Part 4 — Building and Operating

Your LCM Transition Roadmap

This chapter does not summarize the book. The chapters before it covered the architecture, the use cases, the comparison, the tooling, the evaluation, and the governance. The job of this chapter is to tell you what to do next.

The LCM adoption decision is not a research question. It is an execution question. The answer to "should we adopt LCMs?" is almost certainly "yes, for specific use cases, on a specific timeline." The question that matters is: what does that timeline look like, and what do you do at each stage?

Three phases. Each has a concrete objective, a set of activities, and a completion criterion. They are sequential — do not start Phase 2 until Phase 1 is complete. Phase 1 can be done in four weeks.

16.1 Phase 1: Evaluate (Weeks 1–8)

Objective: Identify one use case that credibly hits the token ceiling, validate that claim with an internal benchmark, and produce a go/no-go recommendation for Pilot.

The most common mistake in LCM adoption is skipping Phase 1. Teams read about LCM capabilities, identify a use case that sounds like a match, and begin building. Eight weeks later, they discover that the use case was better served by an LLM, or that SONAR alignment quality for their domain is insufficient. Phase 1 prevents these discoveries from happening in production.

Weeks 1–2: Use case identification and prioritization.

Apply the decision matrix from Chapter 7 to the AI use cases in your organization's backlog. Identify all LCM-candidate use cases: tasks that answer "concept," "yes," and "yes" to the Task Unit Test. If you have no LCM candidates, this book is not relevant to your immediate roadmap — revisit when your use case backlog changes.

For each LCM candidate, assess:

• Business value: what is the cost of the current manual or LLM-based process? What would a successful LCM deployment save or enable?
• Feasibility: does the use case have a defined document corpus? Is the corpus available and in digital form? What languages are involved, and are they covered by SONAR?
• Risk: is the use case in a regulated domain? What are the governance requirements?

Select the one use case with the highest business value and the lowest risk. Resist the temptation to evaluate multiple use cases simultaneously — Phase 1 requires focus.

Weeks 3–5: Internal benchmark construction.

Build a benchmark dataset for the selected use case:

• 20–50 representative input instances (document pairs for comparison tasks, document sets for synthesis tasks, planning prompts for planning tasks)
• A gold standard for each instance: what a correct output looks like, annotated by domain experts
• Clear quality criteria: what metrics constitute a "good" output (Chapter 14)

Run both an LLM baseline and an LCM prototype on the benchmark. The LLM baseline uses your current best approach (best model, best prompt, best RAG setup). The LCM prototype uses SONAR encoding and the concept model's output on the same inputs.

Weeks 6–8: Benchmark analysis and go/no-go recommendation.

Compare LLM baseline and LCM prototype performance on the benchmark using the semantic similarity and task-specific metrics from Chapter 14. The go/no-go recommendation is determined by:

• If LCM outperforms LLM on the primary quality metric by more than 10 percentage points: proceed to Pilot
• If LCM outperforms by less than 10 points: assess whether the improvement is worth the tooling overhead
• If LCM does not outperform: the use case is not an LCM candidate at current SONAR alignment quality; select a different use case or wait for SONAR domain adaptation tooling

The go/no-go recommendation should include: the benchmark results, the estimated Pilot cost (engineering time + infrastructure), the expected business value, and the recommended Pilot scope.

16.2 Phase 2: Pilot (Months 3–6)

Objective: Deploy the LCM as an internal-only capability with human review gates. Build the evaluation harness. Instrument for observability. Produce validated evidence of business value.

Phase 2 is not a production deployment. It is a controlled experiment with real users on real tasks, structured to produce validated evidence before the Scale commitment.

Month 3: Infrastructure and tooling setup.

Deploy the core LCM infrastructure: SONAR encoding server, vector database, concept model inference, SONAR decoder. Use the build-vs-buy guidance from Chapter 13 for each component. Build the minimum viable evaluation harness (Chapter 14). Set up basic observability: log input texts, SONAR embeddings, similarity scores, and decoded outputs for every inference call.

Month 4: Governance and compliance setup.

Complete the LCM governance checklist (Chapter 15). Implement attribution records for all LCM outputs. Get model documentation to the state required by your organization's model risk management process. If the use case is in a regulated domain, initiate the validation documentation process now — it takes longer than the technical work.

Months 5–6: User trials with human review gates.

Deploy to a small group of internal users (5–15 people) from the target business function. Every LCM output goes through a human review step before it influences any decision. Track: user satisfaction with LCM outputs versus previous approach, reviewer time per output, error rate as assessed by reviewers, and the specific error types (semantic inaccuracy, global incoherence, hallucination, bias).

If reviewer time is increasing, the system is adding work. If it is decreasing, the system is saving work. That metric is more honest than any benchmark score.

At the end of Month 6, produce a Pilot report: quantified business value evidence, error rate, user satisfaction, infrastructure cost, and a recommendation for Scale.

16.3 Phase 3: Scale (Months 7–18)

Objective: Extend to production with the hybrid architecture from Chapter 12. Build the governance framework. Track ecosystem maturity signals. Expand to additional use cases.

Phase 3 begins only when the Pilot report demonstrates sufficient business value and acceptable error rates. Define those thresholds in Phase 1 as the criteria that would justify proceeding — do not define them after seeing the Pilot results.

Months 7–9: Production architecture.

Replace the Pilot infrastructure with a production-grade deployment: redundancy, auto-scaling, SLA monitoring, and incident response procedures. Implement the full hybrid architecture for your use case. Add the LLM wrapper components appropriate to your use case profile (concept router, concept elevator, or concept pipeline from Chapter 12).

Months 10–12: Governance maturation.

Complete the full governance checklist. Register the model in the organization's model inventory. Establish ongoing monitoring: track semantic similarity and task-specific quality metrics over time to detect performance drift. Set up a retraining or re-encoding trigger for when SONAR alignment quality degrades for any language or domain.

Months 13–18: Use case expansion.

Return to the prioritized use case backlog from Phase 1. Apply Phase 1 evaluation to the next highest-priority LCM candidate. Run Phase 2 Pilots in parallel with the production Phase 3 deployment. The second use case onboarding should be faster: the infrastructure is already built, the governance process is established, the evaluation harness is already instrumented.

16.4 Your Action Plan: Week, Quarter, Year

This week:

• Apply the Task Unit Test to three AI use cases in your current backlog
• Identify whether any are LCM candidates
• If yes: begin Phase 1 Week 1–2 activities (use case prioritization)
• If no: bookmark the maturity signals from Chapter 8 and set a quarterly reminder to re-evaluate

This quarter:

• Complete Phase 1 (Evaluate) for your top LCM candidate
• Produce the go/no-go benchmark report
• If go: begin Phase 2 (Pilot) infrastructure and tooling
• Brief your technology leadership on the benchmark results and Phase 2 plan

This year:

• Complete Phase 2 (Pilot) and produce the business value report
• Make the Scale decision based on Pilot evidence
• If Scale: begin Phase 3 and add the second use case to the Pilot pipeline
• Track ecosystem maturity signals from Chapter 8 quarterly; update your adoption posture as the tooling landscape evolves
• Contribute your benchmark methodology and evaluation harness to your organization's AI governance documentation — these will be reusable assets for the second and third use cases

16.5 What Success Looks Like

At the end of eighteen months of disciplined execution:

• One or two LCM-powered systems in production, each demonstrating measurable improvement over the LLM baseline on a specific concept-level task
• A validated evaluation harness that your team can apply to new LCM use cases without rebuilding
• A governance framework that satisfies your organization's model risk management requirements
• A monitoring system that tracks semantic similarity and task-specific quality metrics over time and flags drift before it affects users
• A use case backlog with three to five additional LCM candidates evaluated and prioritized, ready for Phase 1 when capacity opens

What success does not look like: an enterprise-wide LCM replacement of your LLM infrastructure, a single concept model handling all AI use cases, or a completed transformation in less than twelve months. LCM adoption is additive, incremental, and evidence-gated. Teams that approach it otherwise will spend more, build more, and get less.

Exercises