Claims A–C

Ext 1: Hierarchical Mixing

Block-diagonal intra-cluster + inter-cluster mixing. Reduces computational cost from O(n²) to O(k² + Σnᵢ²), enabling CCF on resource-constrained embedded hardware with hundreds of contexts.

A
Hierarchical composite mixing operation: intra-cluster doubly stochastic matrices + inter-cluster doubly stochastic matrix, independently projected onto Birkhoff polytope.
B
Clamp activation analysis: approximation error bounded by ε_clamp, zero during normal operation. Three properties hold exactly when clamp is inactive.
C
Computational cost reduction: O(k² + Σnᵢ²) per tick vs O(n²) for flat mixing, where k = cluster count and nᵢ = cluster sizes.
Claims D–H

Ext 2: Environmental Identity Formation

Ontogenetic identity emerges from environmental interaction under manifold constraints. Three irreversibility theorems prove robots in different environments develop formally incompatible identities.

D
Context vocabulary determined by environmental sensor signal distribution, not configuration parameters.
E
Ontogenetic identity formation: vocabulary, coherence landscape, group topology, mixing topology, and compiled repertoire constitute an emergent, irreversible relational identity.
F
Theorem 1 — Mixing matrix incompatibility: different-dimension doubly stochastic matrices cannot be merged while preserving constraints.
G
Identity fingerprint: compact vector derived from accumulated operational state for fleet monitoring and environmental classification.
H
Fleet-scale analytics: environmental classification, relocation detection, deployment verification, and drift monitoring from fingerprint vectors.
Claims I–L

Ext 4: Privacy-Responsive Trust

Privacy invocation increases trust — the paradox. Hardware sensor disconnection + null-content trust events + rarity-scaled increment resolves the privacy/personalisation tradeoff.

I
Hardware sensor disconnection (relay/shutter) within one sampling period. Null-content trust event with zero content bytes. Rarity-scaled trust increment.
J
Rarity factor: trust_increment = base_rate × (1.0 − current_value) × (1.0 − privacy_count/total_count). Rare privacy requests produce larger increments.
K
Earned floor increment: interaction count rises during privacy mode, raising the decay floor.
L
Three concrete technical effects: reduced bandwidth, increased resilience, hardware-verified disconnection.
Claims M–O, Z

Ext 5: Emergent Home Context

The robot discovers its safe haven through accumulation dynamics alone. No programmed coordinates, no beacons. Battery depletion couples to tension, driving return-to-charge through coherence gradients.

M-alt
Emergent safe-haven formation: highest-coherence context emerges from frequency, consistency, low tension, and long dwell — without programmed coordinates.
M
Charging-specific embodiment: battery state-of-charge coupled to tension via monotonically increasing transfer function.
O
Energy-as-tension integration: battery depletion increases total tension, driving preference for highest-coherence context.
Z
Beauregard hardware format: processor, sensor array, coherence function, energy monitor, tension integration, behavior selection — no stored dock coordinates.

Prov 5 — Trust-Constituted Action (US 64/037,374)

25 claims. Action spaces where unsafe behavior is geometrically absent. LLM coherence verification.

Claims 1–8

Trust-Constituted Action Spaces

Action spaces where unsafe behavior is geometrically absent from the behavioral manifold, not merely prohibited by policy. Capability without earned trust produces zero behavioral output.

#1
Trust-constituted action space: effective coherence computed from accumulated context trust, gating behavioral output through minimum function.
#6
Non-amplification: max(C_mixed) cannot exceed max(C_ctx). Trust transfer is bounded by doubly stochastic constraint.
#8
Domain accumulator schema: separate accumulators per conversational domain (task, personal, crisis, relational).
Claims 16–18

Time-Domain & Upstream Routing

Trust accrual capped per unit real time (clock can't be gamed). Distress signals routed upstream to stability computation, not downstream to response generator. Classifiers architecturally independent.

#16
Time-domain bound: trust accrual capped per calendar day, per session, and per minute. Volume cannot accelerate trust.
#17
Upstream affective routing: distress cues reduce C_inst before reaching the response generator. Prevents positive feedback loop.
#18
Provenance separation: signal classifiers run independently of the response LLM. Architecturally distinct signal sources.
Claims 19–25

Semantic Gating & Safety

Monotonic semantic envelope: registers unlock at specific trust thresholds. Grounding always available. Session terminates on acute distress. Safety persists across sessions. Useful cold-start floor.

#19
Semantic action-class gating: monotonic envelope from factual (0.0) through warmth (0.3) to mutuality (0.7) to exclusivity (0.85). Not filtered — constituted.
#20
Grounding override: crisis referral and safety information available at ALL trust levels, including maximum.
#21
Contested provenance default: events below confidence threshold default to model-originated (ineligible for trust accumulation).
#22
Session termination: generative channel closes when C_inst drops below acute distress floor. Cooldown period enforced.
#23
Domain partitioning via doubly stochastic context separation. Evaluation awareness rendered architecturally irrelevant.
#24
Safety continuity: acute risk state persists across sessions. Cooldown and decay carry forward.
#25
Cold-start floor: at zero trust, model provides factual answering, task completion, clarification, and grounding. Useful, not crippled.

Supplement — Enablement Specs (US 64/039,626)

13 claims. Implementation-grade parameters, convergence proofs, self-contained rewrites.

1A, 7A, 9, 14, 24, 29A

Prov 1 Amendments

Implementation-grade amendments: intermediate ceiling constraint, quadrant behavioral profiles with hysteresis, internal state variable enumeration, structural mutual-dependency, observable behavioral hesitation.

1A
Intermediate ceiling constraint: C_eff cannot exceed C_ctx for any value of C_inst. Accumulated trust is the unidirectional ceiling.
7A
Quadrant behavioral profiles: four-quadrant state space with hysteresis thresholds preventing behavioral oscillation at boundaries.
14
Rewritten independent claim: behavioral gating method with ceiling operation and permeability mapping across output modalities.
29A
Observable behavioral hesitation: classification conflict between reflexive and deliberative pathways produces visible amplitude reduction.
A, E, I, M-alt, M, O, Z

Prov 2 Rewrites (Self-Contained)

Self-contained claim rewrites eliminating parent-claim number dependencies. Each claim stands alone for prosecution flexibility.

A
Hierarchical mixing — self-contained rewrite with explicit O(k² + Σnᵢ²) cost and independent Birkhoff projection.
E
Ontogenetic identity formation — self-contained with all three irreversibility properties stated inline.
I
Privacy-responsive trust — rewritten with full technical specificity for §101 defense (hardware disconnect, null-content, rarity-scaled).
§0033a, §0054a, §0055d, §0058a, §0090a, §0098a

Enablement Specifications

Implementation-grade parameter defaults: sensor thresholds (50/300 lux, 40/65 dBA), sigmoid permeability (k=8, c₀=0.5), Sinkhorn convergence (20 iter → 1.2×10⁻⁸), Mahalanobis 3σ halt, loss function weights.

§0033a
Sensor quantization thresholds: 6 sensor dimensions with exact numeric defaults calibrated for domestic indoor deployment.
§0055d
Permeability mapping: sigmoid P(c) = 1/(1+exp(-8(c-0.5))) with tension-modulated midpoint shift ±0.1.
§0058a
Sinkhorn-Knopp convergence: Birkhoff contraction τ ≈ 0.403, 20 iterations, deviation bounded by 1.2×10⁻⁸ on 32-bit float.
§0090a
Distribution signature: mean vector + covariance matrix. Four halt conditions. Mahalanobis distance threshold 3.0σ.

Prov 3 — Fleet Analytics (US 64/039,623)

30 claims (6 independent). Privacy-preserving fleet monitoring from fewer than 20 scalar values.

Claims 1, 1B, 2–9

Fingerprint Computation & Use

Compute an 8-component identity fingerprint (fewer than 20 scalars) from accumulated operational state. Transmit only the fingerprint — no raw sensor data. Perform fleet monitoring functions.

#1
Method: compute fingerprint locally, transmit fixed-size vector, perform classification/relocation/verification/drift from fingerprint alone.
#1B
All four monitoring functions from same fingerprint vector without raw sensor data for any function.
#2
Extended fingerprint: state matrix density, context group count, temporal rhythm, presence pattern.
#4
Privacy-preserving by construction: fewer than 20 scalars, no raw sensor data, no individual identifiers, no interaction content.
#7
Composite identity distance: vocabulary distance + familiarity landscape distance + phase distribution distance.
Claims 10, 10B, 11–17

Fleet Monitoring Applications

Four fleet monitoring functions: environmental classification, relocation detection without GPS, deployment verification, environmental drift monitoring. Adaptive thresholds from fleet-wide statistics.

#10
Fleet monitoring method: each agent computes fingerprint, transmits to fleet server, server performs monitoring with operational outputs.
#12
Relocation detection: fingerprint drift against adaptive threshold (μ + 3σ). No GPS, beacons, or location sensors required.
#14
Environmental drift monitoring: sustained sub-threshold drift identifies staffing changes, schedule alterations, renovations from specific fingerprint component drift.
#15
Adaptive threshold from fleet-wide or cohort-specific drift statistics. Threshold adapts to each environment type.
Claims 18, 18B, 19–27

Fleet Analytics & System

Fleet-scale analytics: environment taxonomy discovery, outlier detection, cohort analysis, quality scoring. System claim with store-and-forward for intermittent connectivity.

#18
Fleet analytics: cluster fingerprints for taxonomy, detect outliers, compare cohort trajectories, compute environment quality scores.
#22
System: fleet server with fingerprint database, reference library, classification/relocation/verification/drift modules, analytics module.
#26
Store-and-forward method: compute fingerprints regardless of connectivity, buffer locally, transmit on reconnect, reconstruct time series without gaps.
#27
Store-and-forward system: agents with local buffer, fleet server with time series reconstruction, transport-agnostic communication.

Prov 4 — Self-Model, Merge/Split, Causation, Bridging (US 64/039,655)

16 claims (5 independent). Extensions 3, 6, 7, 8.

Claims AD–AG

Ext 3: Relational Self-Model

Read-only self-model grounded in earned state. 8-component vector: entitlement, familiarity, floor, maturity, uncertainty, lineage, bridge mass, habit availability. Hardware-enforced write isolation.

AD
Relational self-model: computed from extant mathematical state, read-only (prohibited from increasing any accumulator, floor, or mixing entry). Made available to behavior selector, planner, explanation interface.
AE
Trust-conservative shadow simulation: non-persistent copy of CCF state for counterfactual evaluation. All simulated changes discarded. No trust earned in imagination.
AF
Uncertainty-sensitive restraint: uncertainty from partition ambiguity, mixing entropy, conflict rate. Imposes downward constraint on expressiveness. Produces grounded disclosures.
AG
Grounded self-description: structured self-model variables mapped to verbal/visual/postural outputs. Distinguishes low familiarity vs low stability, continuity vs partial continuity.
Claims AH–AK

Ext 6: Accumulator Merge/Split

Evidence-preserving operations for context evolution. Merge when contexts prove equivalent, split when multimodal. Reversible lineage. Sensor vocabulary migration without erasing earned trust.

AH
Evidence-preserving merge: n_M = n_A + n_B, c_M = (m_A + m_B)/n_M. Evidence mass conserved. Lineage record stored. Birkhoff reprojection of reduced matrix.
AI
Reversible lineage and rollback: reconstruction coefficients stored. Post-merge divergence triggers rollback. Merge is a lineage event, not destructive overwrite.
AJ
Accumulator split: detect multimodality (5 eligibility conditions), distribute evidence proportionally to child accumulators, preserve total evidence mass.
AK
Sensor-vocabulary migration: merge/split when sensors are added, re-quantized, or upgraded. Prior trust carried only where evidence justifies it.
Claims AL–AO

Ext 7: Behavioral Causation Trail

Runtime causation packets tracing the exact chain from sensor to behavior. Privacy-preserving: respects privacy mode. Tamper-evident hash chain for audit. Counterfactual explanations via shadow simulator.

AL
Structured causation packet: 12+ fields per salient event. Active context, C_inst, C_ctx, donors, bridge state, suppression, routines, conflict, envelope, lineage, privacy flag.
AM
Differential explanation: combine causation packet with shadow simulator. Modify one factor, re-evaluate. Grounded 'what would have changed' answers.
AN
Privacy-preserving causation: during privacy mode, packet records metadata only (that privacy occurred, behavioral effect). No private content stored.
AO
Tamper-evident audit: packets linked by cryptographic hash. Human review interface identifies which architectural artifact caused a behavior.
Claims AP–AS

Ext 8: Sponsored Relational Bridging

Trusted entity scaffolds entry into new context. Bounded bridge: b₀ = min(b_max, α × q × C_S × (1−C_T)). Decays 10%/day without sponsor. Promoted only after direct experience. Multi-sponsor quorum for high-stakes.

AP
Temporary sponsor bridge: bounded function of sponsor familiarity, target unfamiliarity, and sponsorship confidence. Reduces withdrawal without granting full familiarity.
AQ
Bridge under manifold constraints: transient transfer structure composed with mixing matrix and projected onto Birkhoff polytope. Non-amplifying.
AR
Confirmation, promotion, decay, revocation: bridge decays absent positive interaction. Promoted to persistent relation only after direct experience. Revoked on contradictory evidence.
AS
Multi-sponsor quorum: high-stakes transitions require 2+ sponsors. Each individually bounded. Convergent trust from multiple independent anchors.

Licensing enquiries

CCF is released under BSL-1.1 — free for evaluation and non-commercial use. Commercial licensing is available. Contact us to discuss integration, white-labelling, or custom deployment agreements.

cbyrne@floutlabs.com
API docs: docs.rs/ccf-core