demon hunter labs
Projects18
Every project, one jump awayChoose what you want to build.
Open the full Atlas

Foundation

Creation

Operations

Research

← Yodi manual

What it is / interactive architecture

Yodi: Inspectable intelligence and control built from specialized Yodis

Yodi is a Yeho-native intelligence engine for compact specialists, Neural Boards, memory, routing, generation, multimodal work, and bounded serving. Specialized Yodis can form a typed team and eventually control admitted application or OS capabilities without bypassing permissions, evidence, or the owning system.

CAST / 03

Yodi field guide

Yuna MnemeThe Memory Seer
Open complete manual
Inspectable intelligence and control built from specialized YodisREV 03
100%
SHAPE KEY
Inspectable intelligence and control built from specialized YodisYodi is a Yeho-native intelligence engine for compact specialists, Neural Boards, memory, routing, generation, multimodal work, and bounded serving. Specialized Yodis can form a typed team and eventually control admitted application or OS capabilities without bypassing permissions, evidence, or the owning system. Every station and route can be selected for more information.DATA REFINERY transfers examples into TOKENIZER + SEQUENCE through a named, inspectable route.R01 / examplesTOKENIZER + SEQUENCE transfers encoded sequences into MODEL + CONTRACTS through a named, inspectable route.R02 / encoded sequencesMODEL + CONTRACTS transfers objective into TRAINING ENGINE through a named, inspectable route.R03 / objectiveTRAINING ENGINE transfers candidates into TOURNAMENT + EVAL through a named, inspectable route.R04 / candidatesTRAINING ENGINE transfers parameters + proof into CHECKPOINT IDENTITY through a named, inspectable route.R05 / parameters + proofCHECKPOINT IDENTITY transfers admitted state into INFERENCE + GENERATION through a named, inspectable route.R06 / admitted stateMODEL + CONTRACTS transfers state into MEMORY + MULTIMODAL through a named, inspectable route.R07 / stateMEMORY + MULTIMODAL transfers context into INFERENCE + GENERATION through a named, inspectable route.R08 / contextINFERENCE + GENERATION transfers result into ADMISSION + ACTION CONTROL through a named, inspectable route.R09 / resultTRAINING ENGINE transfers learning evidence into SPECIALISTS + NEURAL BOARDS through a named, inspectable route.R10 / learning evidenceSPECIALISTS + NEURAL BOARDS transfers promotion candidate into TOURNAMENT + EVAL through a named, inspectable route.R11 / promotion candidateMODEL + CONTRACTS transfers graphs into YEHO BACKENDS through a named, inspectable route.R12 / graphsYEHO BACKENDS transfers execute into INFERENCE + GENERATION through a named, inspectable route.R13 / executeTOURNAMENT + EVAL transfers winner into ACCEPTANCE REPORT through a named, inspectable route.R14 / winnerCHECKPOINT IDENTITY transfers identity into ACCEPTANCE REPORT through a named, inspectable route.R15 / identityADMISSION + ACTION CONTROL transfers containment into ACCEPTANCE REPORT through a named, inspectable route.R16 / containmentYEHO BACKENDS transfers parity into ACCEPTANCE REPORT through a named, inspectable route.R17 / parityVersioned refinery steps establish deterministic examples, splits, provenance, and checksums.INN0101 / CORPUSDATAREFINERYIN 00InputOUT 01Tokenizer identity, sequence framing, masks, positions, and bounded context turn examples into stable model inputs.FXN0202 / ENCODINGTOKENIZER +SEQUENCETokenizer checksum is evidenceIN 01ProcessOUT 01Typed model, tensor, objective, and binary contracts define what training and inference are allowed to mean.FXN0303 / SHAPEMODEL +CONTRACTSIN 01ProcessOUT 03Training executes admitted objectives, schedules, plasticity rules, and reproducible parameter updates.FXN0404 / LEARNINGTRAININGENGINESelected model and loss history...IN 01ProcessOUT 03Pinned candidates, budgets, splits, metrics, and tie-breaks determine which model can advance.QAN0505 / PROMOTIONTOURNAMENT+ EVALHeld-out evidence controls prom...IN 02EvidenceOUT 01Versioned checkpoints bind parameters, architecture, tokenizer, sequence, proof metadata, and corruption rejection to one resumable identity.DBN0606 / STATECHECKPOINTIDENTITYIN 01State storeOUT 02Versioned memory and typed multimodal inputs extend context without collapsing every signal into an unstructured token stream.DBN0707 / CONTEXTMEMORY +MULTIMODALMemory has explicit read and wr...IN 01State storeOUT 01The admitted model performs deterministic inference and bounded generation over selected tokenizer, sequence, and memory state.FXN0808 / BEHAVIORINFERENCE +GENERATIONModel and tokenizer checksums a...IN 03ProcessOUT 01Inference, specialist routing, and future OS actions pass schema, budget, capability, policy, consent, and containment checks before leaving Yodi.APIN0909 / CONTAINMENTADMISSION +ACTION CONTROLIN 01BoundaryOUT 01A bundle of specialized Yodis, inspectable Neural Boards, plastic reasoning, and continual-learning proposals advance only through bounded promotion records.QAN1010 / RESEARCHSPECIALISTS +NEURAL BOARDSSpecialists declare purpose and...IN 01EvidenceOUT 01Yeho owns CPU oracle, Vulkan, and direct-native Kyber execution while Yodi owns intelligence behavior.FXN1111 / EXECUTIONYEHOBACKENDSIN 01ProcessOUT 02One machine-readable artifact records winners, metrics, checksums, repeat flags, containment outcomes, backend parity, and the deliberately narrow claim.QAN1212 / EVIDENCEACCEPTANCEREPORTApplication-generated and verif...IN 04EvidenceOUT 00DHL-VISUAL-PROGRAM / YODI / REV 04VARIABLE NODES / DISTRIBUTED PORTS / INSPECTABLE CONTRACTS