SOCK task guides

Add a source

1. 01

   Confirm authorisation

   Verify that the source, its data classification, and AI processing path are approved for this capture.

2. 02

   Choose the correct access mode

   Use Clone for a repository, Point for a local file or folder, Connect for a supported live database, or Ask for attributed human knowledge.

3. 03

   Select the narrowest useful scope

   Choose an exact repository revision, approved folder, database/schema scope, or bounded knowledge note.

4. 04

   Record intentional omissions

   Name unavailable or excluded sources so the capture boundary remains honest.

5. 05

   Review before freezing

   Confirm every source is correct. You cannot change the set after freeze.

Connect a supported database

The supported connector families named in the current product are PostgreSQL, MongoDB, and Oracle. Availability can still vary by release or deployment.

1. 01

   Check the version floor

   PostgreSQL must be 15+, MongoDB 7+, and Oracle 19+. A below-floor server is refused before capture files are written.

2. 02

   Prepare a least-privilege account

   Grant read-only access to only the database objects approved for the capture.

3. 03

   Connect and verify

   Provide the deployment-specific connection details and test access. After authorisation, Genesis keeps the revealed secret in the operating system’s secure credential store: macOS Keychain, Windows Credential Manager, or a Secret Service-compatible Linux keyring. Confirm TLS, VPN, proxy, and allow-list requirements.

4. 04

   Browse and select scope

   Choose the database, schemas, tables, collections, or objects that belong in the declared evidence boundary.

5. 05

   Decide on sampling

   Sample rows may contain sensitive data. Disable sampling or constrain it to the authorised size when required.

6. 06

   Add the database to Sources

   Confirm the selected database objects and permitted samples appear in the Sources phase. Review this boundary alongside repositories, documents, and tacit notes before freezing.

Add tacit knowledge

1. 01

   Choose Tacit knowledge

   From Add source, open the tacit-knowledge path when important context is not present in repositories, documents, or connected systems.

2. 02

   Give the note a useful title

   Name the decision, operating procedure, glossary, ownership boundary, or historical context being recorded.

3. 03

   Write in Markdown

   State who supplied the information, when it was current, which systems or environments it applies to, and what evidence could confirm it.

4. 04

   Preview before adding

   Use Write, Split, or Preview to check the note. Remove secrets, personal data, and unsupported certainty.

5. 05

   Freeze with the source set

   The note becomes attributed, human-reported evidence inside the same permanent capture boundary.

Annotate a source before freezing

1. 01

   Add and prepare the source

   Wait until the source row shows that preparation or connection has completed.

2. 02

   Choose Note

   When the source row exposes the Note action, add concise source-specific context such as purpose, owner, environment, branch intent, caveat, or expected blind spot.

3. 03

   Keep fact and interpretation separate

   Describe why the source matters without changing what the underlying source says. Use a tacit-knowledge source for longer standalone knowledge.

4. 04

   Review before freezing

   Confirm that the annotation contains no secret or unauthorised data and still applies to the selected revision or export.

Freeze sources and start Survey

• Resolve repositories to an exact commit and create the frozen archive.
• Copy files and folders under guarded rules and record content hashes.
• Query supported databases read-only and write a bounded capture-time export.
• Persist a failure reason when a source cannot be frozen; do not accept a silently empty capture.
• Continue to Survey only after the evidence boundary is durable.

Complete Survey

1. 01

   Confirm every source is frozen

   Survey starts from the permanent evidence boundary. Resolve any failed freeze before continuing; do not proceed with a silently missing source.

2. 02

   Start Survey

   Choose Start survey from the Capture phase. SOCK probes the frozen sources, records their structure and access boundary, and identifies what it can and cannot inspect.

3. 03

   Follow durable progress

   Use the task ledger, processed-source count, walk steps, blockers, and Run activity to judge progress. The Claude Survey terminal explains the live session but is not the completion record.

4. 04

   Review the source descriptions

   Check that each frozen source is described accurately and that unavailable evidence, access limits, or other gaps remain visible.

5. 05

   Continue to Plan

   When Survey shows durable completion, use Finish & continue to Plan. If the assistant is still writing final notes, allow it to settle or confirm the durable artefacts before continuing.

Run and review Plan

1. 01

   Confirm Survey is complete

   Review the frozen-source descriptions, blockers, and known gaps before asking Plan to define the extraction sequence.

2. 02

   Start Plan

   Plan ratifies the coverage frame and builds the ordered work that Extract will follow for the captured source units.

3. 03

   Review scope and dependencies

   Check that the plan covers the intended repositories, documents, connected systems, and cross-source work without implying access beyond the frozen boundary.

4. 04

   Resolve a blocker or stale step

   Use the named blocker and durable run activity to decide whether to resume. Do not start a duplicate planning run because the terminal appears idle.

5. 05

   Continue to Extract

   When Plan is approved and its execution artefact is persisted, choose Finish & continue to Extract.

Run Extract

1. 01

   Confirm the approved plan

   Extract should begin only after Plan has persisted the ordered work for every selected source unit.

2. 02

   Start extraction

   Run the extraction queue. SOCK processes source units sequentially and writes facts, citations, evidence tiers, known unknowns, and generated knowledge files to disk.

3. 03

   Watch jobs, blockers, and knowledge health

   Use durable job counts and the knowledge-health panel to see facts, inferred output, and recorded gaps as completed units land.

4. 04

   Review a blocked unit before resuming

   Read the persisted failure reason and preserve completed work. Resume scheduling from the last durable marker rather than restarting the capture.

5. 05

   Confirm every required unit is complete

   Discover becomes meaningful only after the required extraction units have durable outputs. A quiet Claude terminal alone is not proof of completion.

Preview extracted knowledge and evidence

1. 01

   Wait for a source unit to complete

   Use the Extract queue and durable job count, not terminal silence, to confirm that the unit is ready.

2. 02

   Choose Preview

   Open the completed unit from the extraction queue. The viewer shows the generated artefact tree without changing the capture.

3. 03

   Read the README

   Start with the unit summary and captured scope before reviewing individual dimension documents.

4. 04

   Inspect Evidence

   Check the Claim, Citation, Tier, and Status columns. Follow important citations back to the frozen source.

5. 05

   Inspect Known Unknowns

   Treat missing evidence, unresolved references, and unsupported conclusions as part of the result rather than defects to hide.

6. 06

   Record review context

   Use the current release's annotation or Note action when available. If the source set is already frozen and material context is missing, preserve it in a new capture rather than altering the old boundary.

Run Discover and Infer

Discover reads across completed extraction units and catalogues higher-order surfaces. Infer walks those surfaces to derive deeper candidate knowledge. The current Beta UI may combine them in a single Discover & Infer tab, but they remain two distinct processing responsibilities.

1. 01

   Confirm Extract is complete

   All required units should show durable completion. Review blocked units and known gaps before higher-order processing.

2. 02

   Open Discover & Infer

   Start or resume the combined view when your release groups the two stages. Follow run activity for the transition from surface discovery to inference.

3. 03

   Filter by lens

   Use Structure, Integration, Reliability, Security, Change & risk, or Other to narrow the review. All restores the complete dimension list.

4. 04

   Review dimension coverage

   Expand dimensions and inspect surface counts. Zero surfaces can be honest; it does not prove that the concern is absent from the real estate.

5. 05

   Preview a surface

   Inspect the source evidence and cross-source context before accepting the surface as useful.

6. 06

   Check inference and knowledge health

   Confirm derived output remains marked inferred, evidence-linked, and separate from verified facts. Review recorded gaps before continuing to Ask.

Use the [21-dimension glossary](/sock/reference#dimensions) to understand the exact concern and review question behind every internal identifier.

Ask a grounded question

1. 01

   Open the capture

   Choose the capture whose frozen source boundary matches your question.

2. 02

   Open Ask

   Ask is available once baseline claims exist and is always read-only.

3. 03

   Write a bounded question

   Prefer “Which configuration controls retry behaviour?” over “Explain everything about reliability.”

4. 04

   Review scope and evidence

   Inspect citations and note whether the answer is an observation, interpretation, or known unknown.

5. 05

   Verify consequential claims

   Open the cited source before acting on security, architecture, data, or release decisions.

Use the Claude terminal and run activity

The Claude terminal and Run activity support Survey, Plan, Extract, Discover, and Infer. They explain execution across the lifecycle; they are not a separate lifecycle stage.

• The Claude terminal is a live view of the agent session executing the active stage.
• Use it to inspect commands, active work, diagnostic output, and a pause that needs attention.
• Use Run activity for durable stage events and transitions; expand the terminal only when detailed output is useful.
• Do not infer completion from a quiet prompt or narrative line. Trust persisted artefacts, job counts, lifecycle status, and completion markers.
• If scheduling stops, use the stage's Resume action. Do not start a duplicate run merely because the terminal looks idle.

Resume an interrupted run

1. 01

   Return to the capture

   Review the current durable processing state.

2. 02

   Use Resume

   Continue from the last completed marker rather than starting a duplicate run.

3. 03

   Allow Finishing… to settle

   This state can appear while completion evidence is being committed. A quiet terminal is not itself proof of failure.

4. 04

   Read the blocker

   If processing cannot continue, use the persisted failure reason to correct access, credentials, data, or environment conditions.

5. 05

   Resume Discover or Infer in place

   Return to the combined Discover & Infer view and continue from its last durable marker. Do not discard completed surfaces or restart the entire capture.

Run a meaningful validation capture

1. 01

   Agree the evidence boundary

   Name repositories and revisions, folders, documents, database scope, team notes, and intentionally omitted sources.

2. 02

   Agree the questions first

   Define what you need to learn and what an honest unsupported answer should look like.

3. 03

   Evaluate traceability

   Check frozen provenance, citation reachability, visible unknowns, resumability, and the distinction between candidates and validated relationships.

4. 04

   Measure outcomes, not volume

   Do not use claim count as the success metric. More claims can reflect source size or noise rather than better comprehension.