Paper deep dive
Valhalla: A Layered Knowledge-State and Service-Governance Framework for Long-Term Scientific Knowledge Work
Yuyang Zheng, Nan Li, Wenxia Deng, Lige Yan, Xiang Li, Si Chen
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 95%
Last extracted: 8/18/2026, 5:53:57 AM
Summary
The paper introduces Valhalla, a framework for long-term scientific knowledge work that replaces flat, node-centric knowledge graphs with a layered File-Resource-Entity-Relationship-Graph (FREG) model. This structure preserves provenance and enables the sharing and reorganization of knowledge across researchers. It also proposes a Router-Contract-Workflow service-governance architecture to control LLM access and modification of knowledge states, validated through an antibody-design review task.
Entities (10)
Relation Signals (8)
Valhalla → uses → FREG
confidence 98% · Valhalla replaces flat graphs with layered encapsulation and stable semantic boundaries through a five-layer File-Resource-Entity-Relationship-Graph (FREG) model.
FREG → consistsof → File
confidence 95% · The five-layer File–Resource–Entity–Relationship–Graph (FREG) knowledge model that separately encapsulates raw-material entry points...
FREG → consistsof → Resource
confidence 95% · The five-layer File–Resource–Entity–Relationship–Graph (FREG) knowledge model that separately encapsulates... source identities...
FREG → consistsof → Entity
confidence 95% · The five-layer File–Resource–Entity–Relationship–Graph (FREG) knowledge model that separately encapsulates... knowledge objects...
FREG → consistsof → Relationship
confidence 95% · The five-layer File–Resource–Entity–Relationship–Graph (FREG) knowledge model that separately encapsulates... semantic relations...
FREG → consistsof → Graph
confidence 95% · The five-layer File–Resource–Entity–Relationship–Graph (FREG) knowledge model that separately encapsulates... task-level knowledge networks
Valhalla → uses → Router-Contract-Workflow
confidence 95% · We further introduce a Router-Contract-Workflow service-governance architecture... to constrain how language models access, modify, and extend knowledge states
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:As large language model (LLM) agents are increasingly adopted in scientific research, external knowledge bases, knowledge graphs, and long-term memory have improved information retrieval and task continuity. However, most structured knowledge systems remain node-centric, representing files, concepts, results, and judgments as nodes and relations in a graph. While suitable for personal knowledge management, such structures often depend on individual organizational practices, limiting knowledge sharing, integration, and reorganization across users. This paper presents Valhalla, a layered knowledge-state and service-governance framework for long-term scientific knowledge work. Valhalla replaces flat graphs with layered encapsulation and stable semantic boundaries through a five-layer File-Resource-Entity-Relationship-Graph (FREG) model. File and Resource preserve source identity and provenance, Entity represents knowledge objects, Relationship captures semantic judgments, and Graph provides task-oriented knowledge views, enabling knowledge states from different researchers to be exchanged and reorganized under a unified structure. We further introduce a Router-Contract-Workflow service-governance architecture, inspired by the microkernel paradigm, to constrain how language models access, modify, and extend knowledge states while maintaining structural consistency and auditable operational boundaries. We implement a Valhalla prototype and validate knowledge ingestion, cross-member integration, and scientific writing support through an antibody-design review task comprising 26 paper resources, 80 knowledge entities, and 92 semantic relations. Rather than proposing a new knowledge-extraction algorithm, Valhalla offers a paradigm for organizing collaborative scientific knowledge, transforming individualized knowledge structures into transferable and reorganizable shared knowledge states.
Tags
Links
- Source: https://arxiv.org/abs/2608.15193v1
- Canonical: https://arxiv.org/abs/2608.15193v1
Trouble viewing inline? Open PDF directly →
Full Text
63,407 characters extracted from source content.
Expand or collapse full text
Valhalla:ALayeredKnowledge-StateandService-Governance FrameworkforLong-TermScientificKnowledgeWork YuyangZheng 1 ,NanLi 1 ,WenxiaDeng 1 ,LigeYan 1 ,XiangLi 2, *,SiChen 1, * 1 SchoolofMedicine,ShanghaiUniversity,Shanghai,China 2 SchoolofPharmacy,SecondMilitaryMedicalUniversity,Shanghai,China Abstract Aslargelanguagemodel(LLM)agentsareincreasinglyadoptedinscientificresearch,existingstudieshave demonstratedthatexternalknowledgebases,knowledgegraphs,andlong-termmemorycanimprove informationretrievalandtaskcontinuity.However,mostcurrentstructuredknowledgesystemsstillemploy node-centricgraphorganization,representingfiles,concepts,experimentalresults,andknowledgejudgments collectivelyasnodesandrelationsinagraph.Althoughthisdesignissuitableforpersonalknowledge management,inmulti-usercollaborationandlong-termknowledge-transfersettingsittendstomake knowledgestructuresdependentonindividualorganizationalpractices,renderingknowledgenetworkscreated bydifferentusersdifficulttoshare,integrate,andreorganizedirectly.ThispaperpresentsValhalla,alayered knowledge-stateandservice-governanceframeworkforlong-termscientificknowledgework.Thecentralidea ofValhallaistotransformstructuredknowledgefromaflatgraphintoalayeredencapsulationstructurewith stablesemanticboundaries.Thesystemadoptsafive-layerFile–Resource–Entity–Relationship–Graph(FREG) knowledgemodelthatseparatelyencapsulatesraw-materialentrypoints,sourceidentities,knowledgeobjects, semanticrelations,andtask-levelknowledgenetworks,therebyenablingknowledgestatesproducedby differentresearcherstobeexchangedandreorganizedaccordingtoaunifieddatastructure.Withinthismodel, FileandResourcepreservesourceidentityandprovenancetracing,Entityobjectifiesknowledgecontent, Relationshipretainscross-objectsemanticjudgments,andGraphprovidestask-orientedviewsfororganizing knowledge.Onthebasisofthislayeredknowledgestate,wefurtherproposeaRouter–Contract–Workflow service-governancearchitectureinspiredbythemicrokernelparadigm.Thisarchitectureconstrainshow languagemodelsaccess,modify,andextendknowledgestates,ensuringthatknowledgesharingmaintains structuralconsistencywhilealsoprovidingauditableoperationalboundaries.WeimplementaValhalla prototypeandvalidateitsworkflowsforknowledgeingestion,cross-memberknowledgeintegration,and scientificwritingsupportthroughanantibody-designreviewtask.Theexperimentconstructsastructured knowledgestatecomprising26paperresources,80knowledgeentities,and92semanticrelations,and demonstrateshowknowledgeproducedbydifferentteammemberscanbeorganizedandreusedwithina unifiedlayeredarchitecture.Theprincipalcontributionofthisworkisnotanewknowledge-extraction algorithm,butaparadigmfororganizingstructuredknowledgeincollaborativescientificresearch:by replacingnode-centricflatknowledgegraphswithlayeredencapsulation,Valhallaenablesknowledgeassetsin theLLMeratomovebeyondindividualizedorganizationalschemesandbecomesharedknowledgestateswith aunifiedstructurethatsupportstransferandreorganization. Keywords:largelanguagemodelagents;structuredknowledgestates;layeredknowledgeencapsulation; knowledgetransfer;scientificAIinfrastructure Github:https://github.com/fisherrael123-png/Valhalla 1Introduction AsLLMagentshaveenteredscientificandengineeringsettings,twoprincipallinesof developmenthaveemerged.ThefirstextendssourceacquisitionbyLLMagentsintoaprocessof knowledgeorganization.Thislineofresearchholdsthatsourcematerialsusedinscientificand engineeringtasksarenotthemselvesdirectlyreusableknowledge,butratherrawcarriersof knowledge.Onlythroughentitynormalization,evidencebinding,relationmodeling,andprovenance tracingcandispersed,redundant,andcontext-dependentmaterialsbetransformedintostable structuredknowledgerepresentations.SuchrepresentationsallowanLLMagenttomovebeyond one-offretrievalandshort-termcontext,continuallyreuseestablishedjudgmentsinsubsequenttasks, compareevidencefromdifferentsources,maintaindependenciesamongitemsofknowledge,and supporthigher-levelreasoning,review,andcollaboration[1‑3]. ThesecondlinefocusesonoptimizinghowLLMagentsadvanceprojectsandiscollectively referredtoasAIengineering.Itencompassesseveralsubareas,includingpromptengineering, retrieval-augmentedgeneration(RAG),contextmanagement,skillaccumulationandcontinuous capabilityextension,andconstraintengineering.RAGcombinesparametricgenerationwithexternal documentretrievaltoprovideanupdatableinformationinterfaceforknowledge-intensivetasks[4], whereasGraphRAGincorporatescross-documententitiesandtheirconnectionsintoagraphstructure [5].Researchonexternalmemoryhas,fromanotherperspective,demonstratedtheimportanceof persistentstate[6‑9].Contextmanagementdrawsonhierarchicalstorageandvirtual-context managementtoscheduleinformationbetweenalimitedcontextwindowandexternalstorage[6]. SkillextensionfocusesonenablinguserstoevolveanLLMagent’sfunctionalityduringuse,whereas constraintengineeringfocusesonadvancingprojectsthroughindependentreviewmechanisms involvingmultipleLLMagents[10‑15]. Together,thesetwolinesofdevelopmentestablishtwoessentialfacts:long-termknowledge workrequiresboththestructuringofcontentandevidenceandtheadaptationofLLMagent capabilitiestotherequirementsofknowledgework. 2MainContributions Inthispaper,“long-termscientificknowledgework”referstoresearchactivitiesinwhichsource materials,judgments,andprocessespersistacrossmultiplesessions,tasks,orparticipants.Suchwork mustsatisfythefollowingrequirements: 1.Long-termknowledgemustbereusable,shareable,andreorganizableforapplicationacross diversescenariosorprojects.Reusabilityisthefundamentalpurposeoflong-termknowledge becauseitimprovesanLLMagent’seffectiveness;shareabilityandreorganizabilityfurther transformlong-termknowledgefrommeredataintoascientificasset. 2.Coreusersinscientificteamstypicallypossesssubstantialdomainexpertisebutmaynot havebackgroundsinsystemsengineering,softwaredevelopment,orartificialintelligence. Knowledge-managementandknowledge-useprocessesshouldthereforebereadily understandabletousers,enablingthemtoevolveandadjustLLMagentsforspecificprojects. Table2.1RelationshipbetweenexistingLLM-agentresearchdirectionsandValhalla’sdesignrequirements Theserequirementsyieldthreeinterrelateddesignrequirementsthatmustneverthelessbe validatedseparately: D1:Accumulationandtransferofprovenance-constrainedknowledgestates.Thesystemshould distinguishfileinstances,sourceidentities,knowledgeobjects,semanticrelations,andtaskgraphs whilepreservingpathsbacktotheoriginalmaterials. D2:Auditableextensionofdomainservices.Thesystemshouldrepresenttaskentrypoints, eligibilityconditions,andexecutionstepsasexplicitcomponentsthatresearcherscanreadand modify,ratherthanrelyingsolelyonone-offprompts. D3:Human-in-the-looplong-termgovernance.Persistentwrites,cross-layermodificationsto knowledge,andserviceevolutionshouldbeconstrainedbypermissions,state,confirmation,and stagechecks,withthelocationoffailuresrecorded. Thesethreerequirementsentailinherenttensions:anexcessiveemphasisontransfermayerase sourcecontext,prioritizingrapidextensionmaycreateservicecoupling,andreadablenatural- languagerulesprovideweakerguaranteesthanruntimeisolation. ThispaperproposesValhalla,alayeredknowledge-stateandservice-governanceframeworkfor long-termscientificknowledgework.Centeredonstructuredknowledgewithlayeredencapsulation, Valhallaconstructsasmuchoftheoverallsystemaspossibleinnaturalandstructuredlanguages, followingamicrokerneloperating-systemarchitecture.Beyondthetwoprincipallinesof developmentintroducedabove,Valhallaservestwoadditionalpurposes.First,layeredencapsulation stabilizesknowledgestructures,makingthesharingofstructuredknowledgepossible.Second,it providesextensionmechanismsorientedtowardnaturallanguageanddomainworkflows,allowing researcherstoadjustknowledgestructures,analyticalprocesses,andtaskcapabilitiesaccordingto theneedsofdifferentresearchprojects. Valhallaisdesignedtoplacethetrade-offsamongD1–D3withininspectableobjectsand operationalinterfaces,ratherthantoclaimthatthesetrade-offshavealreadybeeneliminated. 3OverallArchitectureofValhalla 3.1FrameworkDefinition Valhallaisalayeredknowledge-stateandservice-governanceframeworkforlong-term scientificknowledgework.Itrepresentsresearchcontentasmaintainabledataobjectsandrequires languagemodelstoaccessormodifythoseobjectsthroughexplicitserviceinterfaces.“Contentfirst” meansthatthesystemestablishestheidentitiesofsources,knowledge,andrelationsbefore determiningwhichtasksamodelshouldperform;servicesmaynotbypassobjectboundariesmerely foreaseofexecution. Analogiestocomputersystemsprovidealanguageforanalyzingstate,modules,andresource managementinLLM-agentarchitectures.MemGPTlikewiseusesanoperating-systemanalogyto manageschedulingbetweencontextandexternalmemory[6,16].Valhallausesthisanalogymore narrowly:itdoesnotsimulateanunderlyingoperatingsystem,butinsteadadoptsconceptssuchas thedataplane,controlplane,serviceentrypoints,andstatecheckstoorganizeknowledgeoperations. InValhalla,thelong-termknowledgestateconsistsofFile,Resource,Entity,Relationship,and Graph.Thesefiveobjectclassesassumedistinctidentity-relatedresponsibilities:Filepreserves human-facingentrypointstosourcematerials;Resourceprovidesastablesourceidentity;Entity representsaprovenance-constrainedknowledgeobject;Relationshiprecordstypedsemantic judgments;andGraphorganizesentitiesandrelationsrelevanttoatask. 3.2DataPlaneandControlPlane Thedataplaneanswersthequestion“Whatdoesthesystemstore?”,whereasthecontrolplane answers“Howdoesthemodeloperateontheseknowledgestates?”Neithercansubstituteforthe other.Withonlyadataplane,knowledgeobjectsmaystillberewrittenarbitrarilybythemodel;with onlyacontrolplane,serviceslackstable,traceableobjectsonwhichtooperate. Figure3.1ThedataplaneandcontrolplanejointlyconstituteValhalla’soverallframework BeginningwithFilesthatresearcherscanmanagedirectly,thedataplaneestablishesstable sourceidentitiesthroughResources,andthenseparatelyrepresentsknowledgeobjectswithinsources, cross-objectjudgments,andtaskviewsasEntities,Relationships,andGraphs.Whenevidencemust bereviewed,thesystemtracesknowledgeobjectsorrelationsbacktotheoriginalmaterials.When mechanismsorresearchgapsmustbeanalyzed,itinsteadprioritizescoresemanticrelations, avoidingconflationoffilelocationswithknowledgemeaning. ThecontrolplaneorganizesknowledgeoperationsthroughRouter,Contract,andWorkflow. Routerclassifiesrequests;Contractdeclaresinputs,permissions,risks,systemstates,confirmation requirements,andoutputboundaries;andWorkflowspecifiestheexecutionstepsonceeligibility conditionshavebeensatisfied.TheprovenancecontractsandskillregistryofFundaPod,together withprovenancearchitecturesforinteractiveworkflows,providepointsofreferenceforthissystem- levelorganization;Valhalla,however,furtherseparatesknowledgelayersexplicitlyfromoperational eligibility[17‑19]. Together,thetwoplanesformadesignpathwaytowardD1–D3,ratherthanevidencethatthe threeobjectiveshavealreadybeenachieved.Thepresenceofprovenancemappingsdoesnotimply thatEntitiesorRelationshipsarenecessarilycorrect,justasthepresenceofconfirmationrulesdoes notimplythatamodelcannotcircumventnatural-languageconstraints. Figure3.2OverallarchitectureofValhalla 4DataPlane:File–Resource–Entity–Relationship–Graph Valhalla’sdata-planedesigndrawsonrelevantexperiencefromexternalretrieval,graph structures,evidencegraphs,andteamknowledgememory.Itsfocus,however,isnottoproposea newextractionalgorithm,buttodefineobjectidentities,maintenanceresponsibilities,and traceabilitypaths[1,5,20]. Thefive-layermodelcanbedenotedas퐀≝퐀 , 퐀 , 퐀 , 퐀᐀퐀 , 퐀,where: F:File,thesetofrawfiles.Thisfilelayerisresearcher-facingandpreservesentrypointstothe originalmaterialsaswellastracesofhumanorganization; R:Resource,thesetofsourcematerials.Thisresourcelayerissystem-facingandprovidesstable sourceidentitiesandevidenceinterfaces; E:Entity,thesetofknowledgeentities.Thisentitylayerobjectifiesprovenance-constrained unitsofknowledge; Rel:Relationship,thesetofrelations.Thisrelationshiplayerrecordstypedsemanticjudgments acrossobjects; 퐀:Graph,thesetofknowledgegraphs.Thisgraphlayerorganizesrelevantentitiesandrelations aroundspecifictasks. Eachlayerofthefive-layerencapsulationmodelcontainselementswithasingletypeofidentity andresponsibility,givingthestructuredknowledgeconstructedbyeveryuseraunifiedarchitecture. Thismakesthemodelafoundationforsharingstructuredknowledge. Figure4.1Five-layerobjectmodeloftheValhallaknowledgesystem 4.1FileandResource:FromHuman-FacingEntryPointstoStableSourceIdentities Scientificmaterialsfirstentertheworkspaceinformssuchasfoldersofpapers,archived webpages,experimentalrecords,meetingminutes,coderepositories,andprojectdocuments.Because thelocationofafilewithinadirectoryreflectsreadingpaths,taskstages,andhumanclassification, theFilelayerpreservesentrypointsthatresearcherscandirectlyunderstandandmaintain. TheFilelayerdoesnotrequirephysicaldeduplicationthroughthemovementordeletionoffiles. Thesamepapermaybestoredinmultipledirectoriesbecauseitwasfoundthroughdifferentsearch pathsorusedindifferenttaskcontexts;forciblychangingthedirectorystructurewouldimpairthe intelligibilityofprovenanceandworkprocesses.Logicalnormalizationshouldinsteadoccuratthe morestablesource-identitylayer. TheResourcelayerabstractsrawfilesintosourceobjectsthatthesystemcanidentify,reference, andtrace.EachResourcehasastableidentifierandrecordsthetitle,type,origin,publiccopy, lifecyclestatus,andmappingstoFileinstances.Fileanswers“Whereisthesource,andwhywasit placedhere?”,whereasResourceanswers“Howdoesthesystemidentifythesamesource?”[2‑3]. Accordingly,multipleFileinstancesofthesamesourcemayberetainedwhilemappingtoa singlestableResourceidentifier.Thismechanismpreservesresearchers’fileorganizationwhile reducingtheriskthatduplicatefileswillbemistakenfordistinctsourcesofevidence.Identity matchingcanneverthelessbeerroneous,particularlywhenversions,supplementarymaterials,or titlesdiffer,andaninterfaceformanualverificationmustthereforeberetained. 4.2Entity:Provenance-ConstrainedKnowledgeObjects AnEntityrepresentsanindependentlymaintainableknowledgeobjectextractedfromasource, ratherthanthesourcefileitself.Real-worldliteraturesimultaneouslycontainsbackground information,methods,experimentalconditions,authorconjectures,andnoise.Thepurposeof objectificationistodelineatetheboundariesofreusableknowledgeunitswhilepreservingtheir provenance. AResourcemaycontainmultiplerelativelyindependentproblems,mechanisms,ormethods.If theseelementshavedistinctevidencelocations,scopesofapplicability,orconclusions,theyshould formseparateEntitiesratherthanbeingcompressedintoasinglesummary.Thegranularityof decompositionmustsupportsubsequenttracingandrelationconstruction;itcannotbedetermined solelybytitlesorkeywords. Similarknowledgefromdifferentsourcesisnotmergedatthislayer.Similarstatementsmay dependondifferentexperimentalconditionsandargumentativecontexts,anddirectmergingwould erasedifferencesinprovenance.Valhallaretainssource-specificEntitiesseparatelyanduses Relationshipstoexpressequivalence,overlap,support,attenuation,orconflict[1‑3]. TheEntitylayerthereforeperformsthetransformationfromsourcematerialsintomaintainable knowledgeobjects.ItretainsevidentiaryprovenancethroughResourcebutdoesnotindependently performintegrationacrosssources.Objectboundariesandcontentcorrectnessstillrequirereview, andanactivestatusinthesystemcannotsubstitutefordomain-leveljudgmentsoffactualvalidity. 4.3Relationship:TypedSemanticJudgments Thereuseofscientificknowledgedependsnotonlyonindividualobjectsbutalsooninspectable connectionsamongthem.ARelationshipcontains,atminimum,asourcenode,atargetnode,a relationtype,atextualdescription,astatus,andevidentiaryprovenance.Relationtypesmaybe definedaccordingtothetaskandmayincludesupports,contradicts,causes,constrains,detects, implements,anddepends_on[1,3]. Explicitrelationsallowthesystemtodeterminewhetheraconclusionhasachainofsupport, whetheraproposedapproachcontainsgapsinitsevidence,andwhichdependentmodulesmaybe affectedbyanengineeringchange.Theirvalueliesinpreservingthestructureofjudgmentsand providingentrypointsfortracing,ratherthaninsubstitutinglinesinagraphforcausalproof. 4.4Graph:Task-OrientedKnowledgeNetworks AGraphconsistsofselectedEntitiesandRelationshipsandmaycorrespondtoaresearchtopic, project,review,experimentalprotocol,orengineeringmodule.Itdiffersfromfile-linkgraphsin systemssuchasObsidian:nodesinafile-linkgraphspanboththeentityandfilelayers,anditslinks likewisecrosslayerboundaries[21].Bycontrast,Valhalla’scoresemanticgraphpresentsobjectified knowledgeandtypedjudgments. TheroleofaGraphistoprovideastructuredindexforthecurrenttask,ratherthantovisualize everyobjectintheknowledgebase.ResearcherscanfirstlocatekeyEntities,Relationships,and evidencegaps,andthentracethembacktotheoriginalsourcesasneeded.Insufficientgraph coverageornumerousisolatednodesshouldbetreatedasgapsinknowledgeorganization,rather thanasacompletetheoreticalstructure. 4.5CoreSemanticGraphandEvidenceMappings Figure4.2Structuralrelationshipsamongthefileview,coresemanticgraph,andextendedevidencegraph Toavoidconflatingfilelocations,sourceprovenance,andconceptualrelations,Valhalla distinguishesamongthefileview,coresemanticgraph,andextendedevidencegraph.Thefileview retainsthedirectorystructureofFiles;thecoresemanticgraphcontainsonlyEntitiesand Relationships;andtheextendedevidencegraphaddsResourcemappingswhenverificationis required,therebyconnectingknowledgejudgmentstotheoriginalsources. Thesystemcanswitchperspectivesaccordingtothetask:usersinspectthecoresemanticgraph tounderstandknowledgestructure,expandResourcemappingstoreviewprovenance,andlocate originalfilesthroughFile–Resourcemappingswhenreturningtosourcematerials.Suchswitching reducestheconflationofviewsbutcannotcompensateforerroneoussourceregistration,entity extraction,orrelationjudgments. 5FormalArchitectureoftheDataPlane퐀 5.1Five-LayerKnowledge-StateArchitecture 퐀≝퐀 , 퐀 , 퐀 , 퐀᐀퐀 , 퐀 Aconcreteimplementationof퐀inthedataplaneisreferredtoasRoot퐀;thus: 퐀 퐀 ≝퐀 퐀 ,퐀 퐀 ,퐀 퐀 ,퐀᐀퐀 퐀 ,퐀 퐀 where: 퐀 퐀 :thesetofFilesinRoot퐀. 퐀 퐀 :thesetofResourcesinRoot퐀. 퐀 퐀 :thesetofEntitiesinRoot퐀. 퐀᐀퐀 퐀 :thesetofRelationshipsinRoot퐀. 퐀 퐀 :thesetofGraphsinRoot퐀. 5.2Inter-LayerKnowledge-StateTransformationsintheDataPlane, ℳ 퐀 ThefivelayersoftheValhalladataplaneareconnectedthroughasetofmappingswithdistinct responsibilitiesthatperforminter-layerknowledge-statetransformations(Inter-layerknowledge-state transformationsofthedataplane). Thefiveencapsulatedstructured-datalayersarelinkedbyfourtransformationmappingsthat serveasinterfacesbetweenadjacentlayers. 1.File-to-ResourceTransformationMappingℳ 퐀 ThismappingmapsanoriginalFiletoastandardizedlogicalResource. ℳ 퐀 퐀 : 퐀↦퐀 where: 퐀∈퐀 , 퐀∈퐀 MultipleFilesmaybemappedtoasingleResource: 퐀 1 ≠퐀 2 , ℳ 퐀 퐀 퐀 1 =ℳ 퐀 퐀 퐀 2 =퐀 Forexample,Fileinstanceswithdifferentpathsandfilenames,oreveninstancessavedby differentmembers,maybeidentifiedbythesystemasthesamesourcematerial. Thiscanberepresentedasasetmapping(function): ℳ 퐀 퐀 : 퐀↦퐀 2.Resource-to-EntityTransformationMappingℳ 퐀 EntitiesareextractedfromaResource. ℳ 퐀 퐀 :퐀↦᐀ AResourcemaycontainmultipleEntities: ℳ 퐀 퐀 :퐀↦᐀ 1 ,᐀ 2 ,᐀ 3 ,⋯ Thiscanberepresentedasasetmapping(function): ℳ 퐀 퐀 : 퐀↦퐀 퐀 퐀 :thepowersetofE. 3.Entity-to-RelationshipTransformationMappingℳ 퐀᐀퐀 RelationshipsamongEntitiesaresummarizedandorganized. ℳ 퐀᐀퐀 퐀 : ᐀ 1 , ᐀ 2 ↦ℓ TwoEntitiesmayhaveRelationshipsofdifferenttypes: ℳ 퐀᐀퐀 퐀 : ᐀ 1 , ᐀ 2 ↦ℓ 1 , ℓ 2 , ⋯ Thiscanberepresentedasasetmapping(function): ℳ 퐀᐀퐀 퐀 : 퐀×퐀↦퐀᐀퐀 ×:Cartesianproduct. 퐀᐀퐀 :thepowersetofRel. 4.Relationship-to-GraphTransformationMappingℳ 퐀᐀퐀䠀 Distinctknowledgegraphsareconstructedaccordingtotopic. ℳ 퐀᐀퐀䠀 퐀 : ℓ 1 , ℓ 2 ,⋯↦퐀 Thiscanberepresentedasasetmapping(function): ℳ 퐀᐀퐀䠀 퐀 :퐀᐀퐀↦퐀 Because퐀alsorecordsinformationaboutEntities: ℳ 퐀᐀퐀䠀 퐀 :퐀×퐀᐀퐀↦퐀 ×:Cartesianproduct. 퐀᐀퐀:thepowersetofRel. 퐀 :thepowersetofE. 5.CompositeTransformationMappings Thefourtransformationmappingscanbesummarizedas: ℳ 퐀 :퐀↦퐀 ; 퐀↦퐀 ; 퐀↦퐀᐀퐀 ; 퐀᐀퐀↦퐀 Adjacentmappingscanbecomposed.Forexample: 퐀᐀퐀= ℳ 퐀᐀퐀 퐀 (ℳ 퐀 퐀 (ℳ 퐀 퐀 퐀))= ℳ 퐀᐀퐀 퐀 ∘ℳ 퐀 퐀 ∘ℳ 퐀 퐀 퐀 Forreadability,thismayalsobeexpressedas: 퐀᐀퐀=ℳ 퐀 퐀 ℳ 퐀 퐀 ℳ 퐀᐀퐀 퐀 퐀 Similarly: 퐀=ℳ 퐀᐀퐀䠀 퐀 ( ℳ 퐀᐀퐀 퐀 ( ℳ 퐀 퐀 ( ℳ 퐀 퐀 퐀 )))= ℳ 퐀 퐀 ℳ 퐀 퐀 ℳ 퐀᐀퐀 퐀 ℳ 퐀᐀퐀䠀 퐀 퐀 5.3ProvenanceMapping, ℳ 퐀 Transformationmappingsconstructthefiveencapsulatedlayersofstructuredknowledgefrom thebottomupward,layerbylayer.Provenancemappingproceedsfromthetopdownwardtomap eachobjecttoitssources[3]. 1.GraphProvenanceMappingℳ 䠀 퐀 AlthoughaGraphisconstructedfromRelationships,itrecordsboththeRelationshipsandthe Entitiesusedtoconstructit. ℳ 䠀 퐀 : 퐀↦ℓ 1 , ℓ 2 , ⋯ , 퐀↦᐀ 1 , ᐀ 2 , ᐀ 3 , ⋯ Representedasasetmapping(function): ℳ 䠀 퐀 : 퐀 퐀 ↦ 퐀 퐀᐀퐀, 퐀 퐀 2.RelationshipProvenanceMappingℳ 퐀᐀퐀 퐀 ARelationshiprecordstheEntitiesfromwhichitwasconstructed. ℳ 퐀᐀퐀 퐀 : ℓ↦᐀ 1 ×᐀ 2 Representedasasetmapping(function): ℳ 퐀᐀퐀 퐀 : 퐀 퐀᐀퐀 ↦퐀 퐀 3.EntityProvenanceMappingℳ 퐀 퐀 AnEntityrecordstheResourcefromwhichitwasconstructed. ℳ 퐀 퐀 : ᐀↦퐀 Representedasasetmapping(function): ℳ 퐀 퐀 :퐀↦퐀 4.ResourceProvenanceMappingℳ 퐀 퐀 AResourcerecordstheFilesfromwhichitwasconstructed. ℳ 퐀 퐀 :퐀↦퐀 퐀 ,퐀 1 ,퐀 2 ,⋯ Here,퐀 퐀 isthecommoncopyof퐀 1 ,퐀 2 ,⋯.Thiscommoncopyisestablishedsothatℳ 퐀 퐀 can obtainacanonicalrecord퐀 퐀 unaffectedbymanualoperations,therebypreventingchangestoFiles fromrenderingprovenanceunreachable. Representedasasetmapping(function): ℳ 퐀 퐀 : 퐀↦ 퐀 퐀 , 퐀 5.CompositeProvenanceMappings Thefourprovenancemappingscanbesummarizedas: ℳ 퐀 : 퐀↦ 퐀᐀퐀 , 퐀 ; 퐀᐀퐀↦퐀 ; 퐀↦퐀 ; 퐀↦ 퐀 퐀 , 퐀 Adjacentmappingscanbecomposed.Forexample: 퐀 퐀= ℳ 퐀 퐀 ( ℳ 䠀 퐀 퐀)= ℳ 퐀 퐀 ∘ℳ 䠀 퐀 퐀 Forreadability,thiscanbewrittenas: 퐀 퐀= ℳ 䠀 퐀 ℳ 퐀 퐀 퐀 Consequently,whenknowledgeGraphGisused,itsprovenancecan,wherenecessary,betraced backtotheoriginalFiles. 5.4CoreSemanticGraphandExtendedEvidenceGraph ThecoresemanticgraphfortopicTis: 퐀 퐀 ≝퐀᐀퐀 퐀 ,퐀 퐀 where: 퐀 퐀 ⊆퐀 퐀 , 퐀᐀퐀 퐀 ⊆퐀᐀퐀 퐀 TheextendedevidencegraphfortopicTis: 퐀 퐀 ᐀ ≝ 퐀᐀퐀 퐀 , 퐀 퐀 , 퐀 퐀 , 퐀 퐀 where: 퐀 퐀 =ℳ 퐀 퐀 퐀,퐀 퐀 =ℳ 퐀 퐀 ℳ 퐀 퐀 퐀 Thatis: 퐀 퐀 ᐀ =퐀 퐀 ∪ℳ 䠀 퐀 ℳ 퐀 퐀 퐀 퐀 ∪ℳ 䠀 퐀 ℳ 퐀 퐀 ℳ 퐀 퐀 퐀 퐀 6FormalArchitectureoftheImplementedDataPlane 퐀 ∗ 퐀providesthearchitecturalfoundationforsharingstructuredknowledge,butthereorganization ofstructuredknowledgemustalsobeimplemented. 6.1ImplementedKnowledge-StateArchitecture 퐀 ∗ Inpracticalimplementation,thefive-layerknowledge-statearchitecturesupportsthereuseand reorganizationofsourcematerialsattheResourcelayerthroughtheconstructionofmultiple knowledgebases. 퐀 ∗ ≝ 퐀,퐀,퐀 Theknowledge-baseset퐀comprisesoneormoreknowledgebases퐀: 퐀= 퐀 1 , 퐀 2 ,⋯ Eachknowledgebase퐀hasanindependentknowledgespacecomposedof퐀,퐀᐀퐀,퐀.In addition,eachknowledgebasehasavirtualResourcelayerVtosupportconstrainedprovenance tracing. 퐀≝퐀 퐀 ,퐀 퐀 ,퐀᐀퐀 퐀 ,퐀 퐀 Therefore: 퐀 ∗ ≝퐀 , 퐀 , 퐀,퐀,퐀᐀퐀,퐀 퐀 1 ,퐀,퐀,퐀᐀퐀,퐀 퐀 2 , ⋯ 6.2VirtualResourceLayer 퐀 퐀 ThevirtualResourcelayer퐀 퐀 consistsoffourcomponents,allofwhoseentriesareResources ratherthanFiles. 퐀 퐀 :thelocalResourcetable,local_resources.Thistablecontainsmaterialsforgeneraluse. Whenknowledgebase퐀isusedtoadvanceaproject,materialsregisteredhereareincludedwithin thescopeofconsiderationbutarenotnecessarilyused. 퐀 퐀 :therequiredResourcetable,required_resources.Whenknowledgebase퐀isusedto advanceaproject,materialsregisteredheremustbeused. 퐀 퐀 :theexcludedResourcetable,excluded_resources.Whenknowledgebase퐀isusedto advanceaproject,materialsregisteredheremustnotbeused.Thistablehasahigherprioritythan 퐀 퐀 ; thatis,when퐀 퐀 conflictswith퐀 퐀 ,theentriesin퐀 퐀 areretained. B:theblacklistofagivenRoot.ItcontainsmaterialsgloballydisabledwithinthatRoot;for example,ifasourceisfoundtoinvolveacademicfraud,itisexcludedfromtheknowledgesystem throughtheblacklistmechanism. Thus: 퐀 퐀 =퐀 퐀 ∪퐀 퐀 ∖퐀 퐀 ∪퐀 퐀 TheactualResourcesetinuse,퐀 퐀 ,thensatisfies: 퐀 퐀 ∖ 퐀 퐀 ∪퐀 퐀 ⊆퐀 퐀 ⊆퐀 퐀 6.3Resource-to- 퐀 TransformationMapping ℳ 퐀 퐀 ℳ 퐀 퐀 =ℳ 퐀 퐀 , ℳ 퐀 퐀 ,ℳ 퐀 퐀 ResourcesRin퐀aremappedtoentriesinthelocalResourcetable퐀 퐀 : ℳ 퐀 퐀 : 퐀↦퐀 퐀 ResourcesRin퐀aremappedtoentriesintherequiredResourcetable퐀 퐀 : ℳ 퐀 퐀 :퐀↦ 퐀 퐀 ResourcesRin퐀aremappedtoentriesintheexcludedResourcetable퐀 퐀 : ℳ 퐀 퐀 : 퐀↦퐀 퐀 6.4 퐀 -to-EntityTransformationMapping ℳ 퐀 퐀 ℳ 퐀 퐀 :퐀 퐀 ↦퐀 퐀 Itthereforefollowsthat: 퐀 퐀 =ℳ 퐀 퐀 ℳ 퐀 퐀 ℳ 퐀 퐀 ℳ 퐀᐀퐀 퐀 ℳ 퐀᐀퐀䠀 퐀 퐀 6.5SharedDataPlane RootfusionisusedasanexampletoillustratetheroleofthevirtualResourcelayer퐀 퐀 . LetthetwodistinctRootsbe: 퐀 1 ,퐀 2 TheyarefusedintoanewRoot: 퐀 3 1.Resource-LayerIntegration TheResourcelayerof퐀 3 is: 퐀 3 = 퐀Ѐ퐀瀀퐀栀᐀ 퐀 1 ∪퐀 2 where: Normalizedenotes: -thereassignmentofunifiedidentitiesintheResourcelayer; -thecontinuedseparateretentionofdifferentversionsofthesamepublication; -theavoidanceoferroneousmergingbetweensimilarbutdistinctmaterials. 2.BlockingStrategy ForaResourcersubjecttoablockingconflictbetween퐀 1 ,퐀 2 —thatis,risblockedbyonebut nottheother—auserdecisionisdefinedas: 퐀∈ 퐀Ѐ搀퐀 , 퐀Ѐ퐀 , 퐀Ѐ퐀 where: 퐀Ѐ搀퐀 :theResourceisgloballyblockedinthenewRootbyaddingitto퐀 3 ; local:theResourceisblockedonlylocallywithintherelevantknowledgebases.Specifically,it isnotaddedto퐀 3 ;instead,undertheRootwhose퐀 퐀 originallycontainedanentryforr,the knowledge-baseset퐀hasraddedtotheapplicable퐀 퐀 entries,therebyimplementinglocalblocking. Thatis: 퐀 퐀 ∗ =퐀 퐀 ∪퐀|퐀=퐀Ѐ퐀 allow:theResourceisnolongerblocked. 3.FusedKnowledgeBase퐀 Theknowledgebases퐀underthetwoRootsarefusedaccordingtotheblockingstrategy. 7ControlPlane:ConstrainedKnowledgeOperations Figure7.1ThecontrolplaneforknowledgeoperationsinValhalla Valhalla’scontrolplanetransformsnatural-languagerequestsintosystemservicessubjectto expliciteligibilityconditions.Itspurposeistomakerequestclassification,accessscope,risk, confirmation,andoutputvalidationexplicit,ratherthantosimulateaformalauthorizationsystem throughnatural-languagerules[22‑23]. 7.1Microkernel-BasedServiceStructure Valhallatreatsthelanguagemodelasaprocessor,thecurrentconversationalcontextasanin- memoryworkspace,andlong-termknowledgeandservicedefinitionsasprogramsanddatain externalstorage.Inspiredbyhierarchicalmemoryandcomputersystems,thisanalogyexplainsthe useofminimalresidentrulesandon-demandserviceloading;itdoesnotimplythatthesystem alreadyprovidesprocessisolationormemoryprotection[6,16,23]. Valhallaisimplementedasanexplicitlyloadableskill,allowingthesamegovernanceinterface tobeusedacrossdifferentLLMagentenvironments.Theimplementationvehicleitselfisnot presentedasaperformancecontribution;itspurposeistokeepsystemstates,serviceentrypoints, andsafetyboundariesreadable. Afterinitialization,top-levelrulesareresponsibleforself-checks,stateidentification,service routing,andbaselinesafetyconstraints.Theyneitherstoretheentiretyofscientificknowledgenor loadallworkflowsatonce.Theirroleapproximatestheminimalresidentinterfaceofamicrokernel: maintainingonlytherulesrequiredforserviceselectionandboundarychecking. Specificcapabilitiesareloadedondemand.Followingoperating-systemterminology,this capability-executionprocessisreferredtoasasystemservice. Asystemservicecanberepresentedas 퐀≝ℛ,퐀,퐀 where: ℛ:Router,forrequestrouting; 퐀:Contract,theoperationalcontract; 퐀:Workflow,theworkflow. Figure7.2ThesystemserviceofValhalla The“minimalresidency+on-demandloading”designreducestheextenttowhichirrelevant rulesoccupythecontextandfacilitatesthelocalizationofservicechanges.However,thecurrent implementationstillreliesoncontextmanagementbytheunderlyingmodelandcannotguarantee thatcriticalconstraintswillalwaysbepreservedintactafterlong-sessioncontextcompression[24]. 7.2Router:RoutingNatural-LanguageRequeststoServices Researcherstypicallyusenaturallanguagetorequestresourceingestion,entitymaintenance, relationshipconstruction,graphorganization,paperwriting,orsystemextension.Althoughintuitive tohumans,theseexpressionsinvolvedifferentobjects,permissions,andwritescopes.TheRouteris responsibleformappingeachrequesttoanexplicitoperationcategory. TheRouterfirstdistinguishesamonggeneralquestionanswering,knowledge-basereading, knowledgeingestion,structuralmaintenance,projectwork,andserviceextension,andthenselects thecorrespondingContract.Ifarequestmatchesmultipleoperationsoritstargetisunclear,the systemshouldexposetheambiguityandrequestclarification,ratherthanautonomouslycombining multiplewritescopes. TheRouterneithermodifiesknowledgenordetermineswhetherarequestisultimately authorized.Itservesastheservice-dispatchentrypoint,reducingtheriskthatthemodelwilldirectly operateonFile,Resource,Entity,Relationship,orGraphobjectsbasedsolelyonsurfacesemantics. 7.3Contract:OperationalEligibilityandServiceContract TheContractfirstdetermineswhetheranoperationmaybeexecutedunderthecurrent conditions.Itthenauthorizestheoperationbyspecifyingtherequiredinputs,readandwrite permissions,risklevel,permittedstates,prerequisites,confirmationrequirements,accessscope, requiredoutputs,andfailure-handlingprocedures.Becausedifferentservicesoperateondifferent objects,theycannotshareavague,genericauthorization.Finally,theContractprovidestheentry pointtotheWorkflow[19,22‑23]. IfaWorkflowdelegatesexecutiontoanotherservice,thetargetContractshouldperformits checksagainratherthaninheritconfirmationfromtheprecedingservice. TheContractconvertsimplicitrequirementsthatdependonthemodel’svoluntarycompliance intoreadableboundaries,allowingreviewerstodeterminewhatthesystemintendstoreadorwrite, whyconfirmationisrequired,andwhatshouldbereturneduponfailure.Nevertheless,itremainsa normativeconstraintratherthananunbypassablesecurityguarantee. 7.4Workflow:ReadableandModifiableExecutionProcedures OnlyaftertheRouterhasclassifiedtherequestandtheContracthaspasseditseligibilitychecks doesthesystemloadtheWorkflow.TheWorkflowspecifiestheorderofreads,registrychecks,state updates,logging,outputvalidation,andconflicthandling,sothattasksofthesametypeneednotbe replannedfromanadhocprompteachtime[18‑19]. Aliterature-ingestionworkflowmayspecifythatthesystemfirstestablishtheidentityofa resource,extractcandidateknowledgeobjects,checkexistingregistrations,writesourcemappings, andrunvalidation.Arelationship-constructionworkflowmayspecifythatthesystemfirstverify objectidentities,selectrelationshiptypes,bindoriginalevidence,andupdatethetaskgraph.The valueoftheseworkflowsliesintheirreusabilityandauditability,notinanyguaranteethatthemodel willneveromitastep. Naturallanguageandlightweightstructuredfileslowerthebarrierforinterdisciplinaryteamsto modifyworkflows,buttheyalsointroduceambiguityandrefactoringrisks.Consequently,critical isolation,transactionalcommits,andpermissionenforcementshouldstillbeimplementedbyexternal orchestratorsorcode-basedmechanismsratherthanrelyingexclusivelyonWorkflowtext[22‑23]. 8Semi-FormalArchitectureoftheControlPlane ℂ 8.1MicrokernelArchitecture ℂ≝퐀,퐀=퐀, 퐀 1 , 퐀 2 , 퐀 3 ,⋯ where: 퐀:thesystemkernelofthecontrolplaneℂ; 퐀:thesetofsystemservicesinthecontrolplaneℂ,퐀=퐀 1 ,퐀 2 ,퐀 3 ,⋯; 퐀:asystemservice. 8.2SystemKernel 퐀 퐀=搀Ѐ퐀, 퐀᐀퐀䠀퐀蠀, 퐀䠀퐀᐀ where: 搀Ѐ퐀 :startupinitialization,implementedinPythonandexecutedasamandatorystep, includingstablefoundationaloperationssuchascheckingsystemintegrity,resettingsystemstates, andloadingtheRouter; 퐀᐀퐀䠀퐀蠀:safetyconstraints; 퐀䠀퐀᐀:globalrules. 8.3SystemState 퐀 andSessionState 퐀 Valhallaemploystwostatemodels:thesystemstate퐀andthesessionstate퐀. 1.SystemState퐀 ThisstatedescribestheoverallstateofValhallaandisusedtomanageoperationalpermissions. 퐀= 搀퐀᐀ Basestate,inwhichbasicoperationsmaybeperformed 퐀㠀瀀퐀Administrativestate,inwhichadministrativeoperationsmaybeperformed 2.SessionState퐀 ThisstatedescribesthecurrentValhallasessionandidentifiesthetargetofsystemservices. 퐀= 퐀㠀퐀᐀ Idlestate;noservicetarget 퐀搀:<퐀搀 퐀瀀᐀>Activestate;servicetarget:<퐀搀 퐀瀀᐀> 8.3SystemService 퐀 퐀≝ ℛ , 퐀 , 퐀 where: ℛ:Router,forrequestrouting; 퐀:Contract,theoperationalcontract; 퐀:Workflow,theworkflow. Thetransformationmappingℳ 퐀 introducedinChapter5isinstantiatedhereasmultiplesystem services퐀. 8.4Request-RoutingComponent:Router TheRoutermapsauser’snatural-languagerequesttotheentrypointofthecorrespondingtype ofoperationalcontract. ℛ≝ 퐀᐀퐀 , 퐀᐀蠀 , 퐀−᐀퐀᐀ where: 퐀᐀퐀:requesttype; 퐀᐀蠀:keywordsthattriggerthecorrespondingRouterentry; 퐀−᐀퐀᐀:theoperational-contractentrypointfortherequest. 8.5OperationalContract:Contract TheContractusesstructuredlanguagetoconstrainworkflowexecution. 퐀≝퐀䠀퐀Ѐ퐀_퐀᐀퐀,퐀᐀퐀,퐀᐀퐀瀀퐀Ѐ퐀,퐀−᐀퐀᐀ where: 퐀䠀퐀Ѐ퐀_퐀᐀퐀:determineswhetherthesystemservicemaybeexecutedunderthecurrent systemstate퐀andsessionstate퐀; 퐀᐀퐀:determinesthetargetofthecurrentsystemserviceaccordingtothecurrentsessionstate 퐀; 퐀᐀퐀瀀퐀Ѐ퐀:specifiesthepermissionsavailabletothecurrentsystemservice,suchaswrite permissionandread-accessscope; 퐀−᐀퐀᐀:theentrypointtotheWorkflow. 8.6Workflow Aworkflowconstructedinnaturallanguage. 8.7SystemService 퐀 1 CallingSystemService 퐀 2 Topreservethesystem’smicrokernelarchitecture,systemservicesshouldbedecoupledfrom oneanother.Forconvenience,however,averysmallnumberofsystemservicesmaycallother systemservices. Undertheglobalrulesin퐀,whensystemservice퐀 1 callssystemservice퐀 2 ,itmustexecute퐀 2 infullandmaynotdirectlyexecute퐀 2 .Thatis, 퐀 1 =ℛ 1 ,퐀 1 ,퐀 1 퐀᐀퐀1 ,퐀 2 ,퐀 1 퐀᐀퐀2 Because퐀 2 =ℛ 2 ,퐀 2 ,퐀 2 ,thisisequivalentto: 퐀 1 =ℛ 1 ,퐀 1 ,퐀 1 퐀᐀퐀1 ,ℛ 2 ,퐀 2 ,퐀 2 ,퐀 1 퐀᐀퐀2 Itisnotequivalenttobypassingthesecondservice’sRouterandContract: 퐀 1 ≠ ℛ 1 , 퐀 1 , 퐀 1 퐀᐀퐀1 , 퐀 2 , 퐀 1 퐀᐀퐀2 8.8IdempotentOperations Valhallasupportsidempotentoperations:repeatedlyexecutingasystemservice퐀doesnot furtherchangeanyknowledgestate퐀. 퐀 퐀= 퐀 퐀 9Human-in-the-LoopGovernedServiceEvolution 9.1Definition Inthispaper,“governedserviceevolution”referstoextendingormodifyingsystemservices, througheithermanualevolutionorself-evolution,afterauseridentifiesaneedforfunctionalchange. Self-evolutionmeansthattheLLMevolvesservicesinaccordancewithpredefinedrules. Thebasicrulesofserviceevolutionareasfollows: 1.퐀followsthefixedarchitectureℛ,퐀,퐀; 2.퐀hasanexplicit ℛ component 퐀᐀퐀 , 퐀᐀蠀 , 퐀−᐀퐀᐀ ; 3.퐀hasanexplicit퐀component퐀䠀퐀Ѐ퐀_퐀᐀퐀,퐀᐀퐀,퐀᐀퐀瀀퐀Ѐ퐀,퐀−᐀퐀᐀; 4.퐀hasanexplicittarget.IfittargetsstructuredknowledgemaintainedbyValhalla,thetargetmust beexplicitlyidentifiedasoneormorelayerswithinthelayeredencapsulation퐀 ∗ ; 5.Theworkflow퐀of퐀isconstructedinnaturallanguagetofacilitatereviewofthesystem’s operatingmechanisms. 9.2EvolutionProcess Figure9.1Thegovernedservice-evolutionprocessinValhalla Aservice-evolutioninstancetypicallycomprisesrequirementclassification,candidate- componentgeneration,boundarychecking,humanconfirmation,registration,validation,and auditing.Unlikeaone-offtask,aservicechangeaffectssubsequentrequestsofthesametype; therefore,completionofthecurrenttaskcannotserveasthesolecriterionforsuccessfulregistration. Researchonskilllifecyclesprovidesareferenceforcreatingandrevisingexternalizedcapabilities [10‑14],whileValhallafurtheremphasizestheobjectsbeingchangedandthescopeofknowledge writes. 9.3GovernanceBoundaries Theprincipalrisksofserviceevolutionincludeimplicitinheritance,permissionexpansion, cross-layerwrites,andinsufficientvalidation.Valhallaadoptscomponentindependence,explicit scope,non-inheritanceofconfirmation,anddiagnosablefailureasdesignconstraints.These constraintscanbecomeenforceableguaranteesonlywhensupportedbyexternalvalidators,tests,and orchestrationmechanisms;whenspecifiedsolelyinaContractorWorkflow,theymaystillbe misinterpretedbythemodel[22‑23]. Designconstraintsforgovernedserviceevolution ThesourcemappingofanEntitymustberetainedthrougharegisteredResourceandmustnotbeoverwrittenwithcontent lackingprovenance. EachnewwritableWorkflowmustbeboundtoanexplicitContract,withitspermissionsandrisksdeclaredseparately. Theoutputofaone-offtaskmustnotberegistereddirectlyasalong-termsystemservicewithoutreview. Cross-layereffectsmustbedeclaredexplicitly;operationsrequiringconfirmationmustnotinheritauthorizationfromother stages. Serviceextensionsshouldrecordchangesandretainpathsfortermination,correction,andrecoveryintheeventofvalidation failure. 9.4Example:Addingan“Experimental-ProtocolReview”Service ThebasicversionofValhalladoesnotincludeanexperimental-protocolreviewsystemservice; usersmustevolvesuchaserviceaccordingtotheirownrequirements. Supposearesearchgroupsubmitsthefollowingrequest:“Weneedanexperimental-protocol reviewworkflowthatcheckswhetheraprotocolcoversthekeyvariables,whetheritsmethodological choicesareconsistentwithexistingevidence,andwhetheritcontainsunverifiedassumptions.”One practicalapproachistouseanLLMtoformulateacorrespondingreviewprocedure,ortoadoptan establishedprocedureandhaveValhallaevolveinaccordancewithit.Inthisway,theresulting Valhallasystemservicecanuseboththeestablishedprocedureandthestructuredknowledge governedbyValhalla.Valhallafirstclassifiestherequestasacapabilitychangeratherthanone-off textgeneration. ThesystemthendeterminesthatthiscapabilityprimarilyreadsEntity,Relationship,andGraph objectsandmayciteResourceobjectsasevidencesourceswhennecessary,butshouldnotdirectly modifyoriginalmaterialsintheFilelayer.ItaccordinglyaddsaRouterentrysothatsimilarrequests canenterthe“Experimental-ProtocolReview”service.ItthendefinesaContractspecifyingthatthe servicemayreadtheEntity,Relationship,andGraphobjectsrelevanttothecurrentprojectandmay outputareviewreportandalistofhypothesesrequiringvalidation,butmaynotmodifythestatesof existingrelationshipswithoutconfirmation.Finally,itcreatesaWorkflowthatspecifiesthereview steps,includingidentifyingexperimentalobjectives,mappingrelevantentities,examining relationshipchains,locatingevidencegaps,generatingriskalerts,andrecordingthereviewlog. Thisexampledemonstratesthatserviceevolutionoperatesonidentifiablecomponentsrather thanonthesystemasawhole.Whetherthenewcapabilitycanbeexecutedreliablystilldependson testing,inputmaterials,anddomainreview;generatingtheservicefilesalonedoesnotconstitute evidenceofthecapability’seffectiveness. 9.5ComplementarityofAutomaticGenerationandManualMaintenance ResearchersmayalwaysaddormodifyRouter,Contract,andWorkflowcomponentsmanually. Thevalueofnaturallanguageandlightweightstructuredrepresentationsliesinenablingdomain specialiststoparticipateindefiningknowledgeobjectsandreviewingprocesses,ratherthanplacing systemmaintenanceentirelyinthehandsofthemodel. Modelgenerationcanreduceinitialsetupcosts,whereasmanualmaintenancesupportsfine- grainedadjustmenttospecificrequirements. Inonecaseinvolvingtheextensionofaliterature-ingestionsystemserviceforantibodydesign,a candidateservicetriggeredasafetyreviewbecauseofpotentiallyhigh-riskbiologicalinformation. Subsequentmanualrevisionrestrictedthescopetoreview-levelbackgroundinformation,conceptual conclusions,evaluationmetrics,andnon-operationalriskdescriptions,whileexcludingexecutable wet-labproceduresandhigh-riskoptimizationpathways. Thisprocessillustratesthathumaninterventioncannarrowserviceboundaries,butitneither demonstratesthatthesafetyclassificationsoftheunderlyingmodelarestablenorconstitutesa generalcomplianceguaranteeforallbiomedicaltasks. 10CaseStudy:AKnowledge-GovernedScientificWritingWorkflow BasedonValhalla 10.1ExperimentalObjectives Asdiscussedabove,existingLLMagentshavedemonstratedthatstructuredknowledge representations,tooluse,andtaskplanningcanimprovetheautomationofcomplexscientifictasks [15,20,25‑27].Inrealresearchteams,however,long-termscientificworkrequiresmorethanthe completionofindividualtasks:italsorequiresthecontinualaccumulationofdomainknowledge, maintenanceofevidentialrelationships,andongoingadaptationofanalyticalworkflowsasresearch directionsevolve. Valhallaisnotintendedtore-establishthevalueofstructuredknowledgeforscientificagents. Rather,itexplores,onthatbasis,along-termintelligentworkenvironmentforresearchteams.This casestudythereforeexaminesthreecoredesignhypothesesunderlyingValhalla: (1)Structuredknowledgestatescantransformscientificknowledgefromone-offdocument inputsintoknowledgeassetsthatcanbecontinuouslymaintainedandreused. (2)ARouter–Contract–Workflowservicearchitecturecanenableresearchteamstoextend domain-specifictaskcapabilitieswithalowbarriertoentry. (3)Agovernedevolutionmechanismcanpreservethetransparencyandmaintainabilityofan LLM-agentsystemthroughoutlong-termuse. Toexaminethesehypotheses,weselectedacross-domainscientificwritingtaskasthecase studyandevaluatedValhalla’sintegratedsupportforknowledgeingestion,systemevolution,andthe generationandreviewofscientifictext. 10.2ExperimentalTaskandWorkflowDesign WeusedPVRIG/antibodyreviewgenerationastheexperimentalcase[28‑29]. Thetaskspans: tumorimmunologymechanisms; immune-checkpointnetworks; antibodyengineeringdesign; optimizationofFcfunctions;and translationalevaluation. Thetopicencompassesmolecularmechanisms,biomedicalknowledge,andengineeringdesign strategiesandisthereforerepresentativeofcross-domainscientificwriting[28‑29]. Theexperimentcomprisedthreestages: 1.evaluationofdomain-knowledgeingestionandsystem-evolutioncapabilities; 2.evaluationofknowledge-assetsharingandstructuredorganization;and 3.evaluationofknowledge-state-basedsupportforscientificwriting. 10.2.1EvaluationofValhalla’sSystem-EvolutionCapability Becauseresearchdomainsdifferintheirknowledge-organizationpracticesandevaluation criteria,general-purposescientificagentsoftencannotdirectlyaccommodatetherequirementsofa newdomain.WethereforefirstexaminedwhetherValhallacouldextenditssystemservicesin responsetodomain-specificrequirements. Webeganbydevelopingadomain-knowledgeingestionspecificationtailoredtopaperson antibodydesign.Thespecificationdefined: howknowledgeentitiesshouldbeorganized; whichexperimentalinformationshouldberetained; requirementsforrecordingevidence;and theboundariesofdomain-relevantknowledge. ThedomainspecificationwasthenprovidedtoValhalla,whichuseditsbootstrappedevolution mechanismtogenerateanewdomain-knowledgeingestionservice. Theprocesscomprised: 1.generatinganewservicedesignfromthedomainrequirements; 2.creatingacorrespondingRouterentrysothatthesystemcouldrecognizeingestiontasksinthis domain; 3.definingaContractthatconstrainedtheservice’sinputs,outputs,andscopeofknowledge modification;and 4.writingaWorkflowthatspecifiedpaperparsing,Entityextraction,evidencemapping,andlogging. Thisprocessproducedadedicatedknowledge-ingestionservicefortheantibody-designdomain: antibody_design_ingest. Theservicewassubsequentlydistributedtotheresearchersparticipatingintheexperiment, enablingdifferentmemberstoingestpapersfromthedomaininparallel. Duringuse,somematerialstriggeredtheLLM’ssafetyrestrictions.Inresponse,theresearchers manuallyevolvedtheservicebymodifyingitsWorkflowandaddingdomain-specificsafety constraints.Therevisedservicewaslimitedtoconceptualknowledge,evaluationmetrics,research background,andpubliclyreportedexperimentalresults,whileexcludinghigh-riskoperational contentunsuitableforautomatedgeneration. Ingestionoftheremainingmaterialscontinuedaftertheseconstraintshadbeenadjusted. 10.2.2EvaluationofValhalla’sKnowledge-AssetSharingCapability ToexaminewhetherValhallacouldsupportteam-levelknowledgesharing,weintegrated knowledgeobjectsindependentlyingestedbydifferentteammembers. Theresultingknowledgebasecontained: 26paperResources; 80Entities;and 92Relationships. Onaverage,eachpaperyieldedapproximatelythreeknowledgeentities. Valhalla’srelationship-constructionservicewasthenusedtoorganizethesemanticrelationships amongEntities,including: mechanisticrelationships; methodologicaldependencies; evidentialsupportrelationships;and method-evaluationrelationships. Thisprocessproducedaknowledgegraphfortheresearchtopic. Thepurposeofthisstagewastoexaminewhetherknowledgeproducedbydifferentteam memberscouldbeseparatedfromtheirindividualcontextsandtransformedintoashared, transferablestructuredknowledgestate. 10.2.3EvaluationofValhalla’sSupportforScientificWriting Toassesstheeffectofstructuredknowledgestatesonscientificwriting,wedesignedtwo comparativeworkflows. Baselineworkflow 1.CleartheLLMagent’sconversationhistory. 2.Conductthereview-writingtask: (1)loadtheacademic-paper-orchestratorskill; (2)usethe26ingestedpapersasreferences;and (3)useanexistingreviewasastructuraltemplatetogeneratePVRIGAntibodyDesignReview1. 3.ReviewthepaperusingValhalla’spaper_reviewservice. 4.Revisethepaperusingthepaper_polishserviceonthebasisofthereviewresults. ThefinaloutputwasPVRIGAntibodyDesignReview1_RevisedandPolished. Valhallaworkflow 1.CleartheLLMagent’sconversationhistory. 2.Conductthereview-writingtask: (1)startValhalla’spaper_orchestratorservice,asystemservicethatreproducestheacademic- paper-orchestratorskillwhileprovidingaccesstotheknowledgebase; (2)accessthecompletedknowledgebase;and (3)usethesamereviewtemplatetogeneratePVRIGAntibodyDesignReview2. 3.Reviewthepaperusingthepaper_reviewservice. 4.Revisethepaperusingthepaper_polishserviceinconjunctionwiththeknowledgebase. ThefinaloutputwasPVRIGAntibodyDesignReview2_RevisedandPolished. 10.3ExperimentalResults 10.3.1ExecutionTime Thetimerequiredbythetwoworkflowsisshownbelow: ExperimentalstepDuration(min) Writing147.5 Review13.5 Revision112.5 Writing210 Review211.5 Revision210.5 Theinitialgenerationtimewassubstantiallylowerinthesecondscientificwritingworkflow,for whichanexistingknowledgestateandcorrespondingsystemserviceswereavailable. 10.3.2EvaluationofPaperQuality 1.Thegeneratedoutputswereevaluatedbydomainexperts. TheevaluationindicatedthatthereviewgeneratedwiththeValhallaknowledgebaseprovided morecomprehensiveknowledgecoverage,morespecificreferencestoexperimentalresults,and morein-depthdiscussionofmechanisms. Comparedwiththeoutputgeneratedsolelythroughthepaper-writingworkflow,theValhalla- supportedtextdrewmoreextensivelyonexperimentaldataandresearchconclusionsfromthesource papers. 2.ChatGPTwasadditionallyusedforanauxiliaryevaluation,withthefollowingresults: ExperimentalstepDuration(min) Writing147.5 Review13.5 Revision112.5 Writing210 Review211.5 Revision210.5 Detailedevaluation: EvaluationdimensionReview1 Revised Review1 Review2 Revised Review2 Depthofscientificcontent(30)26272829 Logicalstructureofthereview(25)20232324 Rigorofevidence(20)15191819 Expertiseinantibodyengineering(15)12121414 Innovativeframingandpublicationpotential(10)9988 Totalscore82909194 10.4AnalysisofResults TheresultssuggestthatValhalla’sprincipaladvantagedoesnotarisefromanimprovementin one-offtextgeneration.Instead,itstemsfromthelong-termorganizationofscientificknowledge states,operationalworkflows,andsystemcapabilities. First,withrespecttoknowledgestate,Valhallatransformssourcepapersintostructured knowledgeobjectscontainingentities,relationships,andevidencemappings,therebyreducingthe scientificwritingprocess’srelianceonone-offcontextinputs. Intheconventionalworkflow,theLLMmustrepeatedlyidentifyconceptualrelationshipsand experimentalevidenceacrossalargecollectionofpapers.IntheValhallaworkflow,relevant knowledgehasalreadybeenorganizedthroughEntitiesandRelationships,allowingthemodelto constructargumentsaroundstableknowledgeobjects.Thegeneratedoutputwasconsequently strongerinitsuseofexperimentalevidence,explanationofmechanisms,andstructuralorganization. Thisfindingsuggeststhattheprincipalbottleneckforascientificagentisnotlimitedto language-generationcapability;italsoconcernswhethertheagentpossessesaknowledgestatethat canbemaintainedovertime. Second,withrespecttotheextensionofsystemcapabilities,theexperimentdemonstratedthat Valhalla’sservice-evolutionmechanismcansupportdomainadaptation. Conventionalagentsystemsgenerallyrelyontoolsandworkflowspredefinedbydevelopers. Movingintoanewresearchdomainthereforerequirestheredevelopmentofsystemfunctionality. ValhallainsteadencapsulatessystemcapabilitiesasRouter,Contract,andWorkflowcomponents, allowingresearcherstogeneratenewservicestructuresinresponsetodomainrequirementsandto adjusttheirbehavioralboundariesthroughhumanreview. Theobjectsofsystemevolutionarethereforenotopaquesoftwarecode,butreadable,modifiable, andgovernableservicecomponents. Third,withrespecttoscientificwritingquality,theresultsindicatethatsupportfromthe knowledgebaseprimarilyimprovedthedepthofscientificcontent,logicalstructure,andrigorof evidenceratherthanmerelyimprovinglinguisticexpression. Thisresultsuggeststhatstructuredknowledgestatesbenefitscientificagentsbyreducing knowledgelossandbreaksinevidentialchains,therebyenablingthemodeltoorganizenew scientificnarrativesonthebasisofexistingresearchrelationships. Finally,theexperimentillustrateshowValhalladiffersfromconventionalRAGsystemsand single-taskagents. Conventionalmethodsprimarilyoptimizeretrievalandgenerationwithinanindividualtask, whereasValhallafocusesonthejointaccumulationofknowledge,taskworkflows,andsystem capabilitiesoverthecourseoflong-termscientificwork.Theknowledgebase,domain-ingestion service,andpaper-reviewworkflowproducedinthisexperimentcanallbereusedinsubsequent tasks;thesystem’svaluemaythereforeincreasewithcontinueduse. 10.5EfficiencyAnalysis Inadditiontoimprovementsinpaperquality,theexperimentrevealedchangesintheefficiency ofscientific-taskexecution. Comparedwiththebaselinegenerationworkflow,theworkflowsupportedbyValhalla’s knowledgestateandsystemservices: reducedinitialgenerationtimebyapproximately79%;and keptasinglescientificwritingiterationtoapproximately30minutes. Itshouldbenotedthatconstructingtheknowledgestateincurredadditionalcosts.Initial knowledgeingestionrequiredapproximately300minutesinthisexperiment;thecompleteworkflow maythereforeoffernotimeadvantageifonlyonetaskisperformed. Forlong-termscientificwork,however,knowledgeingestionconstitutesaone-timeinvestment inassetconstruction.Whenthesameknowledgebaseissubsequentlyusedforpaperwriting, experimentaldesign,reviewupdates,orprojectdiscussions,theinitialcostcanbeamortizedacross multipletasks. Valhalla’sefficiencyadvantagethereforedoesnotarisefromacceleratingasingletask,butfrom reusingknowledgeandworkflowsthroughoutalong-termscientificprocess. 11Limitations,FailureModes,andGovernanceBoundaries ThecurrentevidenceforValhallaisderivedprimarilyfromitsarchitecturaldescriptionanda singlecasestudy.Itslimitationsincludenotonlyinsufficientinternalandexternalvalidity,butalso engineeringrisksassociatedwithknowledgequality,natural-languageconstraints,servicecoupling, externalsafetypolicies,stateconsistency,andcontextmanagement.Theseissuesmustbestated separatelyfromthedesignintenttoavoidpresentingnormativerulesasruntimeguarantees. 11.1ScopeofEvidenceandInternalValidity ThecasestudyinSection10includednorandomization,independentreplication,strong baselineunderamatchedresourcebudget,componentablation,orauditableexpertevaluation.The model-assistedscoringalsodidnotdisclosethecompletejudgeconfigurationortestsforbias[30]. Thedifferencesintimeandscoresthereforecannotsupportcausalattribution,andD1–D3havenot yetbeenempiricallyvalidated.Futureworkwillevaluatethemthroughlarger-scaleandmore rigorousexperiments. Constructingaknowledgestaterequiresresourceregistration,provenancemapping,object extraction,relationshipconstruction,andqualitycontrol.Theseinitialcostsmaynotberecoveredin one-offorinfrequenttasks.Forlong-termprojects,thebenefitsofreusemustbemeasuredthrough ingestion,maintenance,retrieval,andhandoverrecordscollectedatmultipletimepointsratherthan inferredsolelyfromthesystem’sdesignforreusability. 11.2RisksArisingfromtheLLM’sUnderlyingRules Oneissueidentifiedtodateconcernsbatchingestion.EvenwhenaWorkflowrequires documentstobeprocessedindividually,themodelmaycombinemultipledocumentsinasingle analysis,therebymixingmethods,conditions,andconclusionsacrosssources.Themodelmay interpretsuchcombinationasanefficiencyoptimization,butitunderminesevidencemappingin whicheachindividualsourceservesastheboundary. 11.3Natural-LanguageGovernanceandServiceCoupling Router,Contract,andWorkflowarerepresentedinnaturallanguageandlightweightstructured formats.Althoughthismakesthemreadabletodomainresearchers,italsopermitsthemodelto reinterpretorreconstructtherules.Themodelmayimplementaninstructionto“refertoaservice” bydirectlyinheritingorinvokingthatservice,therebycreatingimplicitdependenciesbetween servicesthatareintendedtoremainindependent. Suchcouplingmayhavenoimmediateeffectonexecution,butitexpandsthescopeof subsequentmodifications:whentheoriginalserviceisupdated,thebehaviorofthenewservicemay changeaccordingly,eventhoughthedependencyhasnotbeenexplicitlyregistered.Similarrisks applytothetransferofpermissionsandconfirmations;thequalificationsofoneservicemustnotbe inheritedbyasubsequentservicebydefault. Thefollowingboundarymustthereforebemadeexplicitwheneveraserviceisaddedor modified: Anewservicemayrefertothestructureanddesignpatternsofanexistingservice,butitmust notinherit,invoke,ordependontheexistingservice’sspecificimplementationunlessthisisdeclared. Alldependencies,read/writescopes,andconfirmationconditionsmustberegisteredandverified separately. 11.4ExternalSafetyPoliciesandPartialCommits Valhallarunsongeneral-purposelanguagemodelsandtheirplatformsandistherefore constrainedbybothitsownservicecontractsandtheunderlyingmodel’ssafetypolicies.Tasks involvingdrugdiscovery,biomedicalanalysis,orexperimentaldesignmaytriggerplatform-level safetyreview. Platformreviewisnecessarytoreducehigh-riskuses,butitsdecisionsoccuroutsideValhalla’s governancesystem.Evenafterataskhaspassedinternalpermissionchecks,theunderlyingmodel maystillrefusetocontinue,resultinginfailure. Valhallamustnotbypasstheunderlyingsafetymechanisms.Amorepracticalapproachistouse stagedcommits,failuremarkers,humanreview,andrecoverabletransactions,whileclearly distinguishingamonga“platformrefusal,”a“servicefailure,”anda“committedknowledgestate.” Thecurrentimplementationdoesnotyetprovidethesemechanismsinfull. 11.5StateConsistency,Snapshots,andRollback Resources,knowledgeobjects,relationships,graphs,andservicespersistasdirectlyeditablefile states.Thislowersthebarriertohumanreviewbutalsocreatestangiblerisksofpartialcommits, concurrentoverwrites,andinadvertentmodification.Thecurrentversiondoesnotyetbindevery operationtoanautomaticsnapshot,transactionalcommit,andverifiablerollback. Atpresent,recoveryreliesprimarilyonlocalbackupsandmanualrestoration.Futureversions mustestablishchangesets,pre-commitvalidation,atomicreplacement,failurerollback,andrecovery exercisesforhigh-riskwrites,whilerecordingwhomodifiedwhichobjects,whenthemodifications occurred,andonthebasisofwhichconfirmation.Withoutthesemechanisms,“governable”can signifyonlythattheprocessisvisible;itcannotbeequatedwithtransactionalsafety. 11.6ContextManagement,EvaluationReliability,andExternalValidity LoadingRouter,Contract,andWorkflowinstagescanreducetheamountofirrelevantmaterial occupyingthecontext,butthecurrentversionhasnoindependentcontext-schedulinglayer.The systemcannotpreciselydeterminewhenparticularrules,resources,confirmations,andtaskstates shouldbeloaded,compressed,evicted,orrestored[24]. Long-runningtasksstilldependontheunderlyingmodel’scontextwindowandautomatic compression.Earlierconstraintsmaybesummarized,provenanceidentitiesmaylosedetail,and stage-specificconfirmationsmaybecomedetachedfromthecurrentoperation.Becauseinternal platformcompressionpoliciesaregenerallyunobservable,thecausesoffailurecanbedifficultto reconstruct. Anindependentcontext-managementlayerrequires,atminimum,thefollowingcapabilities: distinguishingsystemrules,taskstates,knowledgeobjects,andtemporaryreasoningresults; assigningpriorities,lifecycles,andloadingconditionstodifferentcontextobjects; loadingtheContract,resources,andexecutionrecordsrequiredforeachWorkflowstage; persistingstructuredtaskstatesbeforecompressionratherthanrelyingentirelyonnatural- languagesummaries; markingcriticalconstraints,userconfirmations,andprovenanceidentitiesascontentthatmust beretainedorreloaded; evictingirrelevantcontentandloadingthestaterequiredforthenextstagewhenthetask transitionsbetweenstages;and recordingtherulesandresourcesactuallyloadedateachsteptosupportexecutiontracingand erroranalysis. Inaddition,themodel-assistedevaluationinthecurrentcasestudyincludednoindependent blindedassessment,repeatedjudges,biastesting,ormeasuresofinter-rateragreementandtherefore cannotserveasastablemeasureofquality[30].Externalvalidityisalsolimitedbytheuseofa singletopic,asingleteam,andasinglerun.Futureresearchshouldjointlyevaluateknowledge quality,taskperformance,governancefailures,long-termcost,andcross-membertransfer,and shoulddiscloseconfigurationsandrawrecordsinaformthatpermitsindependentreview. 12Conclusion ThispaperhaspresentedValhalla,alayeredknowledge-stateandservice-governanceframework forlong-termscientificknowledgework.ThedataplaneusesFile,Resource,Entity,Relationship, andGraphtodistinguishdocumententrypoints,stableidentities,knowledgeobjects,semantic judgments,andtask-orientedviews,respectively.ThecontrolplaneusesRouter,Contract,and Workflowtoorganizerequestclassification,eligibilitychecks,andconstrainedexecution. Theframework’sprimarycontributionisalayeredarchitecturalprotocolthataddressesa missingdimensionofcurrentstructured-knowledgesystems.Thisprotocolmakesstructured knowledgeashareableandrecomposableasset,extendingLLM-assistedscientificcollaborationfrom thesharingofagentstothesharingofstructuredknowledge. ThePVRIGwritingcaseinSection10showsthattheprototypecanconstructlayered knowledgestatesandinvokerelevantservices,butitstimeandqualityscoresarederivedfroma single,independentlyunauditedworkflow.Theavailableevidencedoesnotdemonstratethat ValhallaoutperformsRAG,GraphRAG,generalmemorysystems,orworkflow-onlymethods,nor doesitestablishlong-termreusebenefitsortherealizationofD1–D3[4‑6,18‑19]. Futureworkshouldconductmulti-taskcomparisons,componentablations,independent replications,blindedexpertevaluations,knowledge-qualityaudits,faultinjection,andlongitudinal costmeasurementsunderafrozenprotocol.Itshouldalsoprovideindependentlyreviewablecode, configurations,corpusinventories,andrawevaluationrecords.Onlyaftertheseevaluationshave beencompletedwillitbepossibletodeterminewhetherlayeredknowledgestatesandservice governanceconstituteageneralizableinfrastructureforscientificagents. References [1]HOGANA,BLOMQVISTE,COCHEZM,etal.KnowledgeGraphs[J].ACMComputingSurveys,2021,54(4):1-37. DOI:10.1145/3447772. [2]WILKINSONMD,DUMONTIERM,AALBERSBERGIJ,etal.TheFAIRGuidingPrinciplesforScientificData ManagementandStewardship[J].ScientificData,2016,3:160018.DOI:10.1038/sdata.2016.18. [3]MOREAUL,MISSIERP,eds.PROV-DM:ThePROVDataModel[EB/OL].W3CRecommendation,(2013-04- 30)[2026‑07‑25].https://w.w3.org/TR/2013/REC-prov-dm-20130430/. [4]LEWISP,PEREZE,PIKTUSA,etal.Retrieval-AugmentedGenerationforKnowledge-IntensiveNLP Tasks[C/OL]//AdvancesinNeuralInformationProcessingSystems33.RedHook:CurranAssociates,2020:9459- 9474[2026‑07‑25].https://proceedings.neurips.c/paper/2020/hash/6b493230-Abstract.html. [5]EDGED,TRINHH,CHENGN,etal.FromLocaltoGlobal:AGraphRAGApproachtoQuery-Focused Summarization[EB/OL].(2025-02-19)[2026‑07‑25].https://arxiv.org/abs/2404.16130.DOI: 10.48550/arXiv.2404.16130. [6]PACKERC,WOODERSS,LINK,etal.MemGPT:TowardsLLMsasOperatingSystems[EB/OL].(2024-02- 12)[2026‑07‑25].https://arxiv.org/abs/2310.08560.DOI:10.48550/arXiv.2310.08560. [7]PARKJS,O’BRIENJC,CAICJ,etal.GenerativeAgents:InteractiveSimulacraofHuman Behavior[C]//Proceedingsofthe36thAnnualACMSymposiumonUserInterfaceSoftwareandTechnology.New York:ACM,2023:Article2,1-22.DOI:10.1145/3586183.3606763. [8]GUTIÉRREZBJ,SHUY,GUY,etal.HippoRAG:NeurobiologicallyInspiredLong-TermMemoryforLarge LanguageModels[C]//AdvancesinNeuralInformationProcessingSystems37.2024:59532-59569.DOI: 10.52202/079017-1902. [9]GUTIÉRREZBJ,SHUY,QIW,etal.FromRAGtoMemory:Non-ParametricContinualLearningforLarge LanguageModels[C/OL]//Proceedingsofthe42ndInternationalConferenceonMachineLearning.PMLR,2025,267: 21497-21515[2026‑07‑25].https://proceedings.mlr.press/v267/gutierrez25a.html. [10]WANGG,XIEY,JIANGY,etal.Voyager:AnOpen-EndedEmbodiedAgentwithLargeLanguageModels[J/OL]. TransactionsonMachineLearningResearch,2024[2026‑07‑25].https://openreview.net/forum?id=ehfRiF0R3a. [11]ZHOUH,GUOS,LIUA,etal.Memento-Skills:LetAgentsDesignAgents[EB/OL].(2026-03-19)[2026‑07‑25]. https://arxiv.org/abs/2603.18743.DOI:10.48550/arXiv.2603.18743. [12]YANGY,LIJ,PANQ,etal.AutoSkill:Experience-DrivenLifelongLearningviaSkillSelf-Evolution[EB/OL]. (2026-03-05)[2026‑07‑25].https://arxiv.org/abs/2603.01145.DOI:10.48550/arXiv.2603.01145. [13]LINH,LIP,SONGJ,JIANGF,ZHANGT.MUSE-Autoskill:Self-EvolvingAgentsviaSkillCreation,Memory, Management,andEvaluation[EB/OL].(2026-07-03)[2026‑07‑25].https://arxiv.org/abs/2605.27366.DOI: 10.48550/arXiv.2605.27366. [14]ZHOUC,CHAIH,CHENW,etal.ExternalizationinLLMAgents:AUnifiedReviewofMemory,Skills,Protocols andHarnessEngineering[EB/OL].(2026-04-09)[2026‑07‑25].https://arxiv.org/abs/2604.08224.DOI: 10.48550/arXiv.2604.08224. [15]GOTTWEISJ,WENGWH,DARYINA,etal.AcceleratingScientificDiscoverywithCo-Scientist[J].Nature,2026, 655:487-496.DOI:10.1038/s41586-026-10644-y. [16]MIY,GAOZ,MAX,LIQ.BuildingLLMAgentsbyIncorporatingInsightsfromComputerSystems[EB/OL]. (2025-04-06)[2026‑07‑25].https://arxiv.org/abs/2504.04485.DOI:10.48550/arXiv.2504.04485. [17]ZHUD,ZHENGLN,CHENZ.FundaPod:AMulti-PersonaAgentPodPlatformwithKnowledgeGraphMemory forAI-AssistedFundamentalInvestmentResearch[EB/OL].(2026-06-18)[2026‑07‑25]. https://arxiv.org/abs/2605.27864.DOI:10.48550/arXiv.2605.27864. [18]SOUZAR,POTEETT,ETZB,etal.LLMAgentsforInteractiveWorkflowProvenance:ReferenceArchitectureand EvaluationMethodology[C]//ProceedingsoftheSC’25WorkshopsoftheInternationalConferenceforHigh PerformanceComputing,Networking,StorageandAnalysis.NewYork:ACM,2025:2257-2268.DOI: 10.1145/3731599.3767582. [19]JONNALAGEDDAP,YAOY,GAOX,etal.ExecutableSchemaContracts:FromAutomaticIngestiontoMulti- SourceRetrieval[EB/OL].(2026-06-03)[2026‑07‑25].https://arxiv.org/abs/2606.05415.DOI: 10.48550/arXiv.2606.05415. [20]WANGZ,CHENZ,YANGZ,etal.EmpoweringBiomedicalEvidenceExplorationandSynthesiswithDeep KnowledgeGraphResearch[J].NatureMachineIntelligence,2026,8:1142-1156.DOI:10.1038/s42256-026-01266-0. [21]OBSIDIAN.GraphView[EB/OL].[2026‑07‑25].https://help.obsidian.md/plugins/graph. [22]TABASSIE.ArtificialIntelligenceRiskManagementFramework(AIRMF1.0)[R].Gaithersburg:NationalInstitute ofStandardsandTechnology,2023.DOI:10.6028/NIST.AI.100-1. [23]SALTZERJH,SCHROEDERMD.TheProtectionofInformationinComputerSystems[J].Proceedingsofthe IEEE,1975,63(9):1278-1308.DOI:10.1109/PROC.1975.9939. [24]LIUNF,LINK,HEWITTJ,etal.LostintheMiddle:HowLanguageModelsUseLongContexts[J].Transactionsof theAssociationforComputationalLinguistics,2024,12:157-173.DOI:10.1162/tacl_a_00638. [25]BRANAM,COXS,SCHILTERO,etal.AugmentingLargeLanguageModelswithChemistryTools[J].Nature MachineIntelligence,2024,6:525-535.DOI:10.1038/s42256-024-00832-8. [26]BOIKODA,MACKNIGHTR,KLINEB,etal.AutonomousChemicalResearchwithLargeLanguageModels[J]. Nature,2023,624:570-578.DOI:10.1038/s41586-023-06792-0. [27]LUC,LUC,LANGERT,etal.TheAIScientist:TowardsFullyAutomatedOpen-EndedScientific Discovery[EB/OL].(2024-09-01)[2026‑07‑25].https://arxiv.org/abs/2408.06292.DOI:10.48550/arXiv.2408.06292. [28]WHELANS,OPHIRE,KOTTURIMF,etal.PVRIGandPVRL2AreInducedinCancerandInhibitCD8+T-Cell Function[J].CancerImmunologyResearch,2019,7(2):257-268.DOI:10.1158/2326-6066.CIR-18-0442. [29]XUEH,ZHANGZ,LIL,etal.CharacterizationofaNovelAnti-PVRIGAntibodywithFc-CompetentFunctionThat ExertsStrongAntitumorEffectsviaNKActivationinPreclinicalModels[J].CancerImmunology,Immunotherapy, 2024,73(5):81.DOI:10.1007/s00262-024-03671-z. [30]WANGP,LIL,CHENL,etal.LargeLanguageModelsAreNotFairEvaluators[C]//Proceedingsofthe62nd AnnualMeetingoftheAssociationforComputationalLinguistics:Volume1,LongPapers.Bangkok:Associationfor ComputationalLinguistics,2024:9440-9450.DOI:10.18653/v1/2024.acl-long.511.