Harness, sources, and locking

A harness is a connected set of knowledge created from your selected sources. It is the object you create, name, run, review, and release.

TermMeaningOperational consequence
HarnessA connected set of knowledge created from selected sourcesThe unit you create, run, release, and publish
SourcesThe preparation destination for repositories, files, knowledge artifacts, connected systems, and attributed human knowledgeReview, preview, annotate, and lock the complete boundary before processing starts
Lock the source listA permanent lock of the prepared source setYou cannot add, remove, or replace sources afterwards
Completed and verifiedThe state a harness reaches when every stage has finished and the chain has been verifiedAsk opens; the Release destination opens

Source types and current connectors

The Add source dialog has four tabs.

TabWhat it acceptsCurrent behaviour
Local RepositoriesRepositories and folders on this machineRead and prepared in place; copied into the harness boundary when you lock the list
Connect SystemPostgreSQL, MongoDB, or OracleConnected read-only; browse the available structure, select an approved scope, and optionally include bounded samples
Knowledge ArtifactsDocuments and other approved local filesCopied and fingerprinted; extraction depth varies by format
Add Tacit KnowledgeAttributed Markdown notesPreserves decisions, context, ownership, and operational knowledge that system artefacts do not contain

The processing stages

The Harness rail destination holds three tabs.

TabPurposeDurable result
Scan sourcesMeasure what was supplied, then analyse a safe work sequenceSource registry, inventory, verification records, structure and technology signals, coverage boundary, and the extraction order
ExtractCreate deterministic facts and validated interpretationsDeclared structures, per-source knowledge documents, official facts, citations, and Known Unknowns
SynthesizeRead across every extracted source at once, then derive deeper candidate knowledgeDiscovered items that no single source shows, plus evidence-linked node and relationship candidates marked as inferred

Two further destinations continue the journey:

DestinationPurposeDurable result
GraphIdentify and classify entities, write capability summaries, map and verify relationshipsA committed knowledge graph of the connected estate
Review · AskRead the counts from sources through to billable CKU; ask bounded questionsCited answers, explicit scope limits, and Known Unknowns

Discovered items, dimensions, and lenses

These terms describe three levels of organisation.

ConceptExampleWhat it is not
Dimensioncall-chain or security-postureNot a promise that every harness contains useful content in every category
Discovered itemA specific payment retry call chainNot another name for a dimension
LensSecurity, or Change & riskNot an engine concept; it does not change stored knowledge

The lenses are Structure, Integration, Reliability, Security, Change & risk, and Other. All removes the filter. Lenses change only what you see.

The 21 authored dimensions

SOCK ships 21 authored dimension readers. Together they provide a consistent way to inspect business behaviour, software structure, data, operations, risk, and change seams.

The 21 SOCK dimensions arranged as two above water, four at the waterline, ten below water, and five at bedrock
Concept model

A conceptual depth view of the dimensions. The layers explain visibility and discovery depth; they are not engine states or UI lenses.

#IdentifierKnowledge focusLens
1call-chainExecution paths across functions, jobs, services, modules, and external callsStructure
2business-journeyEnd-to-end customer, user, and operational journeysStructure
3business-rulesDecision logic, calculations, validation, routing, and policy constraintsStructure
4ui-flowScreens, states, navigation, actions, validation, and supporting behaviourStructure
5data-relationRelationships among entities, records, keys, schemas, and joinsStructure
6data-lineageData origin, movement, transformation, persistence, and downstream useStructure
7domain-modelDomain concepts, vocabulary, responsibilities, and bounded contextsStructure
8capability-mapBusiness and technical capabilities and the components that realise themStructure
9contract-surfaceAPIs, events, files, schemas, commands, and declared contractsIntegration
10interface-mappingsInterface boundaries to the outside world and across internal seams, including partner messaging and batch feedsIntegration
11integration-topologyConnected systems, routes, protocols, directionality, and hubsIntegration
12reliability-mapFailure, retries, timeouts, recovery, observability, and resilienceReliability
13ops-inventoryJobs, schedules, runbooks, controls, utilities, and support proceduresReliability
14infra-inventoryRuntime, compute, network, storage, middleware, and deployment componentsReliability
15security-postureIdentity, access, secrets, trust boundaries, exposure, and controlsSecurity
16compliance-footprintObligations, controls, audit evidence, and regulated-data handlingSecurity
17dependency-blast-radiusUpstream and downstream dependencies that can transmit changeChange & risk
18migration-seamsBoundaries for extraction, replacement, coexistence, and phased migrationChange & risk
19dead-codeImplementation that appears unused, unreachable, superseded, or orphanedChange & risk
20as-is-archaeologyCurrent implementation reality, historical residue, and undocumented structureChange & risk
21legacy-estate-mapMainframe and legacy components behind a modern product, and what resists modernisationChange & risk

Facts, evidence tiers, and honesty

Every accepted fact is retained with a citation back to its source. The system keeps how a fact was obtained visible: a parsed declaration, a live observation, an AI interpretation, and a human report are not equally strong.

  • AI components propose interpretations; validated application code writes the official record.
  • The model cannot promote its own output to a stronger evidence tier.
  • A human report stays attributed and cannot outrank a parsed declaration.
  • Missing, unexamined, unsupported, and unknown are distinct states.
  • Inferred means reasoned from available evidence rather than stated directly in a source.
  • Known Unknowns are questions or gaps the harness knows it cannot currently answer.
  • Evidence-backed does not mean infallible; consequential decisions still require source review.

Known Unknowns are a product outcome

A useful knowledge system must show where its confidence ends. SOCK therefore preserves a Known Unknown when the captured evidence points to a real concern but cannot support a safe conclusion. Examples include a referenced component that is outside the source boundary, an integration endpoint whose owner is not present, conflicting configuration values, or one side of a call chain that has not been captured.

This is different from “nothing found.” It tells a reviewer three things:

  • what the harness could establish;
  • which evidence or relationship is missing or contradictory; and
  • what source, owner, or follow-up could resolve the question.

Known Unknowns turn uncertainty into reviewable work. They reduce the risk of treating an incomplete scan as proof that a dependency, control, or impact does not exist.

Preview, notes, and tacit knowledge

CapabilityUse it forEvidence effect
PreviewInspect locked inputs or generated knowledge files, including README, Evidence, Known Unknowns, and dimension documentsRead-only inspection; it does not promote, edit, or validate a fact
Source noteAdd source-specific context — owner, purpose, caveat, environment, or a deliberate omissionContext accompanies the source; it does not rewrite the underlying file, repository, or database export
Tacit KnowledgeAdd a standalone Markdown note for architecture decisions, operational knowledge, terminology, or history absent from system artefactsLocked as attributed, human-reported evidence and kept distinct from parsed or observed fact

Runs, sessions, and the terminal

A run is one processing execution and its status. A session is an active interaction with Claude. Stages can expose a live terminal beside the stage view, showing the active session, commands, progress, and diagnostic output. The run-activity feed records stage events in a more scannable form.

  • Use the terminal to understand what the active stage is doing and to diagnose a pause or failure.
  • Use the activity feed for stage transitions and completed events.
  • Do not treat a quiet terminal or a final line of text as authoritative completion.
  • The stage status, persisted artefacts, task ledger, and durable completion markers are the source of truth.

Ask is a bounded librarian

Ask is a read-only interface over the facts and knowledge documents of one harness. It does not modify the harness and does not silently search beyond the locked source set. It lives as the second tab of the Review destination.

Durable state and recovery

Progress is judged from durable evidence on disk, not from a terminal appearing to finish. A crash, closed laptop, or network drop does not erase completed work; MSOCK offers the appropriate resume action from the last durable completion marker.

A harness that ran every stage but whose chain never verified is shown as Not verified rather than as a green tick — an honest fifth state, still navigable.