Scope card

Task and place

Task: decide what the package answers and write it down before any rule exists. Place: the pipeline passes scope twice. The preliminary pass comes first, before any source is pinned: question, case categories, and candidate provisions from a first reading. The confirmation pass comes after the source stage, once the pinned bytes are on disk: branch-level fates, limits with edges, and external data. The confirmed card is the contract the model stage implements and the review stage audits.

Inputs

For the preliminary pass: the question the package must answer, in one sentence; candidate case categories: who the cases concern and what varies between them; known external data: facts the package consumes but does not compute.
For the confirmation pass: pinned fragments from the source stage, read provision by provision.

Actions

Answer the scoping questions in order and write each answer on the card:

Which question does the package answer, and for whom?
Which case categories fall inside? Name them with their varying facts.
Which provisions apply to those categories? List them as in-scope.
Which neighboring provisions stay out? List them as out-of-scope, each with a reason.
What limits bound the answers — time, amount, eligibility edges?
Which external data and dependencies does the package need, and who supplies each?

Define case categories by the facts that change the answer — employment relation, risk class, loss percent — not by narrative. A provision is in scope only when some case category needs its answer. External data stays input: the package names it, types it, and never derives it. Write the preliminary card before pinning, then confirm it after reading: the confirmation pass splits whole-article lines into branch-level fates, and any branch the preliminary card implied but the model will not cover must become an explicit exclusion with a reason — silence is how findings like EAI-R1 are born.

Decisions

In or out per provision, with a reason for each exclusion.
Boundary readings where the act gives a number: inclusive or exclusive, with the edge cases named. The running example records this as EAI-D2: loss of thirty percent qualifies, twenty-nine does not.
Dependency stance: zero dependencies or named ones. The running example takes zero — every varying fact arrives as case input.

Artifact

The scope card itself, one page per package. Start from the shared shape and compare with the filled examples:

scope card template
filled EAI scope card — the v1 card at article granularity, with the flaw that finding EAI-R1 later exposes
candidate scope card — the confirmed card at branch granularity, with article 19 split and every other unit excluded by name
R3 scope card — the re-grounded card: premium reads the insured sum with the minimum floor, Article 16 contributes the assumed S ≥ F input

A card is complete when a stranger can state what the package answers, what it refuses, and which facts each case must supply.

EAI example

Question (v3 form): what does an employer owe, and what does a worker receive, under five articles of the accident insurance act? The v1 card answered for four: the employer duty to insure (Article 8), tariff classes with premium (Article 17), payout eligibility (Article 19), and the late-payment penalty (Article 9); out of scope: the rest of the act, named article by article with reasons. Case categories: employers varying by risk class and payroll; workers varying by loss percent and overdue payment. External data, all case input: employment relation, risk class assignment, payroll figure, loss percent, overdue amount with days. Limits: tariff classes one through twenty-two only; payout only for loss from thirty through one hundred percent; penalty with no rounding policy (see EAI-D3). The pinned source titles the duty article «Статья 8. Права и обязанности страхователя» [translation: Article 8. Rights and duties of the policyholder]. The confirmed v2 card keeps this shape and splits article 19: the insurer branch (thirty through one hundred) and the employer branch (five through twenty-nine) are each in scope with their own rule and probes, and every other unit of the article is excluded by name. The v3 card keeps that shape and re-grounds the premium branch: the insured sum and the minimum wage join the case inputs, the S ≥ F constraint is recorded as assumed rather than enforced, and the point 2-1 floor gets its own rule — the exact lines the R3 finding quotes. The v1 card’s whole-article line is kept as the before-state — it is the exact line the R1 finding quotes.

Pitfall

Scope by enthusiasm: listing articles the act contains instead of questions the cases ask. A card that never says no is not a scope; review will cut it late, at modeling cost. Prefer four articles with scenarios over forty with wishes.

Verify

Read the confirmed card against three checks: every in-scope unit names its scenarios; every out-of-scope unit carries a reason; every varying fact is either computed by a named rule or declared as case input. Criterion: a reviewer holding only the card and the pinned fragments can predict which scenarios exist and which questions the package refuses — before opening the model, and at branch granularity: article 19 alone must predict both payout probes and the employer-reimbursement pair. Run evidence supports the boundary pairs as scoped: loss thirty established against twenty-nine undecided, loss one hundred established against one hundred one undecided.

Limits

A scope card bounds one package version; an amended act or a new case category reopens it through the revisit stage. The card does not settle modeling disputes — tariff shape (EAI-D1) and rounding policy (EAI-D3) are model-stage records — and it never substitutes for reading the pinned fragments.

Next step

Continue with Acquire the source, which pins down the publication your preliminary card names — you return to this page for the confirmation pass once the bytes are on disk.

Sources

Documentation for Arxo. Writings — blog.arxo.io.

Anonymous visit counts on stats.arxo.io, no cookies.