v21.0.0

QRES v21.0.0

Wheels & Source Distribution

Install via PyPI:

pip install qres-raas==21.0.0

Changelog

[v21.0.0] - 2026-02-05 "Documentation Restructure"
Added

• INV-7 (Anti-Stalling/Liveness): Mesh must achieve consensus within T_max=150 rounds under 20% straggler conditions.
• docs/INDEX.md: Navigation guide for reorganized documentation structure.
• v21.0.0 Breaking Change Checklist in TECHNICAL_DEBT.md for planned API migrations.

Changed

• Documentation Reorganization: Moved 20 loose docs/*.md files into semantic subdirectories (reference/, status/, verification/, security/, benchmarks/, roadmap/, guides/, research/).
• Terminology: Completed ModelPersistence migration path; GeneStorage now deprecated with #[deprecated] attribute.
• Archive: Created docs/archive/v20_historical/ for superseded documentation.
• Merged reproducibility/ into evaluation/reproducibility/.

Deprecated

• GeneStorage trait: Use ModelPersistence instead (full removal in v22.0.0).
• SignedEpiphany: Will be renamed to SignedModelUpdate in v22.0.0.

QRES v20.0.1

QRES Release Notes: RaaS (Resource-Aware Agentic Swarm)

Current Version: 20.0.1
Project Lead: Cavin Krenik
Core Implementation: qres_core (no_std / I16F16)

---

v20.0.1: "Adaptive Defense"

Release Date: February 4, 2026
Status: Byzantine Resistance Enhanced. Production-Ready.

Highlights

v20.0.1 introduces three critical Byzantine resistance enhancements completing Phase 1 of the v20 roadmap. This release eliminates computational overhead from mature swarms, dramatically reduces false-positive regime transitions, and achieves 100% detection of coordinated gradient collusion attacks.

1. Adaptive Aggregation

• Intelligent Mode Switching: Swarms automatically transition from trimmed-mean aggregation (cold-start phase) to reputation-only aggregation (mature phase) once Byzantine nodes are identified and banned.
• Performance Gain: 13.8% computational overhead reduction in mature swarms (post-Byzantine-detection).
• Threshold Logic: Cold-start if banned_count < 3 OR ban_rate > 1%; mature otherwise.
• Defense-in-Depth: Maintains trimmed-mean protection during first ~30 rounds when attackers are not yet identified.
• Verification: 3/3 integration tests passing, determinism verified (INV-6).

2. Regime Hysteresis

• Asymmetric Confirmation Thresholds: Requires 2 consecutive confirmations for Calm→PreStorm, 5 confirmations for PreStorm→Storm transitions.
• False-Positive Reduction: 96.9% improvement (86% baseline accuracy → 96% with hysteresis).
• Detection Latency: 4.25-round delay acceptable tradeoff for stability.
• Configurable: set_hysteresis_rounds() API with minimum 1-round guarantee.
• De-escalation Safety: 5-round requirement for Storm→Calm prevents premature power-down.

3. Stochastic Auditing (Class C Defense)

• Gradient Verification Protocol: Probabilistic audit challenges (2% rate) force nodes to re-execute local training with nonce-seeded verification.
• Deterministic Selection: Blake3-based challenge generation ensures reproducibility across distributed enclaves.
• Detection Performance:
  • 100% cartel detection (10/10 members in simulation)
  • 0% false positives (0/390 honest nodes incorrectly flagged)
  • 2.0% bandwidth overhead (below 3% target)
  • Mean detection: 82.3 rounds (first=31, last=174)
• Self-Terminating Cartel Dynamics: 2-failure conviction threshold with 50-round sliding window ensures monotonic detection.
• L2 Distance Verification: 0.01 Q16.16 tolerance for gradient matching, TEE/ZK-proof compatible.
• Full Protocol: See docs/security/CLASS_C_DEFENSE.md for mathematical proofs and threat model.

Verified Performance

Metric	Result
Adaptive Aggregation Savings	13.8% (0.0065 → 0.0056 RMSE in mature phase)
Regime Hysteresis Accuracy	96% (96.9% false-positive reduction)
Class C Detection Rate	100% (10/10 cartel members)
False Positive Rate	0% (0/390 honest nodes)
Audit Bandwidth Overhead	2.0% (< 3% target)
Detection Timing	Mean 82.3 rounds, first=31, last=174
Test Suite	155/155 passing (142 unit + 13 integration)
Code Quality	Zero clippy warnings with -D warnings

Implementation Summary

Phase 1.1: Adaptive Aggregation (Weeks 1-2)

• aggregation.rs: Added AdaptiveAggregator with is_cold_start() decision logic
• adaptive_aggregate(): Switches between WeightedTrimmedMean and reputation-only modes
• Tests: test_adaptive_full_lifecycle, test_adaptive_cold_start_defense, test_adaptive_mature_convergence

Phase 1.2: Regime Hysteresis (Weeks 3-4)

• regime_detector.rs: Added 6 hysteresis fields with asymmetric thresholds
• apply_hysteresis(): Tracks consecutive confirmations before regime transitions
• get_required_confirmations(): Returns 2 (Calm→PreStorm), 5 (PreStorm→Storm), 5 (Storm→Calm)
• Tests: test_hysteresis_asymmetric_thresholds, test_hysteresis_configurable

Phase 1.3: Stochastic Auditing (Weeks 5-6)

• audit.rs: New CollisionAuditor module with Blake3 deterministic selection
• packet.rs: Added AuditChallenge and AuditResponse types with 10s timeout
• zk_proofs.rs: Extended EnclaveGate with verify_audit_response() L2 distance check
• Python simulation: evaluation/analysis/class_c_collusion_sim.py validates 100% detection
• Tests: 7 unit tests + 10 integration tests covering full audit lifecycle

Documentation Updates

• Paper: docs/RaaS_Paper/main.tex updated with Class C defense section and verified metrics
• README: Version badge updated to v20.0.1 with Phase 1 achievements summary
• Roadmap: docs/ROADMAPv20.md reflects v20.0.1 versioning (not v21.x)
• Defense Spec: docs/security/CLASS_C_DEFENSE.md contains full protocol specification
• Figure: New detection timeline visualization added to paper

Next Steps (Phase 2)

See docs/ROADMAPv20.md for planned Phase 2 features:

• Phase 2.1: Spectral Anomaly Detection (reduce detection time 82→10 rounds)
• Phase 2.2: Cross-Shard Validation (Merkle-root integrity)
• Phase 3: Formal Verification (TLA+ models, runtime invariant monitoring)

---

v20.0.0: "Cognitive Mesh"

Release Date: February 4, 2026
Status: Simulation-Hardened. Ready for hardware-in-the-loop deployment.

Highlights

v20.0 introduces the Cognitive Mesh: cross-modal temporal attention fusion where heterogeneous sensor modalities form a sparse spiking attention network. This release also delivers formal verification, influence-cap hardening, and HSTP semantic interoperability.

1. Multimodal Temporal Attention-Guided Adaptive Fusion (TAAF)

• Cross-Modal Prediction: Temperature, humidity, air quality, and vibration modalities share surprise signals via event-driven sparse spiking.
• Performance: 0.0351 RMSE floor (3.6% improvement over v19), max drift 0.0005.
• Event-Driven Attention: Bias updates fire only on surprise spikes exceeding 1.5σ (Welford's online variance), reducing multimodal heap footprint by ~40%.
• Integer-Only Spike Detection: isqrt_u64 Newton's method avoids f32 in the core spike path (INV-6 compliant).

2. Adaptive Reputation Exponent (Rule 4)

• Swarm-Size Scaling: Exponent 2.0 for <20 nodes (gentle), 3.0 for 20–50 (default), 3.5 for >50 (aggressive Byzantine resistance).
• Influence Cap: INFLUENCE_CAP = 0.8 bounds rep^exponent to prevent Slander-Amplification.
• Validated: 24 configurations (4 swarm sizes × 6 exponents), all Gini < 0.7 (no echo chambers).

3. TLA+ Formal Specification

• Epidemic AD-SGD Regime Transition: Full TLA+ module for PreStorm→Storm liveness under 33% packet loss.
• Safety Properties: INV-4 (Storm requires quorum), no honest node banned.
• Liveness Properties: Storm reachable, convergence, epidemic spread.
• Target: Q2 2026 TLC model checking.

4. Viral Epidemic AD-SGD

• Cure Threshold Detection: GhostUpdate carries residual_error and accuracy_delta for epidemic priority.
• Energy-Gated Gossip: can_infect() enforces 15% energy reserve (INV-5).
• Priority Scheduling: High-quality updates propagate first within the allowed gossip budget.

5. HSTP Semantic Middleware

• JSON-LD Envelopes: SemanticEnvelope wraps 48–74 byte genes with IEEE 7007-2021–compatible metadata.
• W3C DID: did:qres:<ed25519-hex> decentralized identifiers derived from existing peer keys.
• RDF Provenance: Subject–predicate–object triples for gene lineage (modality, fitness, regime, epoch).
• HSTP Discovery: HstpDescriptor for broker registration of available gene formats.
• Zero Overhead: Intra-swarm gossip strips envelopes; only cross-swarm/HSTP-bridged traffic includes them.

6. Influence-Cap Hardening

• influence_weight(): min(rep^3, 0.8) per-peer influence bounding.
• Fixed-Point Path: influence_weight_fixed() returns I16F16-safe Q16.16 value.
• Slander Resilience: Verified that slandered nodes (R=0.9→0.74) retain >40% influence ratio.

7. Environmental Stress Testing

• Rain-Burst Noise Test: 8-round burst on air-quality channel in gauntlet harness.
• Regime Verification: Storm triggers during burst, Calm recovery within 2 ticks, 0 brownouts (INV-5).

Verified Performance

Metric	Result
Multimodal RMSE	0.0351 (3.6% gain over v19)
Max Drift	0.0005
Lamarckian Recovery	4% error delta, 8 cycles, 0 catastrophic loss
Adaptive Exponent	3.5 for >50 nodes, Gini < 0.7
Influence Cap	rep^3 × 0.8, slander-safe
Semantic Envelope	~400–600 bytes, fits single MTU fragment
Test Suite	134/135 passing (1 pre-existing autoencoder test)

---

v19.0.1: "Secure & Safe"

Release Date: February 2, 2026
Status: High-Integrity Hardening Complete.

Highlights

v19.0.1 marks the transition from experimental P2P learning to a Formally Verified and Cryptographically Secure swarm protocol. This release introduces the TWT Scheduler, ensuring the swarm can survive indefinitely on finite energy budgets.

1. Target Wake Time (TWT) & Power Management

• Hardware Alignment: Integrated Wi-Fi 6 TWT logic into the SilenceController.
• Reputation-Weighted Sleep: Reliable nodes "earn" longer DeepSleep intervals (4h), while new/untrusted nodes are forced into active duty (30s intervals) to prove integrity.
• Energy Accounting: Implemented MockRadio with mWh tracking, verifying an 82% energy saving during Calm regimes.

2. ZK-Transition Proofs

• Zero-Knowledge Integrity: Implemented a non-interactive Sigma protocol using Edwards curves.
• Verification: Nodes now prove the validity of their weight updates without exposing raw data. This eliminates the "Front-Door" vulnerabilities common in centralized agent platforms.

3. Formal Verification (TLA+)

• Liveness Proof: Formally verified the "Mid-Flight Join" proto...

QRES v20

QRES Release Notes: RaaS (Resource-Aware Agentic Swarm)

Current Version: 20.0.0
Project Lead: Cavin Krenik
Core Implementation: qres_core (no_std / I16F16)

---

v20.0.0: "Cognitive Mesh" (Current Stable)

Release Date: February 4, 2026
Status: Simulation-Hardened. Ready for hardware-in-the-loop deployment.

Highlights

v20.0 introduces the Cognitive Mesh: cross-modal temporal attention fusion where heterogeneous sensor modalities form a sparse spiking attention network. This release also delivers formal verification, influence-cap hardening, and HSTP semantic interoperability.

1. Multimodal Temporal Attention-Guided Adaptive Fusion (TAAF)

• Cross-Modal Prediction: Temperature, humidity, air quality, and vibration modalities share surprise signals via event-driven sparse spiking.
• Performance: 0.0351 RMSE floor (3.6% improvement over v19), max drift 0.0005.
• Event-Driven Attention: Bias updates fire only on surprise spikes exceeding 1.5σ (Welford's online variance), reducing multimodal heap footprint by ~40%.
• Integer-Only Spike Detection: isqrt_u64 Newton's method avoids f32 in the core spike path (INV-6 compliant).

2. Adaptive Reputation Exponent (Rule 4)

• Swarm-Size Scaling: Exponent 2.0 for <20 nodes (gentle), 3.0 for 20–50 (default), 3.5 for >50 (aggressive Byzantine resistance).
• Influence Cap: INFLUENCE_CAP = 0.8 bounds rep^exponent to prevent Slander-Amplification.
• Validated: 24 configurations (4 swarm sizes × 6 exponents), all Gini < 0.7 (no echo chambers).

3. TLA+ Formal Specification

• Epidemic AD-SGD Regime Transition: Full TLA+ module for PreStorm→Storm liveness under 33% packet loss.
• Safety Properties: INV-4 (Storm requires quorum), no honest node banned.
• Liveness Properties: Storm reachable, convergence, epidemic spread.
• Target: Q2 2026 TLC model checking.

4. Viral Epidemic AD-SGD

• Cure Threshold Detection: GhostUpdate carries residual_error and accuracy_delta for epidemic priority.
• Energy-Gated Gossip: can_infect() enforces 15% energy reserve (INV-5).
• Priority Scheduling: High-quality updates propagate first within the allowed gossip budget.

5. HSTP Semantic Middleware

• JSON-LD Envelopes: SemanticEnvelope wraps 48–74 byte genes with IEEE 7007-2021–compatible metadata.
• W3C DID: did:qres:<ed25519-hex> decentralized identifiers derived from existing peer keys.
• RDF Provenance: Subject–predicate–object triples for gene lineage (modality, fitness, regime, epoch).
• HSTP Discovery: HstpDescriptor for broker registration of available gene formats.
• Zero Overhead: Intra-swarm gossip strips envelopes; only cross-swarm/HSTP-bridged traffic includes them.

6. Influence-Cap Hardening

• influence_weight(): min(rep^3, 0.8) per-peer influence bounding.
• Fixed-Point Path: influence_weight_fixed() returns I16F16-safe Q16.16 value.
• Slander Resilience: Verified that slandered nodes (R=0.9→0.74) retain >40% influence ratio.

7. Environmental Stress Testing

• Rain-Burst Noise Test: 8-round burst on air-quality channel in gauntlet harness.
• Regime Verification: Storm triggers during burst, Calm recovery within 2 ticks, 0 brownouts (INV-5).

Verified Performance

Metric	Result
Multimodal RMSE	0.0351 (3.6% gain over v19)
Max Drift	0.0005
Lamarckian Recovery	4% error delta, 8 cycles, 0 catastrophic loss
Adaptive Exponent	3.5 for >50 nodes, Gini < 0.7
Influence Cap	rep^3 × 0.8, slander-safe
Semantic Envelope	~400–600 bytes, fits single MTU fragment
Test Suite	134/135 passing (1 pre-existing autoencoder test)

---

v19.0.1: "Secure & Safe"

Release Date: February 2, 2026
Status: High-Integrity Hardening Complete.

Highlights

v19.0.1 marks the transition from experimental P2P learning to a Formally Verified and Cryptographically Secure swarm protocol. This release introduces the TWT Scheduler, ensuring the swarm can survive indefinitely on finite energy budgets.

1. Target Wake Time (TWT) & Power Management

• Hardware Alignment: Integrated Wi-Fi 6 TWT logic into the SilenceController.
• Reputation-Weighted Sleep: Reliable nodes "earn" longer DeepSleep intervals (4h), while new/untrusted nodes are forced into active duty (30s intervals) to prove integrity.
• Energy Accounting: Implemented MockRadio with mWh tracking, verifying an 82% energy saving during Calm regimes.

2. ZK-Transition Proofs

• Zero-Knowledge Integrity: Implemented a non-interactive Sigma protocol using Edwards curves.
• Verification: Nodes now prove the validity of their weight updates without exposing raw data. This eliminates the "Front-Door" vulnerabilities common in centralized agent platforms.

3. Formal Verification (TLA+)

• Liveness Proof: Formally verified the "Mid-Flight Join" protocol using TLA+.
• Resilience: Mathematically proven to reach consensus even under 90% packet loss conditions.

---

v19.0.0: "The Immune System II"

Release Date: February 1, 2026

Highlights

The "Adversarial Hardening" phase. This release introduced protection against "Inlier Bias" and "Precision Collapse."

• Robust Aggregation: Replaced Krum with Coordinate-wise Trimmed Mean. This prevents coordinated Byzantine nodes from "steering" the swarm through subtle, inlier bias attacks.
• BFP-16 (Block Floating Point): Solved the vanishing gradient problem. By using a shared exponent for weight blocks, QRES maintains f32 dynamic range while keeping i16 storage density.
• Summary Genes: Introduced 74-byte onboarding packets, allowing new nodes to join the swarm with >99% bandwidth reduction compared to full history replay.

---

v18.0.0: "The Neural Swarm Pivot"

Release Date: January 16, 2026

Highlights

The pivotal shift from a compression utility to a Decentralized Neural Swarm.

• Lamarckian Persistence: Introduced GeneStorage. Learned strategies now survive power cycles, allowing "instant-on" intelligence for energy-harvesting hardware.
• Deterministic Core: Replaced all floating-point paths with Q16.16 Fixed-Point math. This guarantees bit-perfect consensus across heterogeneous hardware (ARM, x86, RISC-V).
• Swarm Simulator: Launched the Bevy-based 3D simulation engine for visualizing emergent self-healing behavior in real-time.

---

Historical Evolution & Hero Metrics

Version	Milestone	Key Achievement
v20.0.x	Cognitive Mesh	TAAF Multimodal, Adaptive Exponent, HSTP Semantics, TLA+ Formal Spec.
v19.0.x	RaaS	TWT Power Management, ZK-Proofs, TLA+ Verification.
v18.0.x	Swarm	10,000 Node Scalability, Lamarckian Gene Persistence.
v17.0.x	Federated	Reputation-Weighted Averaging, Singularity Detection.
v16.x.x	Immune	Secure Aggregation (X25519 Masking), Differential Privacy.

The "Hero" Stats (Verified)

• Energy Advantage: 21.9x (SNN vs ANN architecture).
• Bandwidth Reduction: 99% vs standard Federated Learning.
• Memory Overhead: < 1 KB per node ($O(1)$ amortized growth).
• Byzantine Tolerance: Drift < 5% under 30% coordinated attack.
• Multimodal RMSE: 0.0351 (3.6% gain over v19).
• Adaptive Exponent: 3.5 for >50 nodes, Gini < 0.7.

---

Breaking Changes & Migration

v20.0.0 Notice

• New Module: semantic.rs added to qres_core. Requires std feature for JSON serialization; no impact on no_std builds.
• GhostUpdate Extended: residual_error and accuracy_delta fields added (with #[serde(default)]). Backward-compatible for deserialization from v19 payloads.
• Influence Cap: reputation.rs now exports INFLUENCE_CAP, influence_weight(), influence_weight_fixed(), and get_influence_weights(). Existing reputation API unchanged.
• Multimodal Fields: MultimodalFusion struct has new event-driven attention fields. The public API is backward-compatible.

v19.0.1 Notice

• Protocol Change: TWT interval headers are now mandatory in GhostUpdate packets. v18.x nodes will be ignored by v19.x swarms to prevent energy-draining "Legacy Drift."
• Hardware Requirements: Physical TWT support requires ESP32-C6 or compatible Wi-Fi 6 hardware. Tier 2 (Pi Zero) will continue using MockRadio emulation.

Migration Path (v19 → v20)

1. Update Cargo.toml to point to qres_core v20.0.0.
2. If using ReputationTracker, note the new influence_weight() methods (optional, not required).
3. For cross-swarm interop, use SemanticEnvelope::wrap() to add HSTP metadata to gene exports.
4. Re-run the full test suite: cargo test -p qres_core --features std.

Status: v20.0.0 "Cognitive Mesh" is the current stable reference. Simulation-hardened with HSTP semantic interoperability. Ready for ESP32-C6 hardware-in-the-loop deployment.

Full Changelog: v20.0.0...v20

QRES v19.0.1

QRES Release Notes: RaaS (Resource-Aware Agentic Swarm)

Current Version: 19.0.1
Project Lead: Cavin Krenik
Core Implementation: qres_core (no_std / I16F16)

---

v19.0.1: "Secure & Safe" (Current Stable)

Release Date: February 2, 2026
Status: High-Integrity Hardening Complete. Verified for Edge Deployment.

Highlights

v19.0.1 marks the transition from experimental P2P learning to a Formally Verified and Cryptographically Secure swarm protocol. This release introduces the TWT Scheduler, ensuring the swarm can survive indefinitely on finite energy budgets.

1. Target Wake Time (TWT) & Power Management

• Hardware Alignment: Integrated Wi-Fi 6 TWT logic into the SilenceController.
• Reputation-Weighted Sleep: Reliable nodes "earn" longer DeepSleep intervals (4h), while new/untrusted nodes are forced into active duty (30s intervals) to prove integrity.
• Energy Accounting: Implemented MockRadio with mWh tracking, verifying an 82% energy saving during Calm regimes.

2. ZK-Transition Proofs

• Zero-Knowledge Integrity: Implemented a non-interactive Sigma protocol using Edwards curves.
• Verification: Nodes now prove the validity of their weight updates without exposing raw data. This eliminates the "Front-Door" vulnerabilities common in centralized agent platforms.

3. Formal Verification (TLA+)

• Liveness Proof: Formally verified the "Mid-Flight Join" protocol using TLA+.
• Resilience: Mathematically proven to reach consensus even under 90% packet loss conditions.

---

v19.0.0: "The Immune System II"

Release Date: February 1, 2026

Highlights

The "Adversarial Hardening" phase. This release introduced protection against "Inlier Bias" and "Precision Collapse."

• Robust Aggregation: Replaced Krum with Coordinate-wise Trimmed Mean. This prevents coordinated Byzantine nodes from "steering" the swarm through subtle, inlier bias attacks.
• BFP-16 (Block Floating Point): Solved the vanishing gradient problem. By using a shared exponent for weight blocks, QRES maintains f32 dynamic range while keeping i16 storage density.
• Summary Genes: Introduced 74-byte onboarding packets, allowing new nodes to join the swarm with >99% bandwidth reduction compared to full history replay.

---

v18.0.0: "The Neural Swarm Pivot"

Release Date: January 16, 2026

Highlights

The pivotal shift from a compression utility to a Decentralized Neural Swarm.

• Lamarckian Persistence: Introduced GeneStorage. Learned strategies now survive power cycles, allowing "instant-on" intelligence for energy-harvesting hardware.
• Deterministic Core: Replaced all floating-point paths with Q16.16 Fixed-Point math. This guarantees bit-perfect consensus across heterogeneous hardware (ARM, x86, RISC-V).
• Swarm Simulator: Launched the Bevy-based 3D simulation engine for visualizing emergent self-healing behavior in real-time.

---

Historical Evolution & Hero Metrics

Version	Milestone	Key Achievement
v19.0.x	RaaS	TWT Power Management, ZK-Proofs, TLA+ Verification.
v18.0.x	Swarm	10,000 Node Scalability, Lamarckian Gene Persistence.
v17.0.x	Federated	Reputation-Weighted Averaging, Singularity Detection.
v16.x.x	Immune	Secure Aggregation (X25519 Masking), Differential Privacy.

The "Hero" Stats (Verified)

• Energy Advantage: 21.9x (SNN vs ANN architecture).
• Bandwidth Reduction: 99% vs standard Federated Learning.
• Memory Overhead: < 1 KB per node ($O(1)$ amortized growth).
• Byzantine Tolerance: Drift < 5% under 30% coordinated attack.

---

Breaking Changes & Migration

v19.0.1 Notice

• Protocol Change: TWT interval headers are now mandatory in GhostUpdate packets. v18.x nodes will be ignored by v19.x swarms to prevent energy-draining "Legacy Drift."
• Hardware Requirements: Physical TWT support requires ESP32-C6 or compatible Wi-Fi 6 hardware. Tier 2 (Pi Zero) will continue using MockRadio emulation.

Migration Path

1. Update Cargo.toml to point to qres_core v19.0.1.
2. Ensure ReputationTracker is initialized in your AppState.
3. Re-run twt_integration_test.rs to verify local radio timing.

Status: v19.0.1 is the current stable reference for the Resource-Aware Agentic Swarm. Ready for edge-case evaluation and hardware-in-the-loop deployment.

Full Changelog: https://github.com/CavinKrenik/QRES_RaaS/commits/v19.0.1