Renamed from Turing’s Witnesses on 11 November 2026, at the author’s suggestion, to depart from the author’s-name-plus-thing pattern of the other titles: the title now names what the player does, in the paper’s own verb. The file is test-it-like-turing.html.
The earlier memos are in previous/ (the first build’s full design memo in previous/memo.md; the five-round rebuild in previous/five-rounds/memo.md; the eleven-question build in previous/eleven/memo.md). This memo records what changed and why, and the build.
The author liked the eleven-question game and asked for two more modules: first sentences of English novels against a machine’s; and a module where the machine goes first and a person answers. Then: a fixed sequence, module 1, 2, 3; five pairs drawn from each, no option of length (“three times five is the sweet spot”); immediate feedback after each choice, no summaries between modules; the score the machine’s, not the player’s, put in words from the machine’s side; the polite convention once, as the coda.
Module 1 is symmetric: two writers, the same prompts, both blind, the machine as itself and within the person’s length. Module 2 pits the machine against people at their best, novelists, who wrote without any machine in mind; the machine gets the year, a premise and the length. Module 3 reverses the order and is Turing’s own game: the machine writes first, told to pass for a person keeping a diary, and a person who has read it answers. So: blind against an ordinary person; blind against a professional; and an imitating machine against a person who is trying. The difference between the three scores is the finding, and the verdict shows them side by side.
witness-b.md (Steffen Huck, one prompt at a time) and witness-a.md (a fresh session, the same rule, as itself, within the person’s length). The builder’s own first set, written before the person’s but several times longer, is witness-a-first.md, not used.module2-candidates.md lists the twenty-four candidates offered; the author chose the first thirteen). The machine’s openings from a fresh session given year, premise and approximate length (novels.md), checked against the originals and the full texts: no shared run of four words. Little Dorrit’s premise gave away Marseilles and the sun, which the memo notes and the machine’s sentence shows.everyday.md: aphorisms (“a sentence about everyday modern life”); confessions (“I have never…”, when allowed to say it had no moment); a novelist’s moments (well written, still a writer). The author then set the diary frame and told the machine to imitate, which is the paper’s game, and answered each line himself.Aim: Fifteen times, two sentences: say which one the person wrote. Yardstick during play: MACHINE TAKEN FOR THE PERSON: n OF m, in the score’s own units, since the score is the machine’s. Verdict: a five-step scale in words by the interrogator’s right answers out of fifteen, the band edges set by Turing’s bar (70 per cent of fifteen, rounded down to ten): 14–15 failed completely; 11–13 failed Turing’s test; 8–10 passed Turing’s test; 5–7 passed easily (no better than a coin); 0–4 the more human (the 2025 result). Under it the machine’s score, the player’s line, the three modules’ lines, who the witnesses were, Turing’s proposal of this experiment and his reply to Jefferson, the fifteen rows; then the polite convention and its readback.
Start screen (forecast; the three settings in a sentence; the aim; the bar), then fifteen questions straight through: a typed line when a module begins, saying who the witnesses are and what the machine was told (the player must be told that the machine imitates in module 3), then the question, X and Y, X IS THE PERSON or Y IS THE PERSON, RIGHT or WRONG with who was who (and the novel, in module 2), NEXT. No paper lines between questions, for pace. The chassis is unchanged: VT323 phosphor on black, the slow line, reverse-video buttons.
Toy bookkeeping around real sentences: the draw, the record, the reveal, the comparison with the bar, the scale. Demonstrates how one interrogator does on fifteen of these pairs in three settings against the 1950 bar, and what the argument says once the score is in. Assumes that these pairs stand for the general case (they do not); that a word limit, a premise and a diary frame are format rules; that fifteen answers from one interrogator measure anything beyond themselves (a coin gets about seven or eight); the scale and the four grounds. The paper establishes an argument and a forecast, nothing about 2026 and nothing about whether any machine thinks. Divergences: fixed questions answered in advance, no interaction; a machine told not to imitate in modules 1 and 2, which is harder than Turing’s game; the bar of ten and the scale are ours.
calibration.js, 105 checks, all passing (a quoting slip in the previous revision had left the file unrunnable while its stale result file still read 106 passed; the build now reads the check’s exit code): the paper’s numbers, the bar and the scale’s band edges; every sentence of all three modules byte for byte against the four files, the module-1 word limit, the module-2 years, authors and Gutenberg numbers and the length rule within four words, the module-3 topics; capitals and full stops throughout; the draw over 2000 seeds (modules in order, five distinct pairs each, all 34 drawn, sides balanced); 300 random games (three introductions, the reveal names the person and the novel, the scores add up); refusals; the five bands at 15, 13, 10, 7 and 4 right; the convention; no decimal.
| rule for judging | right of 5: module 1 / 2 / 3 | of 15 (bar 10) |
|---|---|---|
| always X | 2.48 / 2.48 / 2.46 | 7.43 |
| the longer is the person | 2.06 / 3.04 / 2.21 | 7.31 |
| the shorter is the person | 2.91 / 1.96 / 2.77 | 7.64 |
| the one that says “I” is the person | 2.74 / 3.04 / 2.98 | 8.76 |
| the one with two sentences is the person | 1.60 / 3.40 / 3.95 | 8.95 |
No rule beats the bar. Module 1 is balanced by its word limit; in module 2 the novelist tends to write the longer sentence; in module 3 the person wrote two short sentences where the machine wrote one, which is right about four of five there. These are tells about the writers; the About page names them and the sentences stay as written.
previous/eleven/.mechanism.js generated from the four witness files (module 1 from the previous mechanism’s verbatim strings, modules 2 and 3 parsed from novels.md and everyday.md), with the draw, the judgement, the scale, the summary and the convention. calibration.js and loop.js rewritten; the loop read as the stranger: a module line, the question, two sentences, a choice, RIGHT or WRONG with who was who, the count.template.html: the start screen, help, play flow and verdict rewritten for three modules; the status line in the machine’s units; the reveal only, no paper line, between questions. About rewritten with the provenance of all three modules and the sets not used.session-test.js, session-log.txt): two games (the shorter-is-the-person judge with the help opened mid-play; the longer-is-the-person judge ending as a solipsist), About opened, no page errors, no horizontal scroll.The author’s notes, all taken: the landing page simplified to the test, the three stages of five, and the machine’s score; no scenario repeated on the play screen, only a one-line stage opening (Turing’s questions, novels, diaries) with no set sizes; “human” for “person” on every button and reveal; the verdict the machine’s alone, with the player’s count mentioned once inside the forecast line and nowhere else; the stage-by-stage line and a best-and-worst line (“going by your answers, in 2026 machines are best at diaries and worst at novels”); the witness block, the fifteen-row table and the polite convention coda dropped (the coda’s question could not be answered by someone who had not known which was which; the About page keeps the reply to Jefferson as the reason the test is the only test). The scale is now expressed from the machine’s side, by how often it was taken for the human, with the pass band starting at Turing’s bar (five of fifteen).
The author’s second play found a discontinuity: “Did you enjoy writing those four lines on the bridge?” could come before, or without, the bridge question (so could the machine’s surprise about a nine-word bridge poem). The draw now plays those two only after the bridge: if one is drawn without the bridge, the next undrawn question (not the bridge, so the bridge keeps an even share) takes its place; if the bridge is drawn, it is moved ahead of them. Checked over 20000 draws. The author also saw a stage in which the human was Y five times running; the draw is a fair coin per round (the human is X in 49.5 to 50.4 per cent of rounds at every position), and a whole stage on one side happens in 6.3 per cent of stages, as a coin does. Both checks are now in the calibration file.