The Technical Briefing — A 12-Minute Recorded Defense

“A false balance is an abomination to the LORD, but a just weight is his delight.” — Proverbs 11:1 (ESV)

Chapter: 8 — Measure, Compare, Decide Weight: 8% of the course grade. (The investigation it defends, Project 8, is a separate 20%.) Due: End of Week 8 Submit: One recording — a single video file or a link to one — containing a 12-minute briefing followed immediately by a 3-minute addendum answering the three challenge questions printed below. Plus your slides as a file. Plus one short report.docx section titled “Briefing notes” recording the timestamps of each required element (see Submission). Allowed tools: Any screen recorder or conferencing tool that captures your slides and your voice. Slides in any format. You may have notes in front of you. Everything in Appendix A is available for producing the figures you show. AI policy: You may use a model to rehearse — to ask you hard questions, to critique your slide order, to tell you a sentence is unclear. You may not use one to write the briefing, and you may not read a generated script aloud. Every figure on every slide traces to your own measurement or a primary source already cited in your investigation; no new uncited number may appear on a slide. Disclose your rehearsal use in ai-usage.txt alongside the rest.

The Setup

Your investigation is written. Now four people who did not read it have thirty minutes, and you get twelve of them.

This is not a formality and it is not a summary exercise. It is the moment the work becomes usable by someone else, and it is a distinct skill from writing the work. A written report lets a reader stop, reread, and check a citation. A briefing does not. Everything has to be load-bearing, in order, at a pace a listener can follow, with the conditions attached to each number at the moment you say the number — because there is no footnote in a spoken sentence.

The audience is a skeptical engineering audience, and you should imagine them specifically. Somebody in the room has already seen the vendor’s chart and half-believes it. Somebody else ran a version of your experiment once and got a different answer. Somebody is going to ask what the dispersion was. Nobody is hostile; all of them are busy, all of them are competent, and none of them will accept “it’s faster” without asking measured how, for how long, on what.

It is recorded, which is a gift. You get to watch yourself, notice the forty seconds you spent on a slide nobody needed, and do it again. Take that gift. Almost nobody’s first take is their best one.

What You Are Delivering

One recording, two parts, in this order, without a cut between them:

Total submission: roughly 15 minutes. The addendum is not optional and is worth a quarter of the rubric. Announce it on camera — “that’s the briefing; here is the three-minute addendum” — so the boundary is unambiguous.

You will not have a live audience, so you must simulate the pressure yourself: put a visible timer in the recording or on your slides, and do not edit the addendum. One take for the addendum is the point of the addendum.

A shape for the twelve minutes

This is a suggestion, not a rubric line — but the rubric rewards everything in it, and briefings that wander tend to be briefings that skipped step one.

• 0:00–1:00 — The claim. One sentence, metric named before finding, with the conditions. Then the scope: what this is about and what it is not. Nothing else. Do not open with an agenda slide; do not open with a history of mobile computing. The room needs to know in sixty seconds what you are going to tell them.
• 1:00–3:00 — The subject and the question. The part, workload or product; the constraint you held; why this question is the one worth answering. One block diagram or one specification table, annotated by you, not pasted.
• 3:00–6:00 — The method. What you measured, how, how many repetitions, under what conditions, with what instrument. Show the harness or the log. This is the section engineers actually listen to, and the one students underplay because it feels like plumbing. It is not plumbing. It is your credibility.
• 6:00–9:00 — The evidence. Three results, on screen, each with median, dispersion and conditions visible. Then the mechanism: counters, structure, or both — connected to something this course taught, by chapter number.
• 9:00–11:00 — The Four Questions and the placement. Performance, energy, thermals, placement, each addressed. Defend at least one placement row out loud including its cost.
• 11:00–12:00 — Limits and decision. What you could not measure and what you would need. Then the recommendation, with its tradeoff named and the condition that would reverse it.

On slides

Fewer than you think. A dozen is plenty; twenty is too many for twelve minutes. Every chart that shows a performance number must show its measurement window and thermal condition on the chart, not in a spoken aside — this is a rubric line and it is the single most common thing missing from student decks. Label axes. Use your own figures. If you show anything derived from code/soc-compare.csv or code/sample-perf-stat.txt, the slide must say synthetic on it.

The Three Challenge Questions

These are assigned in advance to every student on every track, and they are the same three for everyone. That is deliberate. They are not a trap; they are the three questions this course exists to make you able to answer, and preparing for them will improve the investigation itself. Answer all three in the addendum, in this order, naming each one before you answer it.

You have three minutes for all three. That is roughly one minute each, which is deliberately not enough to be vague.

Challenge Question 1 — The flattering choice

Name the single measurement or presentation choice in your investigation that most flatters your conclusion. Then state your headline claim again with that choice reversed.

Every investigation has one. A measurement window. A summary statistic. A cooldown policy. Which build. Which cores. Which of your workloads got weighted most heavily. Which axis the chart starts at. The choice does not have to be wrong — most of them are entirely defensible, which is exactly what makes them dangerous, and §8.13 is about precisely this.

A full answer has three parts: name the choice, say why it was defensible when you made it, and state the number your headline claim becomes when you reverse it. Then say whether your conclusion survives.

“I don’t think I made one” is not an answer and scores zero. If you genuinely cannot find one, you have not looked at your own work adversarially, which is the skill being tested.

Challenge Question 2 — The placement row you are least sure about

Take the row of your placement table you are least confident in. Argue the opposite assignment as strongly as its advocate would. Then name the one measurement that would settle it.

Not a straw version. Steelman it: give the best case for putting that workload on the other processor, including the cost it would save and the constraint it would satisfy. Then say why you nevertheless chose what you chose — or, if arguing the other side has changed your mind, say that, which is a strong answer and not a weak one.

Finish with the measurement. Be specific: what you would run, on what, with what instrumentation, and what result would move you. “More testing” scores nothing. “A CPU-only control at the same input size and thermal state, plus the per-operator assignment from the runtime, would tell me whether the accelerated path is actually running the whole graph” scores full marks.

Challenge Question 3 — Halve the envelope

Cut the sustained thermal envelope of your subject in half. Which of your conclusions survive unchanged, which invert, and which can you not answer without new measurement?

Three buckets, and you must put something in all three. Halving the sustained envelope is not a hypothetical — it is roughly the difference between a tablet and a phone, or a phone and a wearable, and it is the fastest way to find out whether your reasoning was about architecture or about a particular operating point.

Survive: which conclusions were about structure and therefore hold. Invert: which were secretly about peak, or about a headroom you no longer have. Cannot answer: which depended on a measurement you took at the original envelope and have no way to extrapolate — and say what you would need. The third bucket is the one that separates strong answers from confident ones. A student who claims all their conclusions survive has not understood Chapter 5.

Recording Requirements

• Audible. Test your microphone before the real take. A brilliant briefing nobody can hear scores what an inaudible briefing scores.
• Legible. Slides readable at the resolution you record at. Small-print charts are the same as no charts.
• You are visible or you are not — either is fine. What matters is the voice and the slides.
• A visible timer, or an on-screen clock, or spoken time checks — something that lets a grader see the 12-minute and 3-minute boundaries.
• No script read verbatim. Notes are fine; reading is audible, and it costs the delivery line. Speak from an outline.
• Do not edit the addendum. Cuts within the briefing are acceptable if the timer stays coherent. The addendum is one continuous take.
• One file or one link. If a link, make sure it is accessible without a login.

Normal-tier rubric (out of 100)

The integrity line (pass/fail, applied over the rubric). A figure on a slide that appears nowhere in your investigation and carries no primary source, a claimed measurement you did not take, or synthetic data presented as real, fails this exam — the same rule as the capstone, for the same reason. Restating a cited figure from your report is fine and expected; introducing a new uncited one is not.

Medium Tier (+up to 25% extra credit)

M1. The two-minute version

Record a separate 2-minute cut — the version you would give in a hallway to someone who has ninety seconds and one question. Same claim, same conditions, no method section, no slides required. Submit it alongside the main recording.

This is much harder than the twelve-minute version and it is where the compression skill actually lives. It must still name its metric before its finding and still state one condition. If you can do this well, you can brief anybody.

M2. The peer cross-examination

Have a classmate or colleague watch your briefing and prepare three questions you did not expect. Record yourself answering them cold, unedited, in under four minutes, and submit that with a one-paragraph written note: which question you handled worst, and what you would change in the investigation as a result.

The note is the graded part. “They asked about my cooldown policy and I realized I never wrote it down” is worth more than three smooth answers.

Hard Tier (+up to 25% additional extra credit)

H1. The re-record, with the diff

Watch your own briefing all the way through, with a notepad, as if you were the skeptic in the room. Identify the weakest two minutes — the place where a listener would stop believing you, or stop following you, or stop caring.

Then do three things:

1. Write the critique, half a page, in the voice of the skeptical engineer. Be specific and be harder on yourself than a grader would be. “Slide 7 shows a 22% difference with no dispersion anywhere on screen, and I moved past it in nine seconds.”
2. Re-record those two minutes, fixed. Submit both versions, clearly labeled.
3. Write the diff, one paragraph: what changed, and — the graded sentence — what the fix revealed about the investigation itself. Often the weakest two minutes of a briefing are weak because the underlying work was thin there, and the honest version of this tier says so.

This is the human-in-the-loop thesis of the whole course, applied to yourself. A model can polish your phrasing. It cannot watch you and know which two minutes a room stopped believing you, because that judgment requires having been in rooms and having something at stake in this one. Make the judgment, act on it, and say what it cost you to admit.

Submission

1. The recording — one file or one accessible link, briefing then addendum.
2. The slides — as a file, in any common format.
3. A “Briefing notes” section appended to report.docx containing:

## Briefing notes

**Recording:**        (file name or link)
**Briefing length:**  mm:ss
**Addendum starts:**  mm:ss     **Addendum length:** mm:ss
**Claim stated at:**  mm:ss     (should be under 1:00)
**Four Questions at:** performance mm:ss · energy mm:ss · thermals mm:ss · placement mm:ss
**Placement row defended at:** mm:ss
**Limits stated at:** mm:ss
**CQ1 at:** mm:ss   **CQ2 at:** mm:ss   **CQ3 at:** mm:ss
**Medium/Hard artifacts:** (file names, if any)
**AI usage (rehearsal):** (one line; full disclosure in ai-usage.txt)

Those timestamps are not busywork. Filling them in is the fastest way to discover, before you submit, that you never actually said the thermals answer out loud.

Hints (Read Before You Begin)

• Write the first sentence before you make a single slide. If you cannot state your claim in one sentence with its metric first, you do not have a briefing yet — you have a pile of results. Fix that before you open a slide tool.
• Build the deck backwards from the evidence slides. Pick the three results you will show, put the conditions on them, and then work out the minimum you must say to make those three land. Everything else is cuttable and most of it should be cut.
• Rehearse with the timer on, out loud, standing up. Silent read-throughs run about thirty percent fast. Your first out-loud run will be long; that is normal and it is information about what to cut.
• Say the conditions in the same breath as the number. “Two point four times faster — median of thirty-one runs, plugged in, after a ten-minute idle.” It sounds fussy the first time and completely natural by the third.
• Prepare the three challenge answers as bullet points, not sentences. One minute each is not enough to recite a paragraph, and recited paragraphs sound recited. Know your three points per question and speak them.
• For CQ1, find your flattering choice now, while you can still do something about it. Medium tier M2 of the capstone (the metric-sensitivity analysis) essentially writes this answer for you.
• Watch one take all the way through before re-recording. Almost everyone’s instinct is to restart at the first stumble; almost everyone’s actual problem is somewhere in minute eight.
• Do not apologize on camera. Not for your voice, not for your slides, not for the scope. State what you did and what it supports. If something is a limitation, it goes in the limits section as a fact, not in the delivery as an apology.

What Mastery Looks Like (Beyond the Rubric)

A great briefing has a quality that is hard to fake and easy to recognize: the audience always knows why they are looking at the current slide. Each one earns its place because the previous one raised a question it answers. The claim arrives first, the method establishes the right to make it, the evidence supports it with its conditions attached, and the limits arrive before anyone has to ask — which is what makes the recommendation at the end land as a decision rather than an opinion.

The addendum is where the grade separates. A strong CQ1 answer is uncomfortable to give and the speaker gives it anyway, plainly, and then says whether the conclusion survives. A strong CQ2 argues the other side so well that the listener briefly wonders which position the speaker holds. A strong CQ3 puts something honest in the “I cannot answer this” bucket and names the experiment that would fill it.

That is the whole professional posture this course has been building for eight weeks, compressed into three minutes: hold your conclusion firmly enough to recommend it, and loosely enough to say exactly what would change it.

Coach’s Note — The instinct under questioning is to defend. Resist it. The best answer you can give a skeptical engineer is often “that’s the weakest part, here’s why, and here’s the measurement that would settle it” — because it tells the room that you know where the edges of your evidence are, which is the only reason to trust what is inside them. Confidence without stated limits does not read as strength to people who know the field. It reads as someone who has not checked.

When You’re Done

1. Watch the whole recording once, straight through, without pausing. Note the timestamp of every moment you drifted.
2. Confirm you said the claim, with its metric first, before the 1:00 mark. If not, re-record the opening — it is sixty seconds and it is eight points.
3. Confirm every performance chart on screen shows its measurement window and thermal condition on the slide.
4. Confirm you said the words for all four questions — performance, energy, thermals, placement — and that a listener could tell them apart.
5. Confirm the addendum names each challenge question before answering it, and that CQ3 has something in all three buckets.
6. Fill in the “Briefing notes” timestamps. If any line is blank, that element is missing; fix it before you submit, not after.
7. Submit the recording, the slides, and the updated report.docx.

A theological footnote. This whole week has been about instruments — the merchant’s stone of Proverbs 11:1, “A false balance is an abomination to the LORD, but a just weight is his delight” (ESV), and then the benchmark, the counter, the chart. A briefing is the last instrument in the chain and the least inspectable of them all, because a listener cannot pause a spoken sentence to check a citation. Everything the room believes about your work, they will believe on the strength of how you said it. That is real power and it is worth naming as such. The ancient world’s remedy for it was not better auditing; it was character — the just weight, kept honestly, in a market where nobody was checking. Say the number with its conditions. Name the choice that flattered you. Put something in the bucket labeled I could not answer this. Do those three things and you will be trusted for the right reason, which is the only kind of trust worth having in a field where almost nobody can rerun your experiment.

That is the course. Go record it, and go do this work well for a long time.

See you out there.