Chapter 3 templates · Ladder Self-Rating Sheet + Seven-Step Workflow Check Card

This appendix is the fillable version of the two maps in Chapter 3.

Template 1 · Ladder Self-Rating Sheet (Step × Level Matrix)

How to use. Score one real project you have in hand, not yourself and not your tools. Fill two cells per step, the current level (how it is actually done right now) and the target level (where you want to climb after reading the matching Part II chapter). Higher is not better for the target. For some steps the right target is to stop at assistant level (see the section 3.5 snapshot). Refill it each time the project goes around the loop.

Level quick reference (read this first, then fill the table)

Use the four criterion questions to rule on the level. When an answer changes, the level changes.

Level	Who drafts	Who reviews	Who decides	Who answers when it's wrong
Tool	You	You (a glance in passing)	You	You; errors visible on the spot
Assistant	AI	You, every part in full	You	You; human review is the only line of defense
Collaborator	AI (with self-checks and alternatives)	You, checkpoints + spot checks + a locked-in process	You	You + the process; errors hit the criteria first
Autonomous researcher	AI (including intermediate decisions)	Mostly AI self-review, the human accepts only the end product	Goals and acceptance criteria stay with the human, the process goes to AI	Nobody (enter serious settings with care)

Self-rating matrix

Project name: __ Date filled: _ Fill-in round _

Step	Current level (tool/assistant/collaborator/autonomous)	Target level	Upgrade precondition, what locked-in criterion or process this step needs first before you dare go up one level	Who (or what) found the last error at this step
1 Master the field
2 Questions and hypotheses
3 Test plan
4 Execution
5 Read and catch errors
6 Deliver
7 Red team

Circle two cells. The step I most want to climb: __. The step I should least let go of (my "operating room"): ____

Self-check

[ ] All seven steps filled with the same level? You are probably scoring "the whole thing." Levels live on "step × task." Go back to section 3.4 and refill.
[ ] A step marked collaborator, but the "upgrade precondition" column has no locked-in criterion at all? Drop it back to assistant. Without criteria you have no standing to spot-check.
[ ] Read and catch errors marked collaborator or higher? Be wary. In the 2026 snapshot no step's errors disguise themselves better than this one's.
[ ] The whole target column says "autonomous"? Reread section 3.4. Turning every knob to maximum is not advanced, it is an operating room that fails the sterilization standard.
[ ] A cell where you cannot answer "who answers when it's wrong"? Use that cell at tool level, without exception, until you can.
[ ] The "who found the error" column is all "I happened to notice"? You have no process-level verification yet, and no step should go above assistant level.

Template 2 · Seven-Step Workflow Check Card

How to use. One card per step. Fill them all at project start, and refill the matching card when you get stuck or sent back. A blank you cannot fill is itself the diagnosis. It tells you where you are really stuck. How to do each step is in the matching Part II chapter (noted on the card). This card only locates, it does not instruct.

Card 1 · Master the field (method in Chapter 4)

This step's goal is to turn "can't read it all" into "can ask it something," and know who claims what, what they are arguing about, and which argument my question lands in.

What I need to produce: ______ (e.g., a controversy map)
Who is downstream (who takes it as input): ______
Drafter (me / AI): __ Review mode (line by line / spot check / criteria): _ Decider: Who answers when it's wrong: ___
The criterion for this step being "done": ______ (e.g., I can predict what new evidence would change my map)
The signal that sends me back to this step: ______ (e.g., at delivery there is a paragraph I cannot write clearly)

Card 2 · Questions and hypotheses (method in Chapter 5)

This step's goal is to grind a blur of curiosity into a question worth answering, one whose answer might embarrass me.

What I need to produce: ______ (e.g., one falsifiable hypothesis, with its scope)
Who is downstream: ______
Drafter: __ Review mode: _ Decider: Who answers when it's wrong: ___
The criterion for this step being "done": ______ (e.g., I can say which observation would kill this hypothesis)
The signal that sends me back to this step: ______ (e.g., the red team points out "the question itself is asked wrong")

Card 3 · Test plan (method in Chapter 6)

This step's goal is to lock in the plan, the method, and the criteria before running, above all "what counts as losing."

What I need to produce: ______ (e.g., a locked-in criteria list + baseline + definitions)
Who is downstream: ______
Drafter: __ Review mode: _ Decider: Who answers when it's wrong: ___
The criterion for this step being "done": ______ (e.g., the criteria were signed off before anything ran, and are not changed afterward)
The signal that sends me back to this step: ______ (e.g., during interpretation I find the criteria did not block some illusion)

Card 4 · Execution (method in Chapter 7)

This step's goal is to turn the design into data (code, pipelines, computational experiments).

What I need to produce: ______
Who is downstream: ______
Drafter: __ Review mode: _ Decider: Who answers when it's wrong: ___
The criterion for this step being "done": ______ (e.g., tests all green + results reproducible from raw data in one command)
The signal that sends me back to this step: ______ (e.g., during interpretation a "finding" turns out to be a bug)

Card 5 · Read and catch errors (method in Chapter 8)

This step's goal is to sort the results into findings, noise, and bugs. The most expensive error looks like the most exciting finding.

What I need to produce: ______ (e.g., a table of "each conclusion + its possible sources of illusion")
Who is downstream: ______
Drafter: __ Review mode: _ Decider: Who answers when it's wrong: ___
The criterion for this step being "done": ______ (e.g., every main conclusion has been checked against the standing list of illusions)
The signal that sends me back to this step: ______ (e.g., the red team finds an alternative explanation I did not check)

Card 6 · Deliver (method in Chapter 9)

This step's goal is to turn "I know" into "others can trust," with the right vehicle and every number traceable.

What I need to produce: ______ (vehicle: paper / memo / report / decision document)
Who is downstream (who the readers are, what decision they make with it): ______
Drafter: __ Review mode: _ Decider: Who answers when it's wrong: ___
The criterion for this step being "done": ______ (e.g., any number challenged with "where did this come from" gets its source within three minutes)
The signal that sends me back to this step: ______ (e.g., I cannot answer the reader's first question)

Card 7 · Red team (method in Chapter 10)

This step's goal is to let the harshest criticism happen at home before release.

What I need to produce: ______ (e.g., an attack list + a disposition record for each item)
Who is downstream: ______
Drafter: __ Review mode: _ Decider: Who answers when it's wrong: ___
The criterion for this step being "done": ______ (e.g., the strongest objection is written into the deliverable, not deleted)
Which step it sends me back to, and the signal: ______ (the red team's job is to send you back, write down the step you are most likely to be sent back to)

Self-check shared by all seven cards

[ ] Is Card 3 (test plan) empty? The step most often skipped entirely. When criteria are added after the fact, the conclusion always "happens to" support the plan finished first.
[ ] Is Card 7 (red team) empty? The step most often omitted. Without an internal red team, your first red team is the real world.
[ ] Was every card's "done" criterion locked in before starting? Criteria written afterward do not count.
[ ] Is there a card where all four criterion questions (draft / review / decide / answer for it) say "AI"? Downgrade it against the ladder self-rating sheet.
[ ] Read the "downstream" line across all seven cards in a row. Where the chain breaks is your project's real bottleneck right now.

Sheet 3 · Five-premise comparison table (Chapter 3, section 3.6)

How to use. Judge one real project you have in hand, premise by premise. Three minutes to fill. How many broke matters less than knowing where. For every broken premise, the method in each later chapter gets converted once by the rightmost column. When done, pin it next to the ladder self-rating sheet.

#	Premise	On my project	If broken, what fails first	My compensation
1	Errors can be found cheaply (the data is still there, it can be recomputed)	Holds / broken	The step 5 interrogation process (Chapter 8)	Move the interrogation forward to step 3. Locking in the criteria is my only chance at interrogation
2	The evidence is machine-readable	Holds / broken	Dispatching mechanical checks (Chapter 12, lesson two)	Cleanup cost goes into the verification budget, and this bill is usually larger than the check itself
3	The verifier is the producer	Holds / broken	L2's independent re-derivation (Chapter 12)	L0/L1 as written, no rerun capability needed; the economics of the spot-check rate are still exploring. Half a rule available, use the four attack surfaces of Chapter 10 as an acceptance checklist
4	The criteria can be written before the run	Holds / broken	Step 3, and with it the whole chain	Lock in a proxy criterion + write "the gap between the proxy and the real target" as a formal limitation
5	There is a next round	Holds / broken	Not one step, the conclusion the whole process produces	The downgrade ladder in section 12.6 of Chapter 12; the three floors of the "no next round" tier

Three things to do once it is filled in.

If you broke two or more, pin this table at the front of the book. Every Part II chapter you read needs one extra conversion, and the chapter where you forget the conversion will look usable and not actually hold.
Readers who broke premise 3, pay special attention. That cell is marked "still exploring" in this book, not "unimportant." The acceptance economics of second-person sign-off (how to set the spot-check rate, what to sample, how to escalate when a check fails) is the most visible gap in this workflow. Do not read "the book didn't write it" as "it can be skipped."
For every broken premise, write one line in your project README or plan document. Same reason as the downgrade record in Chapter 12. A premise that breaks without being written down leaves your deliverable looking exactly like one where every premise holds.

Self-check:

[ ] All five "holds"? Look again at premise 3 and premise 5. These two are the easiest to judge optimistically. At project start you always feel there will be a next round, and on signing day you find you never reran anything.
[ ] Is the compensation column copied from the book? Translate it into concrete actions in your project, otherwise it is only a sentence you once read.
[ ] If you judged premise 4 as "holds," check once more. Is the criterion you wrote a real criterion, or a proxy you wrote down before thinking it through? A proxy is nothing to be ashamed of. Not admitting it is a proxy is.