πŸ€– Models & Intelligence Stack

45+ custom algorithms Β· 207K lines Β· 85% original IP Β· Forward/backward propagation from scratch Β· Self-improving nightly

πŸ† Custom ML Algorithms (100% Our IP)

Every model hand-designed, hand-coded, trained on our data. Forward + backward propagation implemented from scratch where indicated.

ModelBackpropAlgorithm
CIFOG GCN Manual NumPy3-layer Graph Convolution (286β†’128β†’64β†’1). Forward: Z=relu(AΒ·XΒ·W). Backward: MSE gradient β†’ sigmoid deriv β†’ ReLU mask β†’ weight update. 800 epochs, lr=0.003. No PyG/DGL.
Quorum Voting v2 Manual NumPy2-layer neural net (10β†’8β†’3). Forward: ReLU β†’ softmax. Backward: cross-entropy gradient β†’ ReLU mask β†’ weight deltas. 300 epochs, lr=0.05. No sklearn.
FusionWeightLearner Manual NumPyL2-regularized logistic regression. Forward: Οƒ(XΒ·w+b). Backward: dw = X.TΒ·(predβˆ’y)/n + λ·w. 500 epochs, lr=0.1, Ξ»=0.01. 7 dims β†’ scoring weights.
Cortex Master v7 PyTorch Autograd53-dim β†’ 3-layer FFN (128β†’64β†’32) β†’ dual head: Trust (sigmoid, 3-dim) + Strategy (5-class softmax: ADVANCE/HOLD/REJECT/LEARN/QUANTUM). 18 brain signals + 6 teacher + capsule + molecular + target class + NPEAD.
CortexQ Gen v4 PyTorch Autograd6-layer Transformer Encoder (GPT), 8 heads, 256d, 512 FFN. SELFIES tokenizer. Conditioned: pathway(11) + class(8) + pocket(4). 3.2M params, 10,258 training molecules.
MolTransformer v3 PyTorch Autograd4-layer Transformer Decoder, 4 heads, 256d. Cross-attention to 16-dim condition (class + NPEAD + IC50). REINFORCE: reward=βˆ’Vina, token log-prob Γ— advantage, Adam lr=1e-5, grad clip 1.0.
BindNet v1 Sklearn + CustomRandomForestRegressor. Custom features: Morgan FP(r=2, 1024 bits) + MW + QED + LE = 1027 inputs β†’ predicted Vina. Trained on our docking data.
Pocket Encoder PyTorch Autograd3D CNN: Conv3d(4β†’32β†’64β†’128) + BN + MaxPool + AdaptiveAvgPool β†’ FC(1024β†’256β†’64). 24Β³ voxel grid, 4 channels (hydrophobic/polar/charged/aromatic), 1Γ… resolution.
Diffusion Generator PyTorch AutogradConditional denoiser: MLP(64+36+1→128, SiLU→128, SiLU→64). Decoder: MLP(64→128→25×4) for 3D coords. Pocket-conditioned, linear noise schedule, T-step denoising.
NPEAD-DLE v6.1 Custom FrameworkDirected Learning Engine. 28-dim scorer + EIV action selector + 7-dim uncertainty map + counterfactual simulator + meta-learning policy (51 outcomes) + CortexQ validation. Self-improving: learns per-target-class action success rates. 89% accuracy, 63% rescue rate.
RL Docking Trainer PyTorch + REINFORCEGenerate 50 β†’ Dock (Vina) β†’ Reward=βˆ’Vinaβˆ’baseline β†’ Token log-prob Γ— advantage β†’ Adam update. Top-10 update weights. Runs daily 4am on GPU Batch.
Authority Decay Custom FormulaExponential temporal decay: auth *= e^(βˆ’0.01 Γ— days). Floor 0.2. Triggers re-validation after 7 days stale. Nightly after GCN.
NPEAD-MAML Custom Meta-LearnModel-Agnostic Meta-Learning. 67 input dims, 12 tasks, 111 examples. k=1/k=5 few-shot adaptation. AUC 0.982. Adapts to novel target classes in 1 gradient step.
● Manual backprop (NumPy) Β· ● PyTorch (custom arch) Β· ● Sklearn (custom data) Β· ● Custom framework

πŸŽ“ Teacher-Student Knowledge Distillation

Closed-loop: Teacher evaluates β†’ Students update weights β†’ Generators adapt β†’ Outcomes feed back.

πŸ§‘β€πŸ« TEACHER
Claude Opus/Sonnet
Evaluates brain quality, validates scoring, critiques decisions
β†’
πŸ“ DISTILLATION
Phase 14 Nightly
Aggregates by target_class β†’ feature separators β†’ weight adaptations
β†’
πŸŽ“ STUDENTS
All Generation Models
Receive DISTILLED_WEIGHTS β†’ adapt thresholds β†’ adjust explore/exploit
StageTeacher ActionStudent LearnsAlgorithm
Phase 9: TeacherClaude evaluates 20 brains β†’ quality scoreMaster dims 19-24 (teacher_quality[6]) updateLLM β†’ structured score β†’ DDB β†’ input features
Phase 14: DistillNPEAD collects ADVANCE vs REJECT decisionsGEN V3 reads DISTILLED_WEIGHTS per target_classFeature importance β†’ threshold β†’ generation constraints
RL (4am daily)Vina scores = reward signalMolTransformer v3 weights via REINFORCEβˆ’Vina β†’ log_prob Γ— advantage β†’ Adam step
Phase 3: LearnOutcomes (IC50, grades) = labelsFusionWeightLearner β†’ new scoring vectorGradient descent β†’ w β†’ NPEAD + Rust scoring
Phase 4: GCNHero grades = authority labelsCIFOG GCN retrains (286β†’128β†’64β†’1)Manual backprop (MSE) β†’ W1,W2,W3 updated
Phase 13: SELFCortex-SELF diagnoses failuresGateOptimizer adjusts thresholdsFailure clusters β†’ proposals β†’ replay β†’ promote/retire

Key: Teacher never directly modifies weights. It produces quality signals that become input features or labels. Students learn via their own gradient algorithms.

🧠 Core Intelligence Stack

Cortex Master v7 53-dim

3-layer FFN (53β†’128β†’64β†’32), ReLU. Dual head: Trust (sigmoid) + Strategy (5-class softmax). Input: 18 brain + 6 teacher + 1 capsule + 7 molecular + 5 class + 4 meta + 12 NPEAD. NumPy inference.

NPEAD-DLE v6.1 Directed Learning

Evidence-directed action selector. EIV formula (gain Γ— impact Γ— reliability Γ· cost). 7-dim uncertainty map. Counterfactual simulator (blocks worthless actions). Meta-learning policy (learns per target_class). CortexQ validation layer (6 override rules). Immediate learning after every batch job. Auto-kills TSGs + undruggable. 89% accuracy, 63% rescue rate.

CortexQ Gen v4 3.2M params

6-layer Transformer (GPT), 8 heads, 256d, 512 FFN. SELFIES tokens. Conditioned: pathway + class + pocket. Trained 40 epochs on 10,258 molecules. 100% validity.

MolTransformer v3 Conditioned GPT

4-layer Transformer Decoder, 4 heads, 256d. Cross-attention to 16-dim condition. Pretrained 2M+ ChEMBL. Fine-tuned via REINFORCE daily.

Cortex-GEN V3 Production

Scaffold enumeration: 10 scaffolds Γ— 8 transforms + pocket-aware routing. Reads: playbook + DISTILLED_WEIGHTS + failure memory + capsules.

Fast Path V3

Generate β†’ BindNet β†’ Master decides. No GPU for triage. Only ADVANCE gets docked. 90% less GPU waste.

Cortex-SELF Production

Diagnose failures β†’ propose rules β†’ counterfactual replay β†’ promote/retire. Writes improvement proposals to evidence chain.

BWM Production

Biological World Model. Causal graph: targets, pathways, resistance. Adjusts NPEAD strategy per target class from edges.

Knowledge Capsules 38 Live

15 brains β†’ 38 capsules (10 types Γ— 5 cancers). Each = decision-ready claim + confidence. Feeds CortexQ-Gen + GEN V3.

NPEAD Distillation + DLE v6.1

Exports scoring decisions β†’ feature separators per target_class β†’ DISTILLED_WEIGHTS to generators. DLE immediate learning: policy updates after every batch job. Standardized action registry (DynamoDB shared config). CortexQ validates all DLE recommendations.

🧬 20 Brain Signals

#BrainTypeInputOutput
1Cancer BrainKnowledgePubMed binding/structurePocket druggability
2GenomicsOmicsTCGA mutations, CNVDriver/passenger
3TranscriptomicsOmicsExpression z-scoresOverexpression signal
4ProteomicsOmicsUniProt, abundanceTarget accessibility
5PathwayOmicsKEGG/ReactomeCentrality, hallmarks
6GCN (CIFOG)Intelligence276-node graphDRIVER confidence
7ChemistryKnowledgeDrug-likeness rulesScaffold quality
8ClinicalKnowledgeSurvival, drug responseClinical relevance
9BiomarkersKnowledgeCDx panelsDiagnostic signal
10PubMedIntelligenceLiterature (weekly)Evidence validation
11EvidenceIntelligenceHash chain integrityProvenance confidence
12DepMapIntelligenceCRISPR essentialityKnockout dependency
13CuriosityAdvisoryPubMed emergingGap detection
14CompetitorIntelligenceClinicalTrials.govPharma validation
15QuantumIntelligencearXiv quantum+drugQuantum relevance
16SafetyKnowledgeToxicity papersADR/tox signal
17PharmacologyKnowledgeADMET/PK papersBioavailability
18BioBrainIntelligencePDB structuresPocket profiles
19TeacherIntelligenceAll brain signalsQuality corrections
20EGAO FuserOrchestratorAll signalsFused decision

🧬 Tumor Route Intelligence

Tumor Route Graph V1

State-transition graph. Nodes = tumor states. Edges = transitions with blocking genes. Extends CIFOG with escape routes.

Cancer-State Reversal V1

5-dim score: expression(0.25) + pathway(0.25) + dependency(0.20) + safety(0.15) + coverage(0.15). Quantifies malignant→normal reversal.

Bottleneck Detection V1

Betweenness centrality on route graph. Finds genes critical across ALL escape routes.

Pathway Forecasting V1

Predicts next tumor state given intervention. Where does resistance emerge after N steps?

QTEL Landscape Quantum

Full state-transition landscape with probabilities. Evolutionary pressure mapping.

QEBD QAOA

Quantum Bottleneck Search. Minimal gene set blocking most escape routes. QAOA for large graphs.

βš—οΈ Generation & Docking

Cortex-GEN V3 750 lines

15 scaffolds Γ— 12 transforms + SELFIES mutations. Pocket-aware routing. Reads capsules + distilled weights.

Pocket-Conditioned Diffusion 3D

Denoising diffusion. MLP denoiser (64+36+1β†’128β†’64) + decoder (64β†’128β†’25Γ—4). Pocket embedding conditions generation.

Pocket Voxel Encoder 3D CNN

Conv3d(4β†’32β†’64β†’128) + BN + AdaptiveAvgPool β†’ FC(256β†’64). 24Β³ grid, 4 channels, 1Γ…. PDBbind trained.

BindNet v1 RF Β· 1027 feat

RandomForest. Morgan FP(1024) + MW + QED + LE β†’ Vina prediction. Pre-docking filter.

Uni-Dock GPU Parallel Array

GPU-accelerated docking. 3x parallel spot instances per target (AWS Batch array jobs). Each shard generates/docks independently with different seed β†’ 3x faster + 3x scaffold diversity. g4dn.xlarge spot, 5 regions.

ADMET Composite 5-model

bioavail(0.30) + hERG(0.25) + stability(0.20) + PAINS(0.15) + aggregator(0.10). Lipinski/Veber.

Born Machine 4-qubit

Ry+CNOT parameterized circuit (Braket SV1). Target-class biased scaffold generation.

V2 Shadow Gen A/B Test

Read-only comparison pipeline. Design Intent + iterative tightening (3Γ—100). Never touches production.

βš›οΈ Quantum Circuits

VQE Binding SV1 2-4q

UCCSD ansatz. Corrects classical docking (44% on S-C bonds). Routes when Q-VOI β‰₯ 0.75.

QAOA Combo SV1 10-20q

p=2 QAOA on QUBO. Synergy + diversity βˆ’ resistance. Optimal 3-target combinations.

QUBO Feature Select SV1 7q

Maximizes feature-grade correlation minus redundancy. Identifies predictive NPEAD dims.

Grover Conformer SV1 3-4q

Oracle marks low-energy poses, diffusion amplifies. O(√N) speedup.

πŸ” Nightly Pipeline (15 Phases)

PhaseNameAction
1Gates + PQMPermission check, SQS candidates
2Load signals20 brains β†’ EGAO orchestrator
3LearningOutcomes β†’ FusionWeightLearner + corrections
4CIFOG + GCNGraph neural net trains on 276+ nodes (manual backprop)
5Omics refreshUpdate omics signals on brain-bus
6-7Log + BackupActivity metadata + S3 snapshots
8Display syncVerified genes β†’ pipeline tables
9TeacherClaude evaluates brain quality β†’ teacher_score[6]
10Gene scanRecycle + auto-revival + wet-lab upgrade + orphan catch
11Next actionsResolve pending re-docks, permissions
12Brain feedsPubMed/CT/UniProt/PDB/arXiv β†’ capsules β†’ playbooks β†’ V3 discovery
13Cortex-SELFFailure diagnosis, rule proposals, counterfactual replay
14DistillationScoring examples β†’ feature separators β†’ DISTILLED_WEIGHTS
15QuantumRun VQE/QAOA for quantum-routed heroes

πŸ—οΈ Agents

cortex-rs Rust Β· 9K lines

Nightly orchestrator. 15 phases. NPEAD + BWM + SELF + Quantum.

gpu-batch-rs Rust+Py Β· 7K

GPU docking pipeline. Parallel array jobs (3 instances/target). GEN V3 + V4 neural + Uni-Dock + ADMET + scoring. Multi-region spot (us-east-1/2, us-west-2, ca-central-1, eu-west-1).

physician-rs Rust+Py Β· 3.6K

Discovery + Fast Path. Generate β†’ BindNet β†’ Master β†’ parallel dock winners. Submits array jobs. Auto-blocks TSGs/ORs. 5-region spot optimization.

πŸ”¬ Foundation Models (External)

Geneformer V2 104M

NVIDIA. 14M cancer transcriptomes. Expression patterns, perturbation response.

ESM-2 650M Β· SageMaker

Meta. 250M protein sequences. Structure + function from sequence.

CortexQ Brain LLM

Bedrock Haiku orchestration. Full-state reasoning. Failsafe: rule-based.

πŸ“¦ Model Registry (S3 Β· KMS Encrypted)

ModelKeySizeAlgorithm
Cortex Master v7models/cortex_master_v1.pt200KB3-layer FFN + dual head (Trust sigmoid + Strategy softmax)
MolTransformer v3models/mol_transformer_v3.pt12.9MB4-layer Transformer Decoder + REINFORCE daily
CortexQ Gen v4agents/cortexq_gen/cortexq_gen_model_v4.pt12.5MB6-layer Transformer Encoder (GPT) + SELFIES
BindNet v1models/bindnet_v1.pkl1.1MBRandomForest (1027 features β†’ Vina)
SafetyNet v1models/safetynet_v1.pkl570KBRandomForest (tox/hERG/reactive classifier)
GCN Weightsmodels/gcn_weights_latest.json950KB3-layer GCN manual backprop (nightly retrained)
Pocket Encodermodels/pocket_encoder_v1.pt2MB3D CNN Conv3d(4β†’128) + FC(256β†’64)

Bucket: s3://mindnetq-evidence-ledger-865950114883/models/ Β· All models loaded at runtime (not baked into containers)

πŸ“Š Key Metrics

207K

Total lines of code

187

Heroes in pipeline

45+

Custom algorithms

89%

Back-test accuracy

3.2M

CortexQ Gen params

33nM

BRAF validated hit

πŸ’Ύ Data Architecture

TierTablePurposeRetention
πŸ”₯ HOTcortex-brain-busSignals, capsules, playbooksPermanent
πŸ”₯ HOTcortex-heroesAll targets under evaluationPermanent
⚑ ACTIVEcortex-scan-poolGene evaluations (pass/reject)90-day TTL
❄️ COLDcortex-wisdomCompressed insights from scan poolPermanent