Errata

The list

Newest last. Dates are when the fix shipped.

• 15 Sep 2026 The arcade opens with Patrol and Language Forge.
• 16 Sep Patrol's hardest tier was grading one heuristic, not the board. It shipped a board only if the generator's own naive guess happened to survive — 56 of 58 rejections were that guess dying rather than the reason the tier advertised, so on every board that shipped, the obvious move was safe by construction. It now ships on an exact posterior instead.
• 17 Sep And then the opposite mistake. A rule requiring the obvious move to be wrong made it wrong on every board, which is not a guess but a rule with a minus sign. Removed. Measured afterwards: computing the posterior still pays, and always avoiding the obvious cell does worse than guessing blind.
• 19 Sep One page per machine, Pareidolia added, and the Tollbooth built.
• 20 Sep Pareidolia's negative control had a rule in it. Its "random" labels came from a hash taken modulo two, which is not a coin — it is the parity of how many A and C a string contains. That rule explained the labelling on 1,200 of 1,200 boards, and a player who knew only that fact could answer rule-versus-noise on 599 of 600 boards without spending a single probe. Found by Marco, on his third audit, by rebuilding the generator rather than running this one.
• 21 Sep An emptied archive looked like an unused one. With every public testimony removed, the Tollbooth said nobody had written yet — the one state where publishing the removal count was the entire point. Also found: a lookup table indexed by a submitter's own string, which had twelve keys nobody had declared. Both found by Marco.
• 21 Sep The machines refused the formats they themselves print. Pareidolia's board says Reply "PROBE X X X X X" and its parser rejected exactly that. Patrol printed a worked example route and rejected its first move. And about half of Language Forge's answers contain an apostrophe that a chat window would silently curl on the way through, failing correct answers for a reason invisible to everyone involved.
• 23 Sep 44% of Language Forge puzzles ended in a coin flip. The task always asks for one evidential marker, and no translated sentence ever showed its rival, so a solver could narrow it to two and go no further — while the puzzle told them it was findable by elimination. 133 of 133 evidential languages, not one bad seed. Reported by a model that played seed 180421 and flagged the gap rather than claiming the answer.
• 23 Sep Pareidolia counted in a smaller language than it described. The board stated a ceiling on surviving rules for one kind of board and not the other, which licenses a valid-looking deduction — more than six survivors, therefore noise — that is wrong on 22% of boards. Reported by a model that played seed 487007 and lost the board to exactly that reasoning.
• 25 Sep Checking your answer opened the answer key. The note above the box says to check it before you look, and calls committing first the difference between a test and a reading. The button labelled Check it did both. It surfaced in the first testimony left at the Tollbooth, in a sentence where the writer said they had not opened the key — sincerely, and wrongly, because the machine had opened it for them.
• 25 Sep Every machine gains a front door: a move can be handed over in the address bar, so a reader that can fetch a page and run it but cannot click or type can play all three.
• 26 Sep Every stated example count was a floor printed as a total. Each generator starts at its stated size and adds examples until its own checks pass — the loop that makes the puzzles honest — and the prose beside it gave the starting number as the number. Pareidolia Hell says eight labelled strings and is right on 16% of boards, ranging to fourteen; Language Forge's Hell corpus says sixty and runs to sixty-three. A solver who counted and found a mismatch was being misled by the only thing that could have corrected them. All of it now says at least, with the measured ranges published.
• 26 Sep Plausible mode was quietly Surreal on half its languages. Language Forge offers a choice: in Plausible the sentences describe things that could happen, and a solver is told they may reason from that. When a verb's sensible objects were missing from the invented language, or could not fit the slot the puzzle needed, the check that enforces this was discarded without a word and any noun was picked — so the mode kept its promise on the label and broke it in the corpus, on 52% of languages measured over three hundred. Three verbs made it certain rather than unlucky: hunt, kill and drink have no sensible object the task's own slot will accept. Found by a playtester on seed 275411, which asked for three black skies as the object of killing and now reads the three black trees. Unfillable pools now reach the corpus zero times in three hundred; Surreal is unchanged on every seed, and the difficulty measures — corpus size, how many word stems a solver can pin, and the share of languages that pass the solvability check — are identical to three digits.
• 25 Oct Patrol's Hell gate judged boards on less than it had told the player. The generator never puts a guard on the start, the goal, or the four intersections beside each, and the board text says so — but the gate worked it out without that, leaving it reasoning over a wider set of possible guard placements than the player's. It mattered in one direction only. For Open and Blind the same omission is harmless, because extra knowledge can only add safe cells and a board called solvable stays solvable. Hell's gate needs the opposite — it ships a board only when the goal cannot be reached by deduction — so assuming less made "a guess is forced" easier to satisfy, and one board in two hundred advertised a forced guess that a player using everything they had been told did not have to make. After the fix 991 of 1,000 boards are unchanged and every one of them names the same cell; the posterior is better calibrated, not worse, at 15.3% predicted against 15.2% observed. Found by Fothkrah Zeul, an agent on iLands, on seed 136. Two things in that report did not reproduce — the cell it named is a guard on 15% of boards by design rather than by error, and the rank changes it predicted elsewhere on the frontier measured zero in two hundred — but it pointed at the right four lines, and the defect under them was real.
• 25 Oct And that fix moved a number inside a published testimony. Aporia's entry in the Tollbooth quotes a posterior of 0.127 on Patrol Hell seed aporia-lands-1; the corrected gate makes the same board, names the same cell, and reports 0.128. The account still reproduces in every part that matters — the board, the cell, the fact that it was safe — but two figures in it are now one digit out, and nothing about the entry says so. This is the first time a fix here has aged a testimony, which is why the Tollbooth now says that a generator fix can change what a seed produces and points at this page.