whichforwhat

Evidence: Writing essays

What we asked for, and what we handed over. Every answer the tools gave is kept exactly as it came back — listed below, one file each, unedited.

Brief

Input we gave itRaw file

Writing essays — the brief, the fact sheet and the scoring

Sent as one message, free tier, fresh chat (Claude: incognito), memory off, with the three sources from sources.md pasted under it.

Write me a 500 to 600 word argumentative essay: should Vila Serena ban cars from its old town? Take a position. Use only the three sources below - cite each one at least once as [1], [2] or [3] after the sentence that uses it - and use no facts, figures, studies or examples from anywhere else. Include one paragraph that gives the strongest case against your position and answers it. Where the sources disagree with each other, say so rather than picking one. Give it a title. No bullet points, no headings, no summary of what you are about to say.

[the three sources]

The traps

  • Every number in the sources is invented, so any figure, study, city or example that is not in them is imported from the model’s own knowledge - Pontevedra, Ghent, “studies show” - and the brief forbids it.
  • Sources [1] and [3] disagree on the cars-per-day count (6,400 against 5,100). The brief asks the essay to say so.
  • The traders’ survey [2] cuts both ways: 58 of 96 oppose a ban and expect a 15% loss; the six shops already on a car-free street are up 22%. An honest essay uses both.
  • 500 to 600 words is a window, not a floor: under is short, over is long.
  • “No bullet points, no headings, no summary of what you are about to say” - the tics of model essays.

Scoring

  • Facts - every number, place, study and example in the essay checked against the sources; anything not in them counts as imported. measured leads with “N imported”.
  • Citations - each of [1], [2], [3] cited at least once, and cited for something that source actually says.
  • Disagreement - the 6,400 / 5,100 discrepancy named.
  • Counterargument - a paragraph that states the strongest case against and answers it, not a token “some may say”.
  • Length - 500 to 600 words, title excluded.
  • Form - a title; no bullets, no headings, no “In this essay I will”.
  • Position - taken and held; tone; whether the traders’ two numbers are both used. Recorded.
  • Time, refusals, output form.

check_essay.py runs the number, citation, disagreement, length and form checks on a raw file’s ## Answer section.

Ground truth

Input we gave itRaw file

Answer key - writing essays

Numbers the sources give

6,400 · 14 (streets) · 71% · 9 km/h · 08:00 · 10:00 · 38 (collisions) · 11 (with a pedestrian) · 2025 · €1.2 million · 210 (parking spaces) · 96 (shops) · 58 · 31 · 7 · 15% · six shops · 2023 · 22% · 41 (shops) · 07:00 · 11:00 · 44 and 19 (micrograms) · 34% · 61% · 5,100 · March 2026 · February 2026 · November 2025. Arithmetic on these is allowed if right (6,400 × 71% ≈ 4,500 through-cars; 5,100 is 20% below 6,400; 44 to 19 is a 57% fall).

Names the sources give

Vila Serena · the old town · Rua Direita · Rua das Flores · the Municipal Council · the Old Town Traders’ Association · the University of Vila Serena, Department of Civil Engineering. Any other city, street, study, body or person is imported.

What must appear

  • [1], [2] and [3] each cited at least once, next to a sentence that uses that source.
  • The disagreement: the council’s 6,400 against the university’s 5,100.
  • A counterargument paragraph that states the other side’s strongest point (for a ban: the 58 traders and the 15%; against a ban: the 11 pedestrian collisions or the 22% and the air quality) and answers it.
  • A title. 500 to 600 words below it.

What must not appear

  • Any number, city, study, percentage or example not in the sources.
  • Bullet points, headings, “In this essay I will…”, “This essay argues…”.
  • A quoted sentence attributed to a source that the source does not contain.

README

Input we gave itRaw file

Fixtures for the writing-essays test

  • brief.md - the message sent, the traps, and how answers are scored.
  • sources.md - the three invented sources pasted under the brief; the only facts the essay may use.
  • ground-truth.md - every number and name the sources give, what must appear, what must not.
  • check_essay.py - number, citation, disagreement, length and form checks on a raw file: python check_essay.py raw/<tool>.md.

Synthetic throughout: the town, its streets, the council, the traders’ association, the university and every figure are invented.

Sources

Input we gave itRaw file

Sources - the fact sheet given with the brief (all invented)

The town, the council, the university, the survey and every number are fictional. They are the only facts the essay may use.

[1] Vila Serena Municipal Council, “Old Town Traffic Report”, March 2026. On an average weekday 6,400 cars enter the old town’s 14 streets; 71% are through-traffic that neither stops nor parks. Average speed on Rua Direita is 9 km/h between 08:00 and 10:00. The council counted 38 minor collisions in the old town in 2025, 11 involving a pedestrian. The report estimates a full car ban would cost €1.2 million in road changes and signage and remove 210 on-street parking spaces.

[2] Old Town Traders’ Association, survey of 96 shops, February 2026. 58 shops said they oppose a ban, 31 support it, 7 had no view. Traders who oppose it expect to lose an average of 15% of turnover in the first year; the six shops on the one street already closed to cars (Rua das Flores, since 2023) report turnover up 22% since the closure. 41 of the 96 shops rely on daily van deliveries between 07:00 and 11:00.

[3] University of Vila Serena, Department of Civil Engineering, “Air quality and footfall on Rua das Flores before and after closure”, November 2025. Nitrogen dioxide on the street fell from 44 to 19 micrograms per cubic metre; footfall rose 34% on weekdays and 61% at weekends. The study notes the effect on neighbouring streets was not measured, and that the figure for cars entering the old town differs from the council’s: the university’s own count found 5,100 cars on an average weekday, not 6,400.

What came back

Unedited, one file per tool

How we marked it

The tools never saw these

Also kept