Golden Realism

A golden that looks LLM-generated is sent back even when every figure recomputes. Drafting with an LLM is fine; shipping its first draft is not. This is an editing pass over a draft already numerically correct, run after the figures are asserted and beside reduce-house-fixes (which owns the container, H1, and the write-up prose, H2 and H3).

The hard line. Realism is presentation. No edit may change a graded figure, answer or contradict an ask, or open a second reading. A "realistic caveat" that makes a number ambiguous is a Gate E failure; when a realism move and a determinism gate collide, the gate wins.

Part 1. The tells

1. Uniform precision. Real work varies it: headcounts whole, money to the cent at invoice scale and the thousand at budget scale, rates to one decimal. The fastest tell, visible without reading.
2. Default styling. The library's first colour, 8.43-character columns, stock template, legend where the library put it.
3. The four-heading spine (Executive Summary, Key Findings, Recommendations, Next Steps). Take the shape from the genre: a determination has a standard and a finding, a reconciliation a bridge.
4. Sections of equal length. Real documents are lopsided toward the decision.
5. Everything in threes. Real lists are as long as the world is.
6. Bullets and prose inverted. The argument is prose; the enumerable (candidates, periods, exceptions) is the list or table.
7. No trace of use. Real products carry residue: a footnote on a disputed definition, a noted exclusion, a stated rounding convention. Take it from the pack's own fiction (Part 4).
8. The hedge register. "This analysis suggests", "it is important to note", comprehensive, robust, leverage, key insights, actionable.
9. Titles that restate the task. The finding is the title.
10. Generic file names. Not analysis_report.pdf; name it after the decision in the organisation's vocabulary.

Part 2. Per-format passes

Text (PDF, DOCX). The genre's identifying block (addressee, author, date, subject, reference); page numbers and a running header or footer where the genre has one; a finding as title; lopsided sections with the decision longest, in prose; enumerables in tables; a source line under every table in the fiction's vocabulary; at least one true footnote that opens no second reading; precision per quantity.

Data (XLSX). Named sheets, never Sheet1; frozen header (and key columns when wide); widths set to content; number formats per column (currency, %, thousands separators, dates), never a fourteen-decimal float; numbers right, text left; a total row computed from unrounded source that ties to every deliverable citing it; print area and repeat rows where a person would print; a notes sheet with definitions, as-of date and exclusions; autofilter only where one belongs.

Visual (PPTX). Slide titles as finding sentences; the chart is the slide, not bullets about it; slide numbers and a source strip; stock template replaced or colours and type set deliberately.

Charts (PNG, SVG, embedded). Load dataviz first. Units on axes; ticks formatted for reading, never scientific notation; a finding as title; the threshold line labelled with its value; the key point annotated; a deliberate palette with the recommended item distinguished; size and resolution for where it lands.

Data (CSV, TSV, JSON, Parquet). Column names in the source systems' vocabulary; a stated, stable sort; consistent precision per column; ISO dates; booleans as the consuming system writes them; a dictionary alongside where the fiction would have one.

Code (PY, IPYNB, SQL, R). Constants at the top, read from the pack, never hardcoded twice; assertions on control totals; formatted output with units (the deliverable is what it prints); comments only where a choice was made; a notebook's outputs match one top-to-bottom run and its markdown carries the argument.

Part 3. The container

reduce-house-fixes H1 owns it and its tool:

python3 .claude/skills/reduce-house-fixes/scripts/scrub_producer_metadata.py <dir>
python3 .claude/skills/reduce-house-fixes/scripts/scrub_producer_metadata.py <dir> --apply \
    --producer "<the organisation in the fiction>" --stamp <the in-fiction date>

Audit first: the producer and date are the fiction's and cannot be guessed. Never re-save through the writing library.

Part 4. Honest texture only

Every piece of texture must be true of the pack, in this order of preference: the pack's governing documents (a definition, an exclusion, a tolerance, cited in a footnote); the pack's own disputes (a belief someone held and how it was resolved); the analysis's own conventions (rounding, as-of date, population after exclusions, grain), each already pinned.

Never: a fabricated caveat, a fake revision history, invented initials, a data-quality note about a problem the pack lacks, or a value made scruffy (a wrong figure fails Gate A).

Part 5. The pass

1. Figures correct, asserted and frozen first.
2. Read the draft once at speed as a reviewer; note the tells before fixing anything.
3. Run each format's pass.
4. Add texture from Part 4 only, checking each addition against every ask.
5. Scrub the container.
6. Re-run the bundle verifier: every graded figure unchanged in every file at the same rounding.
7. Read once more cold; a draft still reading as a template usually has a structure problem (tell 3).

Checklist

• Figures asserted before and after; no edit moved a value, answered an ask or opened a reading
• Precision varies by quantity; no default styling survives
• Structure from the genre; lopsided sections; lists their natural length; argument in prose
• At least one true note from the pack; nothing fabricated; no hedge register
• Titles state findings; files named after the decision
• Format passes done (workbook affordances, chart pass with dataviz, code assertions and prints)
• Container scrubbed, audited first; bundle verifier green