Every Puzzle ProjectBlog

GAMES magazine: the hard case

After Judge, we chose the hardest pages we could find. GAMES, 1977 to 1982, taught us less about reading crosswords than about running a pipeline: what to record, what never to recompute, and who should be waiting at the end of it.

Alex DeJarnatt · Every Puzzle Project · 6 November 2026

Judge was a kind publication. One crossword to a page, an answer key a week later, clean type, and all of it in the public domain. When its 531 puzzles were in xd, the next publication was chosen for the opposite reasons. GAMES is an American puzzle magazine, first published in September 1977; from 1979 its puzzle pages, "Pencilwise", were edited by Will Shortz, and its constructors include William Lutwiniak, Merl Reagle, Mike Shenk, Henry Hook and Stanley Newman. If the tools could read GAMES, they could read almost anything. They mostly can.

Why GAMES is hard

It is a magazine about puzzles, laid out to delight a reader rather than a scanner.

The pipeline that finished Judge assumed one puzzle, one key, one clue list. For GAMES we rebuilt it as named stages, each writing its own record: harvest, the regions on each page, the parts read in each region, an assembly that links parts into puzzles, a "text step" in which Claude reads the printed lines near a puzzle and says which is the title and which the byline, then proof packets and triage. This is v2.

1977–79: steering by hand

archive.org has 34 issues of GAMES for 1977–82. The first three years became 81 puzzles on v2: 75 from a box review in which a person outlined every puzzle on every page, plus six the pipeline found on its own. Getting them to a reviewable state took fifteen runs, each a person diagnosing a kind of failure, a fix to a stage, and the batch re-run. On the last of them, 10 of 76 puzzles still needed work.

The text step was cheap: about 7,800 tokens a puzzle, 6% of a week's Claude Fable allowance for all 81. On 21 October 2026 the batch went up on the review site: 62 puzzles ready for review, 7 needing work, 3 set aside, 6 dropped as not crosswords.

1980–82: hands off, and measured

Fifteen runs is not a process. So the next three years were run differently: one orchestrating session, no code changes, no person steering, every cost and failure written down as it happened. The questions were what v2 could do on pages it had never been tuned on, and what a puzzle costs.

125puzzles assembled from the stored OCR, in 13 batches of 10
11 minfor the text step on all 125, twelve agents at once
$0.39a puzzle at API prices, all roles, before any review
13%of a week's Fable allowance for the text step
RoleAPI cost, 125 puzzlesA puzzle
Text step: reads the printed lines, names the headers, labels the restFable 5.1$35.27$0.28
Text step again, for 8 restarted puzzlesFable 5.1$1.73$0.01
Remedy agent: three passes over the 68 puzzles that needed workSonnet 5.5$4.12$0.03
OrchestrationOpus 5.5$9.30$0.07
Total$48.69$0.39

The run of 21 October 2026, at first-party API list prices. We ran it on a subscription's allowance, not the API; the dollars are the comparable figure. The text step's 12,100 billable tokens a puzzle count input, cache writes and output; it also read about 135,000 cached tokens a puzzle.

The first triage put 68 of the 125 in Needs work, 52 in Ready for review, and dropped 5 as not crosswords. A remedy agent then went over the 68, restarting a stage for one puzzle at a time. It fixed none outright. Seven it restarted from the parts stage, and six of those came through ready; the rest it could only diagnose and flag. The run ended at 62 Needs work, 58 Ready for review, 5 not crosswords.

Where it failed

Triage records a reason for every held puzzle, so the failures come as counts. Of the 68, 42 had a title or byline the text step had typed that no stage had read; 26 had a part the model's verdict called missing; 25 waited for a page never fetched, mostly cryptic keys in the Answer Drawer; 19 had clues unread; 17 had no key; 17 had a clue list that belonged to another puzzle. Behind the counts are a few kinds of failure, nearly all of them GAMES-shaped:

One finding mattered more than the rest. The text step, asked to label lines, sometimes wrote values no stage had read: titles and bylines for about 20 puzzles, flagged unsure, and for five unkeyed cryptic warm-ups, solved answers. That is transcription by another name. Triage caught every one, because every value records where it came from, and the rule is now fixed: no agent ever types puzzle content from a scan. It looks at a crop to diagnose which stage went wrong; it never produces a clue, an answer or a header the toolchain didn't deliver. A value typed by hand shows as typed by hand, and counts against the puzzle.

The gate nobody staffed

The 1977–79 batch had been on the review site for six days when we counted: eight decisions, all "needs work", from one helper. The ledger held four Ready, none from the maintainer. Our ship rule, one Ready and no Not ready, meant nothing could reach the corpus until someone read every puzzle against its scan, and nobody was going to read 700 of them. The pipeline had been tuned, run after run, to feed a gate that stood empty.

Two experiments on the 62 ready puzzles pointed at the answer. An AI proof review, Claude reading the proof sheets as images, passed 42 and held 20 with yes/no questions, at about 900,000 input tokens a puzzle. A check with no model at all, asking only whether the puzzle is complete and whether any stage recorded a doubt, passed 29 of the 62, and caught 17 of the 20 the AI had held.

So the human gate became a ship gate with three paths. A puzzle that passes every check ships with no model: every slot clued, the grid solved or its key confirmed not printed, numbering consistent, a second clue set present when the page promises one, no recorded doubt, nothing in a clue that a clean transcription never has. A puzzle the check holds goes to an AI proof reviewer that reads the puzzle as text, with crops only of the flagged items, chosen by code, never the proof sheets as pixels. It answers Ready or Not ready with one yes/no question per flag, recorded in the form a person's review takes, with the model's name on it. What it holds, or triage could not finish, goes to a person with a diagnosis and a link to the right row of the proof page. An incomplete puzzle never ships, whoever said Ready.

On 27 October 2026 the check passed 34 of 66 ready puzzles in 1977–79 and 39 of 58 in 1980–82; the reviewer, on Claude Opus at about 20,000 tokens a puzzle, passed a few more. Under the new rule, 1977–79 ships 40 of its 66 ready puzzles (31 by the check, 5 by the AI reviewer, 4 by a person), with 25 held for the maintainer; 1980–82 ships 42 of 58, with 13 held and 3 too crowded for any model to judge.

What is changing

The Judge post ended with four changes for the next publication. GAMES added these, which are the design of v3.

  1. Milestones with durable outputs. Each stage writes a record in a documented format: what it made, and what it was made from. Agents read the record. Nobody recomputes a stage to inspect it.
  2. Views never compute. A page draws only what a milestone recorded, and names the milestone. In October a viewer with its own key-finder, stricter than assembly's, showed a puzzle with no key while the pipeline had one, and nothing on the page could say why.
  3. Run once; re-run only from the invalidated stage. In the hands-off run, assembly ran three times over the same inputs, and a batch snapshot took 38 minutes fetching 1,031 files one by one. Earlier, every packet build re-assembled all 81 puzzles when only assembly's code had changed. Each is a bug, not a cost to live with.
  4. Diagnose a kind of failure, never fix a puzzle. The model is one 1979 puzzle traced back through the stages to its cause, a fix in the tool, and the batch re-run with no model at all; no agent read the other 80. A person's input is which stage went wrong, which restarts the pipeline from there.
  5. Every grid has a fate. A census lists every grid found on every page and what became of it: 226 grids in 1980–82, 119 of them puzzles, 23 keys with no puzzle. "Do we have them all" becomes a question the maintainer answers in one view.

One worry: GAMES may be the wrong test case. Most of what it broke is GAMES-shaped, and tuning assembly to every one of its layouts would be over-fitting to a magazine no newspaper resembles. So the fixes allowed between batches are generic: a heading links to the largest grid beneath it; a block whose numbering doesn't fit the grid isn't its clue list; a puzzle's kind is classified at triage, with a policy per kind. The fold pages, two-player lists and digit keys wait.

Where GAMES stands

As of early November 2026, 1977–79 is open on the review site, and 1980–82 has been through the run and the gate but is not yet published. Nothing from GAMES has reached the xd Crossword Corpus: the export path was built for the first pipeline and is being joined to v2. The plan is one import for the whole series, with one archive of every page it was read from.

The series is roughly 700 puzzles through 1989, about 480 of them not yet on v2, in 68 more issues on archive.org. They go in three hands-off batches, 1983–84, 1985–86 and 1987–89, each measured the way 1980–82 was, with the census review between harvest and text step and the ship gate at the end. After GAMES, the next publication should have one puzzle to a page, to measure the common case this magazine cannot.

Follow the series on its status page, or in the GAMES repo, which lists every puzzle the tools know of and its state.

The numbers come from the project's design notes of 17–27 October 2026, in the blitz repo.