Appendix D — The Grading Contract

Nobody should be surprised in Week 8. That is the whole job of this appendix. Eight weeks is short, and there is no room in it for a student who discovers in Week 7 that the capstone expected a document begun in Week 1, that the honesty sections they skimmed were worth more than the measurement they spent all Thursday on, or that the number they took from a chat window fails the project outright rather than costing a point.

So this is the contract: every weight, every recurring rubric dimension, three full rubrics reproduced word for word, the tier arithmetic, the pace table, the submission checklist, the AI policy, the integrity line, and the two policies you will actually need — what falling behind costs, and what happens when work is late. Read it once, all the way through, in Week 1, then keep it open. It is a reference document, not an orientation.

Effort. This is a graduate course carrying roughly 12 to 15 hours a week, including the reading, the reps, and the lab. That figure is stated once in Chapter 1 and once here, and it is honest rather than aspirational. D.8 spends it for you, week by week.

A note on the noun. The weekly graded builds are labs, numbered 1, 2, 3, 5, 6, 7, plus the capstone as Lab 8. There is no Lab 4 — Week 4 carries the Placement Practical and the Checkpoint instead. Each lab’s assignment document opens with the title Project N — , and the numbers match exactly: Lab 1 is Project 1. Same thing, two names. The lab arc says “lab” throughout, because every one of them requires you to measure something and defend the number — which is not what most people mean by “project.”

D.0 — Where Everything Is

What this appendix does not own. Environment setup is Appendix A. Reading a block diagram, and the provenance of every shipped dataset, is Appendix B. Measurement methodology — warm-up, repetitions, medians, dispersion, the soak protocol, the log template — is Appendix C, and this appendix will keep sending you there rather than repeating it. Vocabulary is Appendix E.

The six deliverable filenames. Across the whole course you produce exactly six document stems and no others.

If a lab’s work needs another document, it is a section of report.docx. Do not invent a seventh stem. This is not fussiness: the site renders each stem into the document format your section actually submits, so a renamed file arrives at your instructor as the wrong kind of artifact — and the Week 8 capstone expects to find soc-architecture-review.docx exactly where you left it in Week 1.

D.1 — The Weights

These numbers are fixed. They appear identically in the lab arc, in every lab document, in the course README, and in the Canvas build notes your instructor works from. If you find a page that says something different, that page is wrong and you should say so.

D.1.1 — The quiz that is also the checkpoint

There are eight weekly knowledge checks and only seven ordinary weekly quizzes, because in Week 4 the checkpoint replaces the weekly quiz rather than sitting beside it.

There is also a Week 4 practice quiz, and it is ungraded. It draws from Chapter 4 alone and exists to warm you up. Do it — it is the fastest diagnostic you get before the checkpoint — but do not confuse finishing it with finishing the week. Chapter 4 says this too, because students miss it every term.

The checkpoint is the second-largest single item in the course, behind only the capstone and ahead of the practical, the briefing, and any individual lab. An 11.5% auto-graded assessment in the middle of an eight-week term deserves an evening of preparation, and rarely gets one.

Coach’s Note — Look at the arithmetic. The six labs are 42% and everybody treats them as the course. The quizzes and the checkpoint together are 22% — more than the capstone — and they are the only points here you can earn with an evening of honest review rather than a week of work. Students who finish at the top of a section are almost never the ones who wrote the best lab. They are the ones who did not leak the easy points.

D.2 — The Five Dimensions Every Rubric Shares

Nine graded rubrics — six labs, the Week 4 practical, the capstone, the briefing. They look different because they grade different material. Underneath they are the same five dimensions, weighted differently:

1. The measurement — taken correctly, and repeatable by a stranger?
2. The analysis — does it name a mechanism and commit to a judgment, or restate output?
3. The citation-or-measurement line — does every number trace to a primary source or your own log?
4. The honest-uncertainty line — “what I could not measure and what I would need.”
5. The disclosure line — ai-usage.txt, honest and specific.

Roughly: about half of every rubric is the analysis. Where there is a measurement, it is a quarter to a third. The three honesty dimensions together are usually 10 to 20 points — more than students expect, and the cheapest block in the course to secure, because none of it requires you to be right about anything. It requires you to be accurate about yourself.

How a rubric is actually applied. Graders do not read front to back, and neither should you when checking your own work an hour before submitting. Filenames first — a missing ai-usage.txt is a visible zero. Then measurements.xlsx, which decides whether the rest is evidence or opinion. Then every number in report.docx: measured, derived, cited, or synthetic? Then the graded core. Then the uncertainty section, then ai-usage.txt, then the extension tiers — and only after Normal is scored.

D.2.1 — Dimension 1: the measurement

Appendix C owns the method. This owns the grade.

A negative result is a full-marks result. “My laptop did not throttle in ten minutes” is real data with real causes: report it, give the two most plausible explanations, and say what would distinguish them. Quietly re-running until you get a curve you like is the failure, not the flat line.

D.2.2 — Dimension 2: the analysis (the graded core)

Every rubric has one line worth more than the others, and it is always the judgment, never the measurement. Lab 1’s is the two bottleneck predictions; Lab 3’s the phone-class prediction; Lab 6’s the two integration failures; Lab 7’s the residual-risk column, worth twenty points alone. D.3 names the graded core for every item so you can protect it.

Coach’s Note — The fastest way to move an analysis from half marks to full is to write the falsification sentence first. Before the claim, write “I would be wrong if ____, and I would find out by ____.” If you cannot fill that in, you do not have a claim — you have a preference. That is not a stylistic view of mine; it is why a prediction is worth grading at all.

D.2.3 — Dimension 3: the citation-or-measurement line

Every number is cited to a primary source or measured on your own machine. D.11 has the consequences; this has the standard. Each number carries exactly one provenance:

That taxonomy is how the headline rule — cited or measured, no third category — survives contact with real work: a derived number inherits its inputs’ provenance, and a synthetic number is not a claim about the world at all. Lab 6 asks for derived / cited / synthetic, Lab 5 for measured / modeled / cited, Lab 8 for “measured or cited?” of every figure. Use the four labels above and all three are satisfied.

Roots worth knowing — search once you are there rather than guessing at a deep path: https://developer.arm.com/ (architecture reference manuals, performance tooling), https://developer.android.com/ and https://source.android.com/ (platform APIs and the platform’s own architecture documentation), https://developer.apple.com/, https://riscv.org/ (ratified specifications and profiles), and https://csrc.nist.gov/projects/post-quantum-cryptography for the standards Lab 7’s Medium tier asks you to cite by section number. If your figure is not in that kind of place, the right answer is very often “not published” — a creditable sentence, and far better than a plausible number you could not source.

D.2.4 — Dimension 4: the honest-uncertainty line

Every rubric pays for a section most syllabi never ask for: what I could not measure, and what I would need. In Lab 1 it is “What I could not find”; in Lab 3 the limitations section; in the Week 4 practical “What I would measure,” worth ten points; in the capstone eight points, and one of the most heavily weighted single lines in the document.

Write it throughout the week, not at the end. Written last it reads as an apology; written as you go it reads as a map of the frontier, and reviewers respect it more than anything else in the document.

D.2.5 — Dimension 5: the disclosure line

ai-usage.txt, on every submission, whether or not you used a model. Graded on honesty and specificity, not abstinence. Format and worked example in D.10.

“I did not use AI on this submission” is a complete, full-marks answer if it is true. Say it plainly and sign it.

D.3 — Every Graded Item at a Glance

What it is worth, what you hand in, and what its rubric weights most heavily. Go to the item itself for the full rubric — every lab carries its own Normal-tier rubric (out of 100) table, and those tables govern.

D.3.1 — Where the deliverables recur

soc-architecture-review.docx is the only document that spans the term. Opened in Week 1, finished in Week 8, and the capstone’s heaviest single line is that it was carried forward and completed rather than written from scratch in the last week. Extend it every week as each chapter teaches you a new block.

placement-decision.docx is written three times, each on a harder problem — four workloads across core classes in Lab 2, six across CPU/GPU/NPU/DSP in the Week 4 practical, four or more on your capstone subject in Week 8. The columns shift slightly; the discipline does not: workload → chosen processor → why, naming what it beat → what it costs, as a number with a unit.

D.4 — The Week 4 Placement Practical Rubric, in Full

Reproduced verbatim from the Week 4 Placement Practical in Chapter 4, so you can read it in Week 1 rather than Week 4. If this table ever disagrees with the source document, the source document governs — and tell your instructor, because that is a bug.

Normal-tier rubric (out of 100)

And one condition that sits outside the rubric. No fabricated or uncited figure, anywhere in the submission. It carries no points because it is not a criterion you earn — it is a condition you meet. A number that traces to nothing, or a citation that does not say what you claimed it said, fails the practical regardless of everything above. This is stated in Appendix D and it applies to every deliverable in this course.

Reading it in Week 1. The lines about running the shipped programs come to 34 points; the lines about what you concluded come to 54. The practical takes four to six focused hours and most of it is writing. It is also the only graded item with no measurements.xlsx — the tools are deterministic, so what is a statistics log elsewhere is a runs/ folder here, one file per run, named so a reader can match each file to the claim it supports.

D.5 — The Capstone SoC Investigation Rubric, in Full

Reproduced verbatim from Lab 8. Read it in Week 4, when you choose your track — not in Week 8, when you are writing.

Normal-tier rubric (out of 100)

The integrity line (pass/fail, applied over the rubric). A fabricated figure, a citation that does not resolve, a quotation nobody wrote, a measurement you did not take, or synthetic data presented as real, fails this project outright. This is not a deduction. It is the one rule in this course with no partial credit, because a measurement report is an instrument, and a crooked instrument is worse than no instrument at all.

Reading it in Week 4. Three of those lines are earned in other weeks: the 12 for soc-architecture-review.docx come from extending it since Week 1, the 10 for real measurements from the habit built in Labs 2, 3 and 5, and the 6 for a stated question from choosing narrowly and early. Students who begin with “I’ll compare these two parts” write surveys; students who begin with a question write investigations — and the investigation is the shorter document.

D.6 — The Technical Briefing Rubric, in Full

Reproduced verbatim from the Technical Briefing in Chapter 8. It is a 12-minute recorded briefing (11:00–13:00 accepted) followed immediately by a 3-minute addendum (2:30–3:30 accepted) answering three challenge questions that are assigned in advance, are the same for every student on every track, and are printed in the briefing document. Read them in Week 5.

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.

Reading it in Week 5. The three challenge questions are worth 24 of 100 and are published four weeks early on purpose: preparing for them improves the investigation itself. CQ1 asks for the choice that most flatters your conclusion, which is exactly what the capstone’s Medium tier M2 — the metric-sensitivity analysis — produces as a by-product. Do M2 and you have written CQ1.

D.7 — The Tiers, and Why Hard Is Always a Memo

D.7.1 — The extra-credit arithmetic

A lab is scored out of 100 on its Normal rubric. Medium can add up to 25 of those points, Hard another 25. A lab with a perfect Normal tier and both extensions fully earned records 150 out of 100, carrying that lab’s ordinary weight — so a 7% lab at 150 contributes 10.5 points to your course total instead of 7. Four rules govern it, and the first is the one people get wrong.

1. Medium and Hard are extra credit, not substitutes. You cannot buy back Normal points with Medium work. A thin Normal tier with a brilliant memo attached is scored as exactly that. Finish Normal first, every week.
2. Extra credit carries the item’s weight. Extension work on a 20% capstone is worth roughly three times the same work on a 7% lab. If you do one Hard tier all term, do the capstone’s.
3. A tier is scored only if it is complete. The Hard tiers are explicit about what fails them — a memo listing five improvements without costing any, a memo that says “it depends,” a memo whose cost figure has no measurement behind it. Half a memo is not half the credit.
4. Whether your gradebook caps the course total at 100 is a Canvas setting your instructor owns. Check your syllabus. The arithmetic above is the book’s; the ceiling is theirs.

D.7.2 — Why Hard is never “more measurement”

Judgment is the thing being trained. A desktop architect asks how fast it can go. This course asks four questions at once — performance, energy, thermals, placement — and holding them simultaneously produces a decision, not a number. You can be handed a measurement. Nobody can be handed a decision that commits to a loser.

More measurement would break the course’s hardware promise. Every Normal-tier requirement is completable on Workbench B — a browser development environment, no install, no admin rights, no phone. If the Hard tiers were extra measurement they would quietly reward students who happen to own an Android device, and the guarantee that a locked-down laptop can earn full marks would become a guarantee that it can earn most of them. A memo costs nothing but thinking and is equally available on every workbench.

It is the part a tool cannot do for you. A model can enumerate options faster and more completely than you can. What it cannot do is decide, for a named product serving named people under a named constraint, which option is not worth its price — and then sign its name to that. It has nothing at stake and cannot be wrong in the way that matters.

D.7.3 — The memo’s shape

All eight Hard tiers ask for the same six moves in different clothes. Use this as a checklist.

Two labs add a seventh move worth stealing for all of them. Lab 1’s memo asks you to say whether what you propose to change is a law — physics, the skin-temperature limit, the energy cost of driving a wire — or a decision somebody made and could unmake; you may only argue with the second kind. Lab 7’s asks who bears the risk, and whether they are in a position to know they are bearing it. That is the sentence most memos omit and the one an ethics review reads first.

D.8 — The Pace Table

D.8.1 — What is due, week by week

Everything is due end of Friday of its own week unless your section’s syllabus says otherwise.

D.8.2 — Where the 12 to 15 hours go

Approximate; the shape matters, not the decimals.

Four observations that will save you a week.

Writing is the largest column in every row. Not measuring. If your week is six hours of running things and ninety minutes of typing, you have inverted the course and your grade will show it — about half of every rubric is the analysis.

Week 3 is the heaviest week in the table; Chapter 5 is the most consequential. Week 3 is the only row that reaches the top of the 12-to-15-hour range, and the measurement is why: three programs — the cache walk, the pointer chase, the loop-order matrix multiply — each run three separate times, with a prediction written down before each. Chapter 5 re-reads Chapters 1 through 4 through the energy lens, which makes a shaky Week 3 far more expensive than the table suggests.

The capstone is not a Week 8 project. Lab 8 tells you to check whether you can already answer the three challenge questions from what is in your report, and says you have four weeks to fix it if you cannot. Those four weeks are Weeks 5 through 8, which is why the table spends one to two hours a week on the capstone from Week 5: naming the question, choosing the track, extending soc-architecture-review.docx. A student who first opens the capstone document on the Monday of Week 8 is attempting a 20% investigation and an 8% recorded briefing in one week, on top of a chapter and a quiz. It does not fit. That is not a warning; it is arithmetic.

The reps are not busywork. Every lab says it in its own words: the reps are the project, done small. Do them and the lab is an afternoon of assembly; skip them and it is a week.

D.9 — The Capstone Submission Checklist

D.9.1 — What you submit

One link or one archive, containing exactly five documents at the top level with these exact names, plus your raw output and your slides.

<your-capstone-repo>/
├── report.docx                      the investigation
├── measurements.xlsx                the measurement log, Appendix C form
├── soc-architecture-review.docx     the Week 1 document, carried forward and finished
├── placement-decision.docx          the placement table and its defense
├── ai-usage.txt                    the honest disclosure
├── runs/                          raw harness output and counter captures, unmodified
├── slides.<pdf|pptx|key>          the briefing deck
└── README.txt                      track, subject, workbench, tier targeted

Raw output may live inline in measurements.xlsx instead of in runs/ — either is fine, as long as a reader can match every claim to the file that supports it. Do not invent a sixth deliverable document; if you need another section, it is a section of report.docx. The briefing recording is submitted as a file or an accessible link alongside this.

D.9.2 — The pre-flight checklist

Do this the day before, not the hour before. Each line is a rubric line in D.5 or D.6.

The measurement log

• Every row: a median and a dispersion figure. A median with no dispersion is half a measurement.
• Every row: machine, compiler and optimization level, power state, thermal condition, repetition count.
• Warm-up runs recorded as discarded, not silently dropped. Raw output committed unmodified.
• At least three distinct kernels or scenarios, each warm-up plus 21 or more timed repetitions, at least one of them a kernel you wrote or adapted.

The numbers

• Every number in report.docx: measured, derived, cited, or synthetic? If none of the four — delete it or source it. There is no third option.
• Every cited figure has a URL and a retrieval date. Click every link; a page that does not contain the figure you attributed to it is the same failure as no citation.
• No bare single numbers anywhere, including in report.docx.
• Every headline number carries its measurement window and thermal condition. If you report a peak, report a sustained figure beside it or say why you could not obtain one.
• Wherever you used the shipped comparison dataset or sample counter output, the sentence that uses it says synthetic and says what that limits.

The documents

• soc-architecture-review.docx is the Week 1 document finished — block inventory, cluster configuration, memory configuration with the bandwidth arithmetic shown, accelerator inventory, storage and radios, every figure marked with its source.
• report.docx opens with the stated question, the subject, the constraint, and what is out of scope.
• placement-decision.docx: four or more workloads, a cost with a unit on every row, at least one rejected placement recorded.
• The bottleneck analysis answers all Four Questions separately, with a heading each, so nobody can mistake a performance answer for an energy answer.
• Counter analysis: IPC plus at least one miss rate per kernel, multiplexing noted.
• “What I could not measure and what I would need” names instruments and experiments, not feelings.
• ai-usage.txt present, specific, signed. (Track 1) No sentence ranks a vendor rather than a choice. (Hard) The memo names what it gives up and what would reverse it.

The briefing

• Briefing 11:00–13:00; addendum 2:30–3:30; one recording; the boundary announced on camera; the addendum one continuous take.
• The claim, metric named before finding, stated before 1:00.
• Every performance chart shows its measurement window and thermal condition on the chart; any slide derived from the shipped synthetic datasets says synthetic on it.
• All four questions said out loud, distinguishably. The addendum names each challenge question before answering it, and CQ3 puts something in all three buckets.
• The “Briefing notes” timestamps in report.docx are filled in. A blank line is a missing element — that is what the block is for.

D.10 — The AI-Use and Disclosure Policy

This is the most important section in this appendix, and the one rule of the course you will be tempted to break by accident rather than on purpose.

D.10.1 — The rule

Use it to explain, never to source.

Use a language model as much as you like to explain a mechanism. Do not use one as the source of a value.

D.10.2 — Why, exactly

Ask a model for the last-level cache size of a specific mobile core, the peak clock of a specific part, or the key sizes in a specific standard, and here is what you get: a number that is specific, plausible, confidently stated, formatted exactly like a fact — and quite possibly invented outright. Not hedged. Not flagged as uncertain. Rendered in the same flat, competent register as a number it actually knows.

That last clause is the whole problem, and it is worth being precise about why.

A mechanism is a pattern. Why a randomized pointer chase defeats a prefetcher, why ikj beats ijk, what break-even residency means, why a tile-based renderer avoids the framebuffer round trip, what an exception level is — these appear thousands of times in the material a model was trained on, stated consistently, because they are consequences of how the machine works rather than facts about one part. A model is genuinely excellent at them, and more patient than any textbook: you can ask the same question four different ways at eleven at night until it lands.

A value is a single fact about one product in one generation. It appears rarely, inconsistently, often in tables the model saw as flattened text, often wrong in the source. And the model’s job is to produce a fluent continuation either way. When it knows, it produces a number. When it does not, it produces a number, because a number is what fluently continues “the L2 cache on that core is.” There is no internal marker separating the two — and, the part that costs students the course, no external one either. The tone is identical, the confidence is identical, the formatting is identical. You cannot tell by looking, which means you cannot tell at all.

So: an excellent explainer of mechanisms, an unreliable source of values. The policy follows directly. It is not a moral position about AI; it is a statement about which of its outputs you can verify at a glance and which you cannot.

Chapter 7’s reps make you prove this deliberately: ask a model for the post-quantum key and signature sizes before you look them up in the standards, then look them up, then count how many it got right — and, the real question, ask whether you could have told the wrong ones from the right by tone. Almost nobody can. A rule you have tested yourself is a rule you keep in Week 8.

D.10.3 — The zones

Green — always allowed. Disclose it in ai-usage.txt like everything else.

Amber — allowed, with disclosure, and only if you read every line and can defend it.

Amber items go in ai-usage.txt by name. That is the entire cost of using them, and it is a fair price.

Red — fails. Not a deduction on a line; see D.11.

The line between green and red is exactly this sharp, and Lab 1 states it in one pair of sentences worth memorizing: “Explain what a system-level cache does and why it sits in front of the memory controller” is a great question to ask a model. “How big is the system-level cache on part X” is the question that will cost you the project.

D.10.4 — The ai-usage.txt format

One file per submission — not one per document. A short header block and one honest paragraph. Length is not the point; specificity is.

# AI Usage — <lab number and name>

**Models used:**            <product name and version as the product reports it, or "none">
**Dates:**                  <when>
**Zone:**                   green / green + amber (name the amber items)
**Figures sourced from a model:**  none
**Signed:**                 <your name>, <date>

<One paragraph. What you asked, what you did with the answer, and — the part
almost everyone omits, and the part that is graded — anything it told you that
you had to verify, override, or discard.>

Four rules about that file.

1. It is graded on honesty and specificity, not abstinence. A submission disclosing four uses honestly scores better than one claiming none and vague about the code it did not write.
2. The “what I discarded” sentence earns the line. A disclosure with no friction in it reads as a disclosure nobody thought about.
3. “I did not use AI on this submission” is a complete, full-marks answer if it is true. Write it, sign it, move on.
4. Never write “I used AI for research.” Research is exactly the activity the policy prohibits, and the sentence tells a reader nothing except that you did not read this section.

D.10.5 — A worked example

Full marks on the disclosure line. One paragraph, three specific uses, and the third is the reason the file exists.

# AI Usage — Lab 3, The Memory Wall Lab

**Models used:**            one general-purpose assistant (product name and version
                            as reported in the product's own interface)
**Dates:**                  Tuesday and Thursday of Week 3
**Zone:**                   green, plus two amber items named below
**Figures sourced from a model:**  none
**Signed:**                 A. Student, <date>

I used the assistant three times. First, green: I did not understand why a randomized
pointer chase defeats a hardware prefetcher, so I asked for the mechanism twice, in
different words, until I could restate it without the transcript in front of me; the
explanation was consistent with the chapter and I took no number from it. Second,
amber: I asked it to restructure my measurement log so run conditions were a column
rather than a paragraph, then read the result, renamed two headings, and confirmed
every value still came from my own raw output files. Third, and this is the one worth
recording: I asked what the last-level cache size was on my laptop's processor, and
it gave me a specific figure, stated flatly, with no hedge at all. I could not find
that figure on the vendor's own specification page for my exact part, so I discarded
it and reported the boundary I inferred from my own data, with a confidence
statement. Its number was close to my inferred boundary, which is precisely why
discarding it was right — a plausible number I cannot source is more dangerous to me
than an obviously wrong one, because I would not have gone looking.

D.11 — The Integrity Line, Stated Plainly

Every number in every deliverable is cited to a primary source or measured on your own machine. There is no third category.

That is the whole rule. Its consequences are graded in two very different ways, and the difference matters enormously.

Why fabrication fails rather than deducts. A measurement report is an instrument. Almost nobody who reads yours will be able to rerun it — they will trust the instrument, and the instrument is you. A crooked instrument is worse than no instrument at all, because no instrument produces caution and a crooked one produces confidence. That is also why the consequence cannot scale with the size of the lie: a report with one invented figure and nine honest ones is not 90% trustworthy. It is a document whose numbers all have to be checked by hand, which is the same as a document with no numbers.

Why unrepeatable data is not evidence. This catches honest students. You ran the benchmark, the number is real, you wrote it down — but you did not record whether you were on battery, what else was running, how many repetitions, or the thermal state, and now nobody, including you next Tuesday, can produce that number again or say what it was a number about. Nobody will accuse you of anything. It simply cannot do the job you wanted it to do. That is why measurements.xlsx has the shape Appendix C gives it, and why “conditions” is a graded column rather than a courtesy.

The correct move, every time, is the cheap one. Write “I could not establish this” or “not published by the vendor” and move on. Vendors publish what helps them sell; cache sizes, interface widths, sustained power envelopes and thermal design targets are frequently absent, and saying so demonstrates exactly the judgment being graded. That sentence has never once cost a student a grade in this course. An invented figure has.

Coach’s Note — Nobody fails this line on purpose. It happens like this: it is Thursday night, you have one gap in a table, you half-remember a number, it sounds right, and filling it takes four seconds while sourcing it takes twenty minutes you do not have. Every single time. The defense is mechanical, not moral: write the citation in the same motion as the number. Every time you type a figure, type the URL and the date beside it before you type anything else. Reconstructing citations afterward is miserable, and it is exactly how honest people end up with unsourced figures.

D.12 — What Falling a Week Behind Costs

One week is one eighth of this course. The direct cost is 8.5% — a 7% lab and a 1.5% quiz — roughly a full letter grade in most schemes. The compounding cost is larger and less obvious:

1. The reps are the lab, done small. Skip a week’s reps and the lab stops being an afternoon of assembly and becomes a week of construction. You do not save three hours; you spend six.
2. soc-architecture-review.docx is cumulative, and the capstone’s heaviest line is that it was carried forward and finished. A week you do not extend it is a week you pay for in Week 8, while also writing an investigation and recording a briefing.
3. Week 5 re-reads Weeks 1 through 4 through the energy lens. It does not work on a shaky Week 3.
4. The capstone consumes everything — measurements you learned to take in Weeks 2, 3 and 5, a placement table you learned to build in Weeks 2 and 4, a bottleneck analysis you have rehearsed since Week 1. A missing week does not vanish; it reappears in Week 8 at a worse exchange rate.

And here is the arithmetic people avoid. To catch a missed week up while also doing the current one you need 24 to 30 hours in a single week. You do not have them, and you knew that before you read this sentence. So you do not catch up by working harder. You catch up by cutting scope deliberately — a professional skill, graded in miniature all term.

D.12.1 — The triage order

If you have less than a full week’s hours, spend them in this order. Do not spend them evenly.

Three rules override the table. Never skip the measurement to save time — it is the one part you cannot reconstruct later, it is what the analysis is made of, and it is usually the smallest column in your week anyway. Protect soc-architecture-review.docx above everything else, because it is the only artifact whose absence compounds. And tell your instructor in writing before Friday, not after: “I will submit Normal-tier only, without the Medium section, and here is why” is a professional communication and is usually answerable. A silence followed by a request three weeks later is neither.

D.13 — Late Work, Resubmission, and the Week 4 Checkpoint

Your section’s syllabus governs where it differs from this. Where it is silent, this is the policy.

D.13.1 — Due dates and late work

Everything is due at the end of Friday of its own week, in your section’s time zone: the weekly labs, the Week 4 practical, the capstone, the briefing, and the weekly quizzes, which close at the same moment.

An extension requested before the deadline, with a reason and a proposed new date, is a different conversation from a late submission, and instructors treat it as one. Ask early.

D.13.2 — Resubmission

One lab per term may be resubmitted once, under four conditions: the original was submitted on time and in good faith (an honest attempt at every required item, not a placeholder); the resubmission arrives within seven days of the graded original being returned; it is capped at the Normal tier’s full marks, with no extra credit on a second pass; and it carries two or three sentences saying what changed and why.

Resubmission exists to fix an analysis, not to supply work you never did. Rewriting a residual-risk column that hedged, re-deriving a placement row you got backwards, adding dispersion figures you left out — that is what it is for. Running the experiment for the first time is not a resubmission; it is a late submission, and D.13.1 covers it.

An integrity failure is never resubmittable. D.11 is a pass/fail condition, not a rubric line, and a document that failed it cannot be repaired by adding a citation after the fact — because the question was never whether the number could be sourced, but whether it was when you published it.

D.13.3 — The checkpoint as an early-warning signal

The Week 4 checkpoint is auto-graded, cumulative across Weeks 1 through 4, and it replaces Week 4’s ordinary weekly quiz rather than sitting beside it. The Week 4 practice quiz, drawn from Chapter 4 alone, is ungraded and is a genuinely useful warm-up. Do it first; do not confuse it with the checkpoint. Each checkpoint attempt draws a random subset from a larger pool, so a retake is real practice rather than memorization — whether your section allows more than one attempt is a Canvas setting your instructor owns, so ask rather than assume.

And here is what the checkpoint is for, which is not the 11.5%. It lands at the exact midpoint, on cumulative material, with four weeks left. It is the moment the course tells you honestly whether the first half landed, at the last point where the answer is still actionable. Read your score as a diagnostic, not a verdict.

D.14 — The Contract in One Page

Roughly 12 to 15 hours a week. Six labs, numbered 1, 2, 3, 5, 6, 7 — there is no Lab 4. Week 4 carries the Placement Practical and the Checkpoint, and the Checkpoint stands in for that week’s quiz rather than sitting beside it.

The nine rules.

1. Every number is cited to a primary source or measured on your own machine. No third category. Derived numbers inherit their inputs’ provenance; synthetic numbers are labeled synthetic where they are used.
2. An uncited figure scores zero on its line. A fabricated one fails the project outright, and no quality of write-up compensates.
3. A measurement nobody can reproduce from your stated conditions is not evidence. Median, dispersion, repetitions, conditions — or it is not a result.
4. Use AI to explain, never to source. Green: mechanisms, prose review, practice questions, rehearsal. Amber: drafted code you read line by line, disclosed. Red: sourcing a figure, generating a measurement, writing an analysis you did not perform.
5. ai-usage.txt goes with every submission, whether or not you used a model, graded on honesty rather than abstinence.
6. Six document stems, and no others. Anything else is a section of report.docx.
7. Normal is the grade; Medium and Hard are extra credit, and Hard is always the memo — one recommendation, its arithmetic, its cost, who loses, the strongest case against it, and the observable condition that would reverse it.
8. Every Normal-tier requirement is completable on Workbench B. No phone, no admin rights, no installation, no exceptions.
9. The graded core of every rubric is the judgment, not the measurement. Half of every rubric here is what you concluded and how honestly you bounded it.

And the one sentence underneath all nine. You are being trained to produce the third document — not the vendor’s deck with its single number, not the competitor’s rebuttal with a different single number, but the one somebody can actually decide from. It has fewer numbers than either of the others, every one of them carries its conditions, it says plainly what could not be measured and what that would take, and it ends in a recommendation its author is willing to defend to a room that would prefer it were simpler.

Everything in this appendix is a mechanism for making that document the one you habitually write.

Read this again in Week 4. See you at the briefing.