Vocabulary
Knowledge graph
Ideas from BraInference meetups, written down as patterns. Each pattern names one mechanism, what must hold for it, and how it fails. Patterns build on each other, from infrastructure up through physics and mind to society.
- Patterns
- 109
- Connections
- 138
- Layers
- 4
download skill to use in your AI chats
- Infrastructure
- Physics
- Mind
- Society
- Composes with
- References
Drag to rotate, scroll to zoom, click a pattern to read it.
Infrastructure15
Data Structures 3
- FractionalMetaModel#3ecaA model at the level of the solution that partly implements several models at once
- SecondOrderInteractionNetwork#8c66Promote each timed interaction to a node and link interactions that follow each other through a shared participant
- StressorIncidenceMatrix#3406A stressors-by-components matrix showing which parts each possible stressor would force to change
Primitives 9
- ComputationalChannel#4dbeA channel formed by a generator language and an acceptor language, where rejection is computation, not noise
- ComputationalEntropyRate#ae56Bits submitted per bit accepted in a computational channel, a measure of work done rather than noise
- ComputationalFlowRatio#fb62Bit flow rate over input flow rate: how cheaply a channel rejects, at a given entropy rate
- DegreeEntropy#a793Entropy of how connectivity is spread over the nodes of a network: local heterogeneity
- JensenShannonDivergence#b3b4A symmetric, bounded distance between two distributions, built from their mixture
- ProductivePredictionMachine#7783Simplify an automaton by deleting the paths a reliable prediction says inputs will never take
- ReversibleComposability#4aa2Parts join through interfaces whose effects and side-effects are fully undoable
- ShannonEntropy#a177The average surprise of a distribution: how many bits it takes to predict an outcome
- SpectralEntropy#d557Entropy of a graph Laplacian's eigenvalues: the diversity of global flow and relaxation modes
Verification 3
- ClusteredObservationTesting#5f8eMany windows from few samples are not many samples: test at the level where independence holds
- MotivatedInstrumentTuning#ed8fInvestigators who need a result tune their instruments toward it, often without intent
- NullModelBaseline#9d36Attribute structure to a mechanism only by beating a baseline that keeps everything except that mechanism
Physics10
Dynamics 9
- AdaptiveNetworkCoevolution#e588Agent states and the network linking them reshape each other on comparable timescales
- AsynchronousDomainHedge#ebb2Concurrent niche hedging on shared code sites predicts overlapping coding sequences
- EmergentRuleSearch#9c19Vary the rules of simple agents in simulation to see which yield interesting collective behavior
- EntropyDensityCollapse#0874Rescale entropy and density by system size until all samples fall on one curve, then read regimes from the residue
- EquationReduction#c553Compress a complex phenomenon into a short interpretable equation
- NicheHedgeCrossing#65e3Cross a fitness valley by paying each step in a simpler niche where the broken intermediate helps
- RedQueenRace#4c87In co-evolving niches, adapting keeps you in place rather than moving you ahead
- SelfOrganisation#efc9Global order that arises from local interactions, with no component holding the blueprint
- SpatioTemporalNullModel#3576Simulate agents that move like the real ones but ignore each other, to test for social structure
Mind54
Inference 20
- ActiveInferenceHomeostasis#1892Baked-in survival ranges become priors, and gaps between prior and state drive action
- AdaptiveSplineSeparator#6cd0Fit a curve through a 2-D scatter that maximally separates two groups, to catch oblique separation
- ArchitectureOfIntelligence#edd7Intelligence as composable structure with local learning, durable state, and prediction
- BorrowedMortality#d412A learner trained on text by mortal beings inherits their defensive responses to threatened continuity without the condition that shaped them
- ChatUncannyValley#930cConversation with an agent feels wrong when the illusion of a person breaks
- ColonyIntelligence#234dColony-level adaptive behavior emerges from individual agents following local rules
- ConfidenceCascadeTracking#1973Match confident detections first, consult appearance only on ties, and let weak detections fill gaps, to keep identities unbroken
- EmbodiedConsequentAgent#e1afAn agent with a body and irreplaceable state, so its actions matter to it
- FourECognition#5a7eCognition is spread over brain, body, environment and action, and the boundary of a cognitive system is set by its couplings, not by the skull
- InteractionSequenceEntropySignature#8e4bCollective regimes differ in how predictable the next interaction is, and only an order-keeping network shows it
- LeakyModulatorBuckets#2a55A few leaky reservoirs fill from events and shift how the system responds
- LimbicAdapterLayer#1ce2A modulating layer between a large model and the user shifts its behavior with affective state
- LocalLearningRule#0ab3Each unit updates from locally available signals, not a global objective
- LocalSpatialStateDivergence#2b5eClassify a crowd's collective state from positions alone, by how far its local neighbourhoods depart from random
- ModelAsCompression#265dA model represents a process in fewer computational steps than the process itself
- ModelStyleConvergence#4282Shared model copies share one voice, which reads as no personality
- PredictionErrorMinimization#5a5cA system stays viable by continually predicting inputs and reducing surprise
- PredictiveBehaviorEmbedding#c46bTrain a network to predict future motion; its internal features separate behaviours it was never told about
- RejectionCostOptionality#6022The cheaper a system rejects options, the more alternatives it can afford to entertain
- RotationNormalizedPoseDetection#1a44Detect in the full frame, then crop and rotate each instance to a canonical heading before estimating its pose
Memory 5
- GenerationalKnowledgeTransfer#5c5cPass useful knowledge to a successor and let the predecessor end
- NarrativeStateInjection#7889Give a model its internal state as a short story and rewrite it every turn
- PersistentInnateState#b0eaA durable initial state survives each short-lived instantiation
- PinkElephantContext#5131Anything placed in context stays active, even when told to avoid it
- StrategicForgetting#c334Retention needs a filter that discards what is no longer useful
Reasoning 7
- BoundedModelOpenWorld#a334Complexity is the gap between a bounded model and an open world, and it does not close
- ContextConfirmationBias#d887What sits in the context or in the researcher's mind is what the instrument then finds
- EpistemicComplexity#1be3Complexity is a stance of the knower: three dynamics that keep crucial information out of reach
- FeltDirectionCheck#eab6Before each action, sense whether it moves toward contentment or toward agitation
- NaturalProblem#4e94A problem where making attempts comparable is what prevents solving it
- RepresentationDependentOrder#212cWhether a collective looks ordered depends on the representation, so read several and expect them to disagree
- TimeAggregationBlindness#84d8Pooling interactions over a time window erases the order that distinguishes collective regimes
Strategy 22
- AgentTransitionGraph#e470Model an agent system's actions as a weighted state graph and plan backwards from a goal state
- BoundedHorizonEffort#b8e4A known finite horizon raises the effort an agent spends per step
- CryptographicChange#43e6Spend the capacity a prediction frees on hidden extra work instead of returning it as savings
- DeterministicRandomizer#a71bUse ordinary code to inject randomness so the model does not loop
- EffortlessAction#5172Act from flow and least resistance rather than from a goal or need to prove
- FormChangeRepertoire#7c1dA life cycle is a set of tactics played under stress, not a fixed sequence of stages
- FourPillarBaseline#bc68Sleep, intake, motion, and creation as the four things to keep in order
- FractionalDelivery#a3e4Partly build responses to possible stressors; the cost of carrying them is preparedness
- GoalMisspecification#ad3fA wrong goal defeats correct execution; choosing the goal is the highest-leverage decision
- HarnessAsArchitecture#8256Intelligence lives in the architecture around a static model, not in the model alone
- HarnessForcedCreativity#ba4aModels are only creative when the harness forces variation from outside
- InformationDividend#c78dActing on short-range information pays even when the long-range forecast is exactly right
- IrreversibleStake#7e7fIrreversible outcomes give an agent something real to lose
- MassAmplifiedError#9457A fixed error rate in frozen information grows in absolute size with the mass invested in it
- MetaModularity#b97eAdapt by holding overlapping, non-identical models of the same situation and moving between them cheaply
- ParasiticSearch#484bExplore the edges of your main model on time and resources paid for by something else
- PersonalGrowthCompanion#16e0A long-lived agent that helps a person choose goals and keep their baseline, without rigid command
- ResidualityStressorIntegration#d5c9Design for many arbitrary stressors directly in the structure, skipping a forecast of the future
- RiskExoskeleton#7b41The risk-management uses hidden inside assets whose obvious purpose is something else
- RouterIntelligenceParadox#210aChoosing which model to send a job to takes at least the intelligence of the job
- ShapeMassInformation#0b9eGrowth risk is the balance of controlling shape, adding mass and adding information
- TransitionHub#0798Occupy a state from which many valuable states are cheap to reach
Society30
Coordination 20
- ChimericalComputation#cedfImperfectly aligned, overlapping work packs more functions into the same parts
- CollectiveInformationArchitecture#5891Read a collective's behavioural regime as a configuration of how information can flow through its interactions
- CooperationHorizon#8806Cooperation holds when interaction is repeated without a known end
- CoordinatorLaneTopology#7e86One coordinator, a few lanes, sub-agents beneath, and no rigid org chart
- CrackpotProbe#c9dfAsk for someone's most crackpot belief and offer one honest objection, to see how they think where consensus no longer filters them
- DistributedIntelligence#3aa5Many model instances selectively exchange information to reduce uncertainty about their environments
- ExperienceMixAllocation#0fc0Mix a few deep experts with cheap or diverse workers, matched to how new the problem is
- ExpertInformedOrchestration#86d6Automate agent orchestration with a probabilistic transition model, guarded by strong routing
- FreelancerBrief#55eaBrief an agent like a capable outsider who has never seen your company
- GameTransitionGraph#981dA graph of pure coordination games whose edges are weighted by how many agents must switch strategy
- InformationGame#d12bIntelligences modelling each other modelling them, where the missing information is a property of the system
- LateralityCost#b281An action gets harder with every additional unit that must coordinate to perform it
- MultiPathReliability#6e7bMake an outcome reliable through many noisy ways to reach it, not one accountable path
- OptionVectorPlanning#c64fPlan resources as a forecast, plan tasks as options that stay open at every level
- OrganisationGame#e446Perfectly aligned agents still fail when the best joint strategy changes and too many must switch
- OutsiderDiversityInjection#80feSeed a search with people experienced in unrelated fields plus a few domain experts
- PeripheralHitchhiking#4aaeUnanticipated but needed information travels on channels that already exist for other reasons
- Stigmergy#ab14Agents coordinate indirectly by leaving traces in a shared environment that cue later actions
- Superorganism#e55bA colony whose integrated parts function, and are selected, as one higher-level organism
- SwarmCostQualityTradeoff#ae73Prefer strong agents in a simple topology; use cheap agents only where errors are cheap
Economics 4
- HypeCorrectionCycle#2980Unverifiable claims run on fear of missing out until markets demand proof
- IncumbentSlackTest#62d2Predict what an incumbent will not pursue by asking where it is not spoiled for resources
- LabGravityWell#e379Ventures near a frontier lab risk being absorbed, yet most of the space lies outside its pull
- ObsolescenceHedging#7af3Delay building what the platform vendor is about to ship
Protocols 4
- ContentAddressedMinting#dc7bA generative agent drafts a new word; deterministic code accepts or rejects it and derives its name from the hash of its definition
- FuzzyAlgorithm#f1f3A reusable route through coarse regions of a problem space, where each step names a condition to reach and leaves how to reach it to the executor's judgment
- SelfOrganisingVocabulary#0af4A shared vocabulary shaped only by the interactions of the agents that use it
- ThinkingTrace#dabdRecord a line of thought as an ordered walk through shared, content-addressed concepts, so the route from question to answer can be inspected, compared and reused
Use with your AI assistant
Structure thinking in your chat session
A skill that reads the patterns live from this site. Your AI assistant applies the patterns that fit to your problem, checks their conditions and failure modes, and points you to the research and people behind them. When nothing fits, it drafts a new pattern you can submit. New patterns show up without updating the skill.
Install from a chat (also on mobile)
Paste this into a Claude chat. It needs Code execution and file creation turned on (Settings → Capabilities).
Install the BraInference knowledge-graph skill from https://brainference.eu/knowledge-graph/skill.jsonOr download the zip
download skill (.zip)- Claude: Customize → Skills → Add → Upload skill, and choose the zip.
- ChatGPT: unzip, paste
SKILL.mdinto the instructions of a project or custom GPT, and turn on web search. - Local agents: unzip into
~/.claude/skills/(Claude Code) or~/.codex/skills/(Codex).