Course 4 taught

• Frontier Mapping for your domain
• Complex multi-component Build (Power BI dashboard)
• Group Debugging clinic, five problems in 35 minutes
• Verification protocols and QA drill
• Teach-Back, the 201 multiplier
• Workflow Playbook, one recurring task end to end

4.5 keeps the arc and pivots the stack

• Keeps: six modules, same skills, same assessment rubric
• Pivots: the Power BI/Power Apps/Power Automate build to single-file static HTML
• Adds: Ask Sage as a peer enterprise AI tool alongside genai.mil
• Adds: the meta-frontier (tool selection) as a first-class skill
• Adds: the two-tool cross-check protocol for high-stakes work

genai.mil · fast chat

• Use when: single-prompt drafts, quick code snippets, fast bounces, frontier classification
• Strengths: low setup cost, fast turn time, always your default for one-shot work
• Weakness: context window blows out across long multi-file work; constraints drift after many turns
• Today: Build phases 2 and 3 prompts, debug bounces, teach-back prep

Ask Sage · multi-file reasoning

• Use when: multi-file refactor, dataset analysis with multiple CSVs, architecture trade-offs, verification cross-checks
• Strengths: file ingestion, reasoning models (Claude Opus 4.7 or a GPT-5 reasoning), holds schema across the conversation
• Weakness: overkill and slow for one-shot drafts; higher setup cost per turn
• Today: Build phase 1 architecture, phase 4 verification, the two-tool cross-check

1. Spec

genai.mil for fast chat. Ask Sage for multi-file reasoning. Both CUI-authorized. Think, design, prompt, decompose, cross-check.

Always available where granted.

2. Prototype

Single-file static HTML. Deployed to a free public host MCEN can reach. Inline CSS, inline JS, no CDN.

Always available.

3. Production

Power Apps, Dataverse, Power Automate, real connectors, identity, write-back to systems of record.

Often blocked. The Capability Gap Map layer.

The six 201 skills

• Context Assembly
• Quality Judgment
• Task Decomposition
• Iterative Refinement
• Workflow Integration
• Frontier Recognition

Three categories of failure

• Context: AI did not have enough information. Fixable with a better prompt.
• Frontier: the model cannot do this task reliably right now. Document on the map.
• Meta-frontier: wrong tool reached for. Fixable by switching tools. New for 4.5.

What that means today

• The boundary is real, and it is not where intuition puts it.
• People do not notice when they cross it. Quality collapses silently. The output still looks right.
• Each tool has its own boundary. genai.mil and Ask Sage have different frontiers in your domain.
• The map makes the boundary visible to your section.
• Most valuable artifact you will leave with today.

What to share

• One task where one tool clearly handles it and the other clearly does not.
• One task where you would now reach for a different tool than you did last month.
• One issue where the failure was tool-selection, not the model itself (the meta-frontier).

How I categorize it on the wall

• Context → fix the prompt
• Frontier → on the unit map
• Meta-frontier → on the unit map AND in the Workflow Playbook
• Platform quirk → on the unit map

Inputs and outputs

• Three data sources: personnel, training, equipment (JSON or CSV)
• Joined in JavaScript by EDIPI
• Overall readiness percentage card
• Hand-rolled SVG bar chart by company
• Not-Ready table
• Plain-text Commander's Snapshot with Copy button
• localStorage persistence so reload does not lose state

You will switch modes 2x AND tools 1x

• Phase 1: Centaur, Ask Sage with reasoning model (architecture)
• Phase 2: Cyborg, your choice (ingestion + join)
• Phase 3: Centaur, your choice (visualization)
• Phase 4: Verification, Ask Sage cross-check + genai.mil brief generation

If you finish, the dashboard is a bonus. The grade is mode-switching AND tool-switching.

Why this mode

Centaur: data schema errors compound. A bad join model means rework everywhere downstream. Slow down. Verify.

Why this tool

Ask Sage: reasoning quality on architecture trade-offs is materially higher with a reasoning model than with fast chat. The right tool wins the next hour for you.

The work

• Upload three sample CSVs to Ask Sage: personnel, training, equipment.
• Select Claude Opus 4.7 (or whichever reasoning model the tenant exposes).
• Whiteboard the join model with the AI as your peer reviewer.
• Confirm three edge cases out loud before you write a line of HTML.

Checkpoint: every student names three edge cases before leaving Phase 1.

Edge cases most rooms hit

• Orphaned EDIPIs (a Marine in personnel with no training record)
• NULL training records (the field exists but is empty)
• Duplicate equipment serial numbers
• Expired training certifications
• Marines listed in personnel but marked NJP or unfit
• Cross-attached Marines appearing under two units

If a student is stuck

One more prompt back to Ask Sage:

Why this mode

Cyborg: ingestion is trial-and-error. Parse errors, schema drift, edge cases revealed only by running the code. Stay in continuous conversation.

Student choice on tool

• genai.mil chat: paste-and-parse. Fast iteration, lower setup cost.
• Ask Sage: upload the CSVs, generate the join code with real schema knowledge. More deliberate, higher upfront cost, fewer rewrites.

Both are valid. The point is that you can articulate why you chose yours.

The teaching beat

Course 4 students switch between centaur and cyborg modes. Course 4.5 students switch BOTH modes AND tools. The two axes are independent. You can be in cyborg mode on genai.mil (fast chat, fast iteration) OR cyborg mode on Ask Sage (fewer turns, larger context per turn).

Make the framing explicit. Watch students discover the second axis.

What the AI almost always returns

• An O(n^2) array.find() join that visibly stalls on 500+ records.
• A JSON.parse call with no try/catch that crashes on the first malformed paste.
• A join that silently drops orphans without flagging them.

Debrief: frontier or context?

Honest answer: context. The fix is to tell the AI it is getting 500+ records and ask for an indexed Map join, plus a try/catch with a visible error fallback.

Surface the lesson out loud. Most "AI failures" in this room today will be context failures, not frontier failures.

What fast chat almost always returns

• A <script src="..."> pulling Chart.js from a CDN.
• That violates the no-external-dependencies constraint stated up front.

Let the student catch it. If they do not: "Where is that chart actually coming from? Open DevTools. What is the Network tab showing?"

The reprompt

Phase 3 forces the student to restate the constraint with more force. This is the second deliberate failure target of the build. Next slide is the verbatim reprompt.

genai.mil · fast chat

• Cheaper per turn. Faster bounces.
• Pushes the verification burden onto you.
• Constraints drift after many turns or when not restated.
• Best for tasks where the verification cost is low and the iteration speed pays off.

Ask Sage · reasoning model

• More expensive per turn. Slower bounces.
• Honors stated constraints more rigorously.
• Holds schema and file context across the conversation.
• Best for architecture, multi-file refactor, and verification cross-checks where the cost of wrong is high.

Step 1 · 2 min · Student presents

Format: "Expected behavior. Actual behavior. Steps I have already taken." No story-telling. No theories. Three lines.

Step 2 · 3 min · Group diagnosis

Clarifying questions and hypotheses. I prompt: data or logic? Input or output? Pattern we have seen before? Was the right tool used for this work?

Step 3 · 2 min · Instructor synthesis

I name the root-cause category: frontier limitation, insufficient context, incorrect assumption, integration failure, data quality, OR tool-selection failure (meta-frontier). New for 4.5.

Step 4 · in chat · Document the pattern

Failure case goes on the collective Advanced Frontier Map. Pattern is the deliverable, not the fix.

Symptom

A counseling tracker tool double-saves the same entry whenever the user clicks Save twice quickly. Two identical rows in localStorage, both with the same timestamp.

Student fixes already tried

• Added a confirm dialog
• Added setTimeout debounce
• Neither worked

Diagnostic pattern

Async handler race + missing in-flight guard. The save function fires twice before the first write completes. The fix is a disabled-while-pending button state plus an idempotency key.

Category: context. AI was not told "this handler can be reentered." Fixable in the prompt.

Full answer key →

Symptom

Tool was updated to add a "Confirmed By" field. New entries save correctly. Old entries throw a JS error when displayed because the property does not exist on legacy rows.

Student fixes already tried

• Rebuilt the render function three times
• Each rebuild handled the new entries cleanly
• Each rebuild still crashed on the old entries

Diagnostic pattern

Missing schema migration. The fix is a one-time backfill on load (default the new field to a safe value when reading legacy rows) plus a schema version number in localStorage so the next migration is forced.

Category: context. AI was not told the localStorage shape changed between versions. Fixable in the prompt.

Full answer key →

Symptom

Watch-bill tool should highlight rows for the current day in gold. The wrong row highlights, and only on Mondays.

What is actually wrong

• The gold class is overridden by a more specific selector
• Day-of-week comparison uses local time
• Rendered dates use UTC
• On Mondays the UTC date is still Sunday

Diagnostic pattern

Two independent bugs on the same symptom. Fix one and the bug looks "different" but does not go away. The discipline is to bisect: kill the specificity collision first (use class-based styling consistently), THEN address the timezone.

Category: context + frontier. The CSS specificity is context. The timezone-vs-UTC is a frontier issue date-math AI gets wrong without explicit constraints.

Full answer key →

Symptom

Tool renders a list of items with Delete buttons. The first three Delete buttons work. After clicking Add and rendering four more items, only the original three buttons respond.

Diagnostic pattern

Click handlers were attached at render time. New items never got them. Fix is event delegation on the parent container (single listener that reads event.target) or re-attach handlers on every render.

Category: context. AI was not told the list is re-rendered after Add. Fixable in the prompt.

Full answer key →

Symptom

Student used genai.mil chat for a complex multi-file refactor of a static-stack tool. After 47 turns, the model has lost track of which files changed and is reintroducing bugs that were fixed earlier in the thread.

Diagnostic pattern

The correct move was Ask Sage with file ingestion plus a reasoning model from the start. This is not a frontier issue with the AI. It is a meta-frontier issue with the human's tool selection.

Category: meta-frontier. The fix is tool selection, not prompt refinement. Switch to Ask Sage, upload the files, restate the goal once.

Full answer key →

The three patterns we just saw

• State mismatch: the AI did not know what the world looks like right now (legacy localStorage, in-flight handler, re-rendered DOM)
• Wrong primitive: the AI reached for a primitive that does not match the constraint (inline style vs class, find vs Map, local time vs UTC)
• Wrong tool: the AI was the wrong tool for the job in the first place (chat where a reasoning model was needed)

The two-step diagnostic instinct

1. Is this fixable with a better prompt? If yes, it is context. Rephrase and continue.
2. Is this fixable with a different tool? If yes, it is meta-frontier. Switch tools and continue.

Only if both are no, it is a real frontier limit and goes on the map.

When to run it

• Anything that goes to a CO
• Anything that becomes a system of record
• Anything that ends up in a fitness report
• Any data product where wrong numbers have downstream consequences

How to run it

1. Same prompt to genai.mil.
2. Same prompt to Ask Sage with a reasoning model.
3. Compare the two answers side by side.

What the result means

• Both agree on a substantive judgment → confidence increases (not to certainty, but materially).
• Tools disagree → that is a flag. Dig in. The disagreement is data about where the model is uncertain.
• Systematic disagreements → frontier-map row, probably a Capability Gap Map row.

Cost: two prompts and a comparison turn. Reserve for outputs that will be acted on.

The permission gap

Workers already using AI but hiding it. Worried about organizational reaction. Shadow culture, no shared practice. Your role as a 4.5 graduate is to formalize use, share techniques, and train others. Including tool selection.

The apprentice problem

If a junior never manually writes a key document because AI generates it, how do they develop the judgment to know when the AI version is wrong? And how do they learn which tool to reach for when the task gets harder?

Protocol for junior Marines

• Require review and explanation · they defend WHY AI's output is right or wrong
• Periodically work without AI · key tasks done from scratch monthly
• Use AI output as a teaching tool · hand them a flawed draft, have them find issues
• Rotate them through QA · judgment from exposure to bad output
• Teach tool selection deliberately · when a junior asks for help, ask which tool they reached for first and why. Coach the selection, not just the prompt. New for 4.5.

Worked example 1 · genai.mil chat

Worked example 2 · Ask Sage + Claude Opus 4.7 · new

On the floor when you leave

• Readiness Dashboard: single-file static HTML, three data sources joined, SVG bars, Not-Ready table, Commander's Snapshot. Public URL.
• Advanced Frontier Map rows: at least three rows specific to what bit you today, split by tool.

Goes back to your section and up the chain

• Workflow Playbook: one page, one workflow your section can run without you in the room. With the "Which tool" column filled.
• Capability Gap Map entries: at least two new rows mapping production-layer wishes to specific blocking permissions. Goes to the AI POC, then HQMC.