KCC 07

Inspector Pipeline

The Inspector Pipeline is the structural learning loop by which patterns observed in cells mature into shared standards in the capability catalog.

InspectionTrace streamPattern detectionProposalsHuman review gate
Created 2026-06-08 · v0.4.0

The learning loop

The Inspector Pipeline is the structural mechanism by which patterns observed in cells mature into shared standards in the capability catalog. It is the learning loop that makes the organization smarter from what its cells do.

Observe. Detect. Propose. Review. Promote.
Diagram
Inspector Pipeline
Observe
Detect
Propose
Review
Promote

Stage 1 — Observe

Input: decision traces (Surface 9) from all cells, in real time. The Inspector reads trace metadata, not agent outputs directly — which agent ran, what tools it invoked, what confidence it declared, the resource consumption, whether it succeeded, and whether humans intervened. This preserves privacy while keeping pattern detection feasible.

Stage 2 — Detect

The Inspector identifies patterns in the trace stream, automatically where possible.

  • Prompt patterns — a prompt structure used repeatedly with high success and confidence
  • Tool composition patterns — sequences that produce consistently good results
  • Confidence calibration patterns — cells whose declared confidence matches actual outcomes
  • Failure mode patterns — multiple cells encountering the same failure
  • Workslop and bypass signatures
  • Cost anomaly and trust-drift patterns

Stage 3 — Propose

The Inspector produces structured proposals: the pattern observed (with evidence), which cells and how many invocations over what period, the proposed action (new capability, new version, kernel modification, advisory, or deprecation), the expected impact, and the risks.

Stage 4 — Review

Maintainers evaluate the proposals. Review is human-led — maintainers are accountable for what enters the catalog or kernel.

OutcomeMeaning
Accept as proposedThe change is implemented as described
Accept with modificationsReviewers suggest changes; the proposal is revised
Reject with reasoningNot implemented; reasoning recorded for future learning
Request more evidenceReturned for more data
DeferScheduled for future review when conditions change

Stage 5 — Promote

The accepted change is implemented: a new L1 capability is added for opt-in cells, a new L2 version is published and announced, a kernel modification is scheduled, or a deprecation is announced with a migration window. Adoption of L2 capabilities is opt-in. See the Maturity Ladder for promotion criteria.

Stage 0 — Import (Optional Intake)

Reference pattern, non-normative. The five stages mine the trace stream conformant agents already emit. Stage 0 is a disciplined landing zone that ingests a foreign agentic workflow (a .cursorrules, a CLAUDE.md, a folder of prompts) and produces a draft capability. One rule is load-bearing: the draft is never auto-promoted — a human reviews and rewrites it, and it enters the normal Review and Promote stages like any other candidate.

Pipeline Operators and Telemetry

The pipeline is operated by 1-3 named Inspector Operators who monitor the trace stream, tune detection rules, and manage the proposal queue (they have access to metadata, not outputs). The pipeline itself emits telemetry: detection_rate, proposal_rate, promotion_rate, time_to_promotion, adoption_rate. A pipeline with many proposals but few promotions has a review bottleneck; few proposals but high promotion has a detection bottleneck.

Why It Matters

Without the Inspector Pipeline, cross-team learning depends on social mechanisms — demos, wiki pages, brown-bags — which are slow, lossy, and ungoverned. The Inspector Pipeline makes cross-team learning structural: automatic, at scale, with named decision rights and an audit trail. For a worked example, see 13.3 — Inspector Pattern Detection.