Writing the Submission

The write-up is the answer key. The rubric is generated from it, the prompt and the goldens, never hand-edited, and needs 25 or more criteria; a wrong rubric is fixed in the prompt or the golden. Asks carry about 55 to 60 percent. Spec: guidelines/rubric.md.

Short is the standard

The in-task grader recomputes every figure in every block from the shipped files ("reproduce, do not assume", guidelines/determinism_judge_system_prompt.md). Every extra figure is one more claim that can fail, so a long write-up fails determinism even when the answer is right.

• Cut words and incidental figures, never owed figures. Every figure a component or an ask owes stays (each is a criterion). A figure that only shows working (a sub-count, a denominator, a rival's intermediate) goes.
• Reproducible, not low-level. Steps name the file and rule that produce each result; they do not walk arithmetic, restate the files' conventions or narrate cleaning. A convention needing a paragraph belongs pinned in a shipped file.

Not written: Deliverable Format Rationale, Justification, Golden Solution listing, Key Assumptions, Determinism block. Never copy one from an older task or from guidelines/.

Every figure is emitted, never typed. The generator writes <task>/generator/figures.json (the call, each component, each step result, each deliverable answer, as the write-up states them); blocks 1 to 4 are drafted from it, and this fails the build on any number absent from it:

python3 .claude/skills/determinism-check/figures_check.py taskNN

Typed figures and steps that read differently when followed literally were twelve of sixteen write-up findings in one window of judge reports, so each step uses the verifier's words and figure.

Order: build, determinism-check, this skill, golden-realism; never before the analysis is done.

The blocks, in order

#	Block	Length
0	## Tags (unnumbered)	2 to 3 lines
1	## 1. Final Recommendation	1 sentence, the deciding figure, 1 clause per rival
2	## 2. Critical Components	2 to 5 one-liners
3	## 3. Step-by-Step Solution	at most 8 one-liners
4	## 4. Deliverable Answers	grouped by file, answer lines only

Nothing else goes in the file.

0. Tags. The Axis 0 domain and the Axis 1 objective in the taxonomy's exact wording (guide-to-prompt); the objective is one of the eight labels or the task is rejected on the tag. No scenario recap.

Domain: Economics, prices and inflation (inflation decomposition). Analytical objective: Forecasting and predictive modelling.

1. Final Recommendation. The call in one sentence, no conditions, no "subject to". A hold is stated as a call ("We should fund none of the three programs this cycle"). Then the deciding figure and the margin, then one clause per rival: what it leads on and why it loses. No figure blocks 2 to 4 do not carry.

We should not acquire Data Corporation at the asking price. The price is 20x normalized EBITDA against the 12x mandate ceiling. Not the reported revenue case, which counts one-time and related-party sales. Not the trailing margin, which defers R&D.

2. Critical Components. 2 to 5 intermediate results every valid methodology must compute for the main call, one line each, figure in bold at its rounding everywhere else. Test: if a defensible method could skip it and still land the call, cut it.

1. Recurring revenue is USD 38M
2. Normalized EBITDA margin is 13%

3. Step-by-Step Solution. At most eight one-sentence steps, the last being the recommendation. Each is the method, the shipped file or rule, and the one result it produced.

2. Removed one-time and related-party sales per the revenue policy: recurring revenue is USD 38M.
3. Divided the asking price in the term sheet by normalized EBITDA: 20x, above the mandate's 12x ceiling.

• No sub-bullets; never pad to eight or cram two steps into one line.
• Carry only the result the next step uses or the call rests on; owed detail lives in block 4, once.
• The one exception: an ask whose device is silent gets a step naming the handling the golden uses and the files it joins, ending in the ask's figures, so the grader can verify what the solver is never told (supplemental-stumping Part 3). It counts toward the eight.
• Cite where a convention is pinned ("per the revenue policy"), never restate it. A convention a competent analyst could reverse is an unpinned fork, fixed in a shipped file.
• Name only files the pack ships, by their pack names, or the grader finds missing input.
• End in a figure or finding, never an intention ("Analyzed the cohort data" is not a step).

4. Deliverable Answers. Graded at the main call's bar: each reproduces and is uniquely forced. Grouped by file under ### <file.ext>, in the prompt's order of files and asks. Answer lines only, in the exact unit and rounding asked: no Basis line, no restated ask, no method.

• Concise is not summarised. An ask owing a row-set gets every row as its own point, every column value named, in the deliverable's order, one level of points under the numbered item.

1. Transfers, 42 rows in the file's order, whole units:
   • Store 118: send 340 to store 204, 1,120 on hand now, 780 after

• Never a markdown table (pipes in names break columns, wide tables wrap, odd tables are an LLM tell).
• A content requirement gets one line saying what the deliverable contains and where ("Bar chart of margin by category on page 2 of the memo, runner-up highlighted").
• A figure in two files has the same rounding in both and here; compute it once in code and carry it.

Golden deliverables

• One to three files, every one the prompt named, in its format. Four is a spec failure. No format family is assigned; prefer a visual where it makes the decision read at a glance. Together they answer every ask and the call.
• Business-realistic is a gate: once figures are frozen, run every golden through golden-realism.
• The recommendation is unmissable, never inferred from output.
• The code deliverable prints the critical components' figures, not only the answer.
• Prefer code-produced artifacts (a library chart, a script-written workbook, an emitted CSV); a chart in a text deliverable still comes from the script.
• Consistency is graded: every shared figure agrees, same rounding, in every file and in block 4.

Rejection causes

LLM tells (em dashes, odd tables, white space, extra headers, filler such as leverage, robust, seamless, comprehensive, hedging near the call); a figure that does not recompute; padding; a component another valid method skips; an unpinned convention; a conditional call.

Checklist

• Exactly the five blocks; none of the banned blocks
• Tags in taxonomy wording; block 1 one committed sentence, deciding figure, one clause per rival
• Block 2: 2 to 5 components every method computes; block 3: at most 8 one-line steps naming shipped files
• Block 4: every ask and every owed figure, answer lines only, exact unit and rounding, no table
• No incidental figure; shared figures agree everywhere to the same rounding
• One to three goldens, code prints the components, charts rendered by code, golden-realism run after figures froze
• figures_check.py exits 0; no em dashes