# AI SDLC and agentic engineering: an inspectable delivery method

Presentation guide · 7 October 2026 · LockedIn Labs

Use this eight-minute walkthrough to examine the publication's method and browser-local reference tools. The synthetic simulation and prepared implementation references are distinct from the historical executed Java example. No step calls a production agent, authenticates evidence, grants authority, or activates enterprise controls.

## 1. Establish the argument — one minute

Open [the publication and guided demonstration](https://aisdlc.ai/#demonstration). Describe the operating problem: delegate iteration while keeping acceptance criteria and authority separately controlled. The AI SDLC includes specification, context, architecture, evaluation, governance, release preparation, operation and evaluated improvement.

Ask the audience which decision they would delegate and what consequence they would retain. The unit of delegation is a defined workflow, rather than an enterprise-wide promise of autonomy.

## 2. Change the boundary — two minutes

In the guided demonstration, select **Within scope**. This fixture declares a current identity, permitted policy, passing verifier results, complete evidence and a named disposition. The deterministic reference engine returns **Simulated allow**. Those declarations are not measurements of a running system.

Select **Identity revoked**. The same proposed action now returns **Block**. Select **Evidence missing** to produce **Hold**, then **Policy prohibits** to produce **Block**. Each case changes one posture field and retains the same task and verifier declarations.

Expand the ten trace stages. Inspect the decisive rule and the missing obligations; export the unsigned JSON trace. The artifact carries the scenario, engine revision, rule evaluations and explicit non-authorizing status. Ask how a real implementation would intercept the effect and propagate revocation. The browser simulation demonstrates the decision model, not that interception.

## 3. Carry a plan into implementation — three minutes

Open [Project Dossier](https://aisdlc.ai/project-dossier). Its built-in example is synthetic. Inspect the declared plan, passport and source identities. Use **Review in Workbench** to inspect the exact plan; return it unchanged before continuing. Editing risk inputs requires renewed classification and can invalidate derived references.

Follow the dossier stages into assurance, policy references and the Lab. Show missing evidence and named decision duties, rather than presenting prepared documents as accepted proof. Use the dossier's Implementation Studio path to carry the source package into [Implementation Studio](https://aisdlc.ai/implementation-studio).

Inspect a deployment template, verifier responsibility and evidence connector contract. Download the generated reference assets. They are starter documents for an operating team to implement and validate. They are not executable enforcement, authenticated evidence or release approval. Inputs remain browser-local; use synthetic data throughout a public demonstration.

## 4. Show what was actually executed — one minute

Open [the historical Java verification trace](https://aisdlc.ai/spec/aisdlc-cq#java-gate-evidence). The defective candidate ran nine tests with one cross-tenant failure. The repaired candidate ran nine tests with no failures, errors or skips, under unchanged acceptance checks.

State the limit precisely: architecture checks passed in the failed run, and branch protection was enabled afterward. The example does not establish an enforced merge rejection at the time, complete specification conformance, customer results, or today's repository protection settings. The supporting source history is private; the public page preserves its evidence summary.

## 5. Connect the method to current research — one minute

Open [the current practitioner edition](https://aisdlc.ai/news#edition-2026-10-07). Source reporting preserves original publication dates and measurement limits. The edition's synthetic exercise is proposed and unexecuted. Its decision note asks teams to measure allowed work and prohibited effects separately, preserve failed attempts and denominators, and keep proposed policy changes outside the acting agent's authority.

Close with an implementation decision: select one reversible workflow, fix its acceptance conditions and permissions, rehearse failure and recovery, then use retained evidence to decide whether a wider boundary is justified. Some workflows can move to human-on-the-loop operation; a hybrid approach can remain appropriate where ambiguity or consequences call for closer review.

## Questions to expect

**Is this a live autonomous runtime?** This site is the knowledge publication and reference workspace. Its planning tools and deterministic simulation are browser-local. The publication does not demonstrate a running enterprise enforcement service.

**What makes a model judgment trustworthy enough to act on?** Evaluate the particular judgment against representative cases, preserve uncertainty and disagreements, and enforce action permissions through a protected path. An evaluator can also fail. An explanation or benchmark score alone is insufficient authority.

**Can the system improve itself?** Feedback can create candidate changes to prompts, tests, tools or policies. Promotion requires a separately controlled evaluation and authorization process. A production implementation must show that the executing agent cannot silently expand its permissions or redefine acceptance.

**What would we need to prove in a pilot?** Accepted outcomes and verification cost; unauthorized effects and missed failures; permission revocation; recovery and rollback; evidence provenance; and operational ownership under realistic dependencies. The publication supplies reference methods and artifacts, not a guarantee that those properties hold in a particular deployment.
