Place external systems

Type: kb/types/instruction.md

Give the artifact a map of named systems against its criteria, with every placement supported by a code-grounded main-review output or an explicitly identified source ingest the KB already holds.

Effort: simple. Lookup and summary from generated main-review files; the operator checks the contestable list.

  1. Select generated reviews directly under kb/agentic-systems/reviews/ with generated-by: analyse-agentic-system, analysis-run, source-identity, and reviewed-revision. Read the evidence basis and admit only findings supported by inspected code at the named revision. Record each input path, SHA-256, run ID, and source revision in the report. A kb/sources/ ingest may independently supply code-grounded findings; identify it as an ingest, not a main review. Do not select a legacy memory review or use it to fill gaps in a main review.
  2. Read the selected file directly. For each proposed placement, identify the supporting section and sentence before drafting. If the compact review is insufficient and its exact result.md is available, run commonplace-agentic-analysis-handoff <sibling-run-state-path> from the repository root. Require exit status zero, substantive result disposition complete, and matching run, source revision, and generated-review path. Read the exact result's relevant records and limitations and record its SHA-256 and canonical record IDs. If the required support is still missing or validation fails, omit that placement and report the gap. Do not acquire new sources or infer a mechanism from the system's claims or from omission.
  3. For each supported placement, state in one or two sentences what the system retains, what admits a change, and which criteria the evidence supports or rules out within its inspected boundary. Mark uninspected criteria as uninspected. Say what a gate judges and keep wiring, observed operation, and causal effect separate.
  4. Close with what the selected placements show together and which criteria their evidence does not test. Do not generalize that limit into a claim about every system.
  5. Link each placement to the tracked main-review file or ingest that supports it. A tracked artifact must not depend on an ignored exact result being present. If the needed support exists only in local state, return the proposed placement and its retention or regeneration requirement in the report; do not insert it into the artifact or point to an insufficient compact review as though that file carried the evidence.
  6. Before returning or applying the paragraph, recheck the input hashes and any completion check used in step 2. If an input changed or validation fails, withhold the affected placement and report the changed input.

Report: the paragraph; input identities and hashes; withheld placements and their missing support; and placements a system's author would contest, each with the supporting section and sentence for the operator to check.

Preserve: the artifact's criteria as stated. A system that does not fit them is evidence about the criteria for the operator, not a reason to adjust them in this pass.