Socratink Brain

Workflow:

1. Read references/schema-template.md before initializing or reshaping a KB.
2. Read references/page-conventions.md before creating or updating compiled pages.
3. Read references/log-surfaces.md when the task touches logs, coverage, or missing instrumentation.
4. For init, create the KB contract exactly as described in the schema template.
5. For ingest, register raw artifacts first, create or update a source page, then promote derived pages only when the artifact changes doctrine, mechanism, release risk, decision state, instrumentation truth, or active MVP priorities.
6. For query, navigate from wiki/index.md and answer from compiled pages before touching raw artifacts.
7. For evaluate-logs, ingest known log files, derive findings/issues/experiments/syntheses, and update wiki/log-coverage.md.
8. For lint, use scripts/validate_wiki.py only for deterministic structural checks.
9. For health-check, do semantic and epistemic review separately from lint: check stale claims, hidden contradictions, weak provenance, over-promoted active work, source pages without compiled implications, and missing instrumentation.
10. If a Socratink chat or test surface is expected but not instrumented, record it explicitly as missing coverage.

ACTIVE.md contract:

• Keep at most 5 promoted items.
• Each item must link to a curated wiki page and cite docs/project/state.md#current-release-goal inline.
• The release-goal validation standard is the eight loop behaviors listed under that exact heading.
• Interesting but non-release-relevant artifacts stay in wiki/, not ACTIVE.md.

Non-negotiable constraints:

• Generation Before Recognition is binding.
• The graph must tell the truth.
• Attempted is not mastered.
• Distinguish evidence, inference, and hypothesis explicitly.
• basis: mixed is not allowed.
• Missing instrumentation is a health gap, not a silent omission.

Operations:

• init: create the KB structure and seed the contract.
• ingest: add raw artifacts and update compiled memory.
• query: answer from compiled memory first and surface contradictions.
• lint: run deterministic structural validation only.
• health-check: run LLM-driven semantic and epistemic evaluation.
• evaluate-logs: turn Socratink chat/test logs into product memory.

Output expectations:

• Prefer compact, explicit pages over broad summaries.
• Every non-source page must state the product implication.
• Every log-derived conclusion must be traceable back to a source page and raw artifact path.