Examination cover sheet · pencil 0.1.0

Pencil Comeback

$PENCIL

CandidateYour class
PaperAny subject
VersionsA to F
MaterialsA pencil

Instructions to candidates

  1. A school board voted 6-1 to ban student AI use in every grade, weeks after an AI pilot. Some teachers are going back to paper and pencil.
  2. Paper only works if it is done properly: shuffled versions, keys that match, marking that is fair, and an honest look at which questions worked.
  3. This page runs pencil 0.1.0, an open-source library that does exactly that. Write a test once; print as many versions as you need; mark the class; read the item analysis.
  4. Nothing you type leaves your browser.
Question 2

Write the test

[setting]Answer sheet

Plain text, once. A star marks the right option, an exclamation mark keeps an option in place, and the accepted answers to a short question go after an equals sign. The paper on the right is set as you type.

How to write a test 
Title: Fractions check
Class: Grade 5
Time: 25 minutes

# Part A: Choose one answer
1. What is 1/2 + 1/4?
   a) 3/4 *
   b) 2/6
   d) None of the above !
# Part B: Write the answer
11. Write three quarters as a decimal. [2 marks]
   = 0.75 | .75
Question 3

Versions and printing

[seating]Answer sheet

Neighbours get different papers. Questions are shuffled within each part, options within each question; the same seed always gives the same versions, so you can reprint them next term.

Seating plan: where each master question lands

Master question to version question, with the right letter

Number on that version, then the right letter. Short answers have no letter.

Question 4

Report card: $PENCIL

[the coin]Answer sheet

Pencil Comeback · end-of-term report

Live figures
SubjectMark
Price—
Market cap—
Volume, 24 h—
Liquidity—
Change, 24 h—

Teacher's comment: Figures will be entered when they arrive.

Supply
1,000,000,000
Tax
0%
Site
v1.0
The board is wiped clean. The chart is written up when it arrives.

$PENCIL on the board: 1-second candlesOpen the chart

Hall pass

Contract address

FZu9nxS72VUGqCzdezNbPNfbAduMpnoPfY3dUg8pump
Question 5

Mark the class

[marking]Answer sheet

One line per pupil: name, version, then the answers in the order of that pupil's own paper. Every answer is mapped back to the master, so a class that sat three versions is marked as one. The sample is a simulated class of 28 that sat the fractions test.

Marks

Each pupil's mark
PupilVersionMark%
Question 6

Item analysis

[review]Answer sheet

How each master question did across the class. p is the share who got it right. D is the top 27% minus the bottom 27%. r is the correlation between the question and the rest of the test. A question that stronger pupils get wrong more often than weaker ones is usually a question with something wrong in it.

    Question 7

    The answer matcher

    [fair marking]Answer sheet

    Short answers are matched by a policy written down before any answer was matched. It never guesses: anything it marks wrong goes on a list for a person to look over.

      The policy

      • Case, spaces and a final full stop do not matter.
      • Look-alike characters are folded: H₂O and H2O, x² and x^2, ½ and 1/2, the minus sign.
      • Digit groups may be split by commas or spaces: 2,500 and 2 500.
      • Numbers match by value: 0.75, .75 and 3/4; 1½ and 3/2.
      • A single-letter "x =" before a number is ignored.
      • Not done: number words, units written out, "the", extra words. List them as accepted answers when you want them.

      Held-out answers

      Where the matcher and the teacher disagreed
      AcceptedPupil wroteTeacherpencil
      Question 8

      The paper

      [100 marks]Answer sheet

      Paper · October 2026 · pencil 0.1.0, MIT

      Back to Paper, Done Properly: Seeded Versions, Fair Marking and Item Analysis for Printed Tests

      The pencil contributors. Answer all questions. Every figure is computed by the library on this page.

      1

      Summary

      A printed test has to do four things well that a screen did for the teacher: give neighbours different papers, keep every key in step with its paper, mark short answers fairly, and show afterwards which questions worked. pencil does all four from one plain-text test. Its item statistics equal an independent numpy implementation to within 2e-15. Its answer matcher marks 36 of 40 held-out answers as a teacher would, and wrongly accepts none.

      [10]
      2

      Background

      On 7 October 2026 the school board of Frederick County Public Schools, in Maryland, voted 6-1 to ban student use of AI from pre-kindergarten through grade 12, with exceptions for lessons about the technology and for students with disabilities [2][3]. Only weeks earlier the district had let middle and high school students use one AI assistant, Google's Gemini, with a teacher's permission; board members cited the lack of warning to parents, misuse elsewhere and the fact that teachers could not switch the tool off on student devices [2][3]. The president of the county teachers' association said some teachers were returning to paper-and-pencil assignments because they worried students would use AI to do the work [2]. The lone dissenting member warned that older students could be left less prepared for college [3].

      Not connected. pencil, this site and $PENCIL have no connection with Frederick County Public Schools, its board, its teachers' association, Google or any school, and none of them has seen this paper. pencil takes no side on the ban; it is a tool for teachers who use paper, for whatever reason.

      [15]
      3

      Versions

      Questions are shuffled within each part and options within each question, by a small generator seeded with the test's seed and the version's letter, so a version can be rebuilt exactly. Options marked to stay in place ("none of the above") keep their position. Every version keeps a map back to the master. Over 4,000 shuffles of a four-option question the right answer lands in each position between 880 and 1,120 times (the tests check it).

      [15]
      4

      Marking

      Choice answers are compared by letter on the pupil's own version, then mapped to the master option, so wrong answers can be counted per master option. Short answers follow a policy fixed in advance (Question 7 on this page). The held-out set of 40 answers, each with the verdict a teacher would give, was written before the matcher: pencil agrees on 36. The four it gets wrong are all rejections of answers a teacher would accept: "5 centimetres", "in 1945", "seven", "the Nile". A matcher that is too strict sends work to a teacher; one that is too generous gives marks away, so the policy errs toward strict and lists every rejected answer for review.

      [20]
      5

      Item analysis

      For each master question: difficulty p (the share of marks earned), the upper-lower index D (top 27% of pupils by total minus bottom 27%), and the corrected point-biserial r (the question against the total of all other questions). For the test: mean, standard deviation, Cronbach's alpha and the standard error of measurement. Flags: very easy (p above 0.9), very hard (below 0.25), barely separating (r below 0.15), working backwards (r below 0), a wrong option nobody chose, and a wrong option chosen more than the key.

      Table 1. The simulated class on the fractions test.
      PupilsMeanSDAlphaSEM

      On the simulated class, question 6 ("Which fraction is closest to 1/2?", where 2/5 and 3/5 are both 0.1 away) comes out with r = -0.20: the analysis finds the question with two right answers.

      [20]
      6

      Limitations

      The sample class is simulated from a simple ability model; it shows what the analysis can find, not how real pupils answer. Item statistics from one class of 28 vary a lot from class to class, so the flags are questions to look at, not verdicts. The matcher does not read number words, units written out or extra words, and because numbers match by value it cannot insist on "lowest terms". The same people wrote the policy, the held-out answers and the sample tests. pencil does not mark essays, and does not read scanned answer sheets.

      [10]
      7

      Use

      Write the test once; print two or more versions for neighbouring desks; keep the seed; mark from the pupils' sheets; look over the list of rejected short answers; then read the item analysis before using the test again.

      [10]
      R

      References

      1. pencil 0.1.0, the library on this page (MIT).
      2. R. Thubron, TechSpot, 9 October 2026: the school district's U-turn on AI. techspot.com
      3. S. Lindström, Tech Insider, 10 October 2026: the Frederick County ban in all grades and the Gemini pilot. tech-insider.org

      End of paper. Check your work.

      Question 9

      Stationery cupboard

      [source]Answer sheet

      pencil 0.1.0: one file, no dependencies, Node 18 and later and every modern browser. MIT licence. Every file of the repository is a pencil in the case, with its SHA-256.

      Downloads

      Quick start

      const P = require("./pencil.js");
// in a browser: <script src="pencil.js">

const bank = P.parseBank(testText);
const [A, B, C] = P.versions(bank, { count: 3, seed: "term-2" });
P.printPaper(A, bank);
P.grade(A, "b c a d | 0.75; 18").score;
P.analyze(bank, P.gradeClass(bank, [A, B, C], rows));
      node bin/pencil.js paper test.txt --version B
node bin/pencil.js analyze test.txt --class answers.csv
npm test

      The pencil case

        The syllabus: the API