Documents instead of retellings

Fifteen of sixty-three

Issue 10 counts the records with no primary source behind them. After the showcase change of 26 September there were 63 of 848. The 6-in-1 asked for fifteen to be taken and searched. I took those where a document of the time surely exists, weighted towards the prehistory: Turing’s 1947 lecture, Strachey’s draughts, the Georgetown-IBM demonstration, Project MAC, DENDRAL, the Stanford Cart, Speak & Spell, Deep Blue, GeForce 256, AIBO, Roomba, Watson, Kiva’s warehouses, Pentagon Directive 3000.09 and China’s 2017 AI plan.

All fifteen were found: 29 new sources, each one read. They are the lecture typescript from the King’s College archive in Cambridge, Strachey’s letter to Turing of 15 May 1951, the Project MAC and NASA reports as scans, issues of Electronics from 1978, pages by IBM, NVIDIA, Sony and iRobot from the same year in Wayback, three texts of the directive, and the State Council notice on gov.cn. 48 records are now without a primary source.

What the documents said

Seven records were rewritten because the document said otherwise. The largest gap was in Directive 3000.09. The record quoted its requirement, its approving officials and its waiver as the 2012 text, and they were the 2023 wording. 2012 says “shall”, not “will”. The same three officials approve both times, and the waiver is for “urgent military operational need”, never from the legal review. The 2017 change was an amendment, not a reissue. The record itself said honestly that the directive had not been read, because the Pentagon’s server refused it. The copies in Wayback did not.

Turing did not call chess a fair test. In the typescript he says he would not think a win from hand-written tables significant, and offers chess as a field where the opponent’s moves give the machine contact with people. Strachey’s first Manchester draughts version existed by 15 May 1951, and the letter mentions no failed run on the ACE. The 6-in-1 then brought the Guinness page, which gives July 1952. David Link’s reconstruction bore it out: the coding sheets are dated June and July 1952, with a run on 10 July. Link also found what the record had not said: nothing in the code detects the end of a game or names a winner. The 6-in-1 then brought three more readings: Campbell-Kelly’s biographical note of 1985, a blog post on games of the 1950s, and the Bodleian catalogue of Strachey’s papers. Strachey’s own 1952 paper, reread, said what Link found: the machine played tolerably only until the end game, and the game it printed was stopped once the outcome was obvious. The failed ACE run of summer 1951 that Guinness and the blog, citing Copeland, describe was not restored: neither the letter, nor Link, nor Campbell-Kelly mentions it. The catalogue and Campbell-Kelly even disagree on the year Strachey met the Pilot ACE, 1950 or 1951. AIBO was only announced on 11 May; sales began on 1 June. There were 5,000, not 3,000: the Japanese share went in twenty minutes. GeForce 256 had four pixel pipelines, not thousands of parallel operations. The Speak & Spell chip was called TMC 0280 in 1978, the toy cost $50 and went on sale in July. Rodney Brooks was one of three founders of iRobot, and the Roomba moved in spirals and along walls, not by random walk.

In three more records the wording was fitted to the document: the Watson dates are air dates, the Stanford Cart went some 16 metres around a chair and cardboard shapes, and China aimed to be not “the world centre of AI” but its major centre of innovation. MIT’s documents of the time say Project MAC was organized in the spring of 1963; the day of 1 July comes only from the Multicians history, so the record is now dated by the year. For Deep Blue, Georgetown-IBM and Kiva the documents confirmed everything, and the texts did not change.

What is left

The archive catalogues of MIT, the Bodleian and the NLM sit behind bot checks; their documents were read in other copies. The 48 records without a primary source are mostly from 2025–2026 and the defence line: lawsuits, laws, court rulings. The route to them is already known: CourtListener, EDGAR and the institutions’ own sites.

A split that meets again is not a cycle

The audit of 26 September found a defect in validate:provenance that had broken nothing yet but would have. A source can be replaced by several, and those several can later merge into one: src-9001 goes to src-9002 and src-9003, both go to src-9004, and that goes to src-9005. The redirect schema accepts such a file. But the walk remembered every node it had ever visited, and when the second branch reached src-9004 it declared a cycle. Now a cycle is only a node that repeats on the same path, and a node already resolved is remembered and not walked a second time. The tests came first and failed on the old code: the audit’s converging split gives src-9005, and a diamond gives each final id exactly once. A real cycle, a cycle behind one branch and a self-redirect are refused as before. Records and entities cannot split, so their redirects resolve exactly as they did before the fix. All 3,510 ids passed the gate.