Waiting for better models
The restart moved at the speed of the models. Twice it stalled for weeks, and twice a new Claude model got it moving again.
The restart did not move at a steady pace. It went in bursts, and the quiet stretches between them were not holidays. Twice the work reached a point the model of the day could not get past, and twice it started again when a new Claude model arrived. The prompt history shows it plainly.
Show the numbers
| Week of | Prompts |
|---|---|
| 16 Mar | 565 |
| 23 Mar | 1115 |
| 30 Mar | 739 |
| 6 Apr | 118 |
| 13 Apr | 38 |
| 11 May | 6 |
| 8 Jun | 335 |
| 15 Jun | 210 |
| 29 Jun | 9 |
| 6 Jul | 10 |
| 13 Jul | 553 |
| 20 Jul | 184 |
| 27 Jul | 11 |
| 3 Aug | 128 |
| 9 Nov | 258 |
| 16 Nov | 77 |
Spring: close, and stuck
The first push, in March, ran on Claude Opus 4.6 with the effort setting at its maximum. It went fast at first. In a fortnight every version was disassembled and decompiled, Rogue 3.4 was read through function by function, and a rebuild of Rogue 5.4's object file was under way. On 2 April that rebuild came within two small patches of matching.
Then it stopped converging. The last differences in 5.4 resisted everything, and the model kept concluding that what was left could not be fixed, or was close enough. The person running the work kept insisting it could be, and that "close enough" was not the standard. By the end of March it was stuck all the same. In mid-April the effort setting was turned down, turned back up, and then the work stopped. For seven weeks there was almost nothing.
June: a new model, and then the right one
Work started again on 9 June with a different approach: instead of reading Rogue 3.4 harder, build the 2.8BSD compiler and prove each function by compiling it. Claude Opus 4.8 appears in the commits the next day, and within two days the C library, curses and termcap libraries all matched the binary exactly.
On 12 June Claude Fable 5 came out, and the problems that had been set aside were handed to it the same day. The next week was the breakthrough of the whole project: the remaining differences in Rogue 3.4 were worked through one after another, with the effort setting raised again when a problem needed it.
About a week later Fable was taken away for a while, and the work waited for it. For two weeks at the end of June there are almost no prompts at all.
July: everything falls
When it came back, things fell quickly. Rogue 3.4 matched byte for byte on 3 July. In the next three weeks so did Rogue 3.6 on both machines, 4.22, 5.3 and 5.2, and the toolchains and simulators grew into the general-purpose tools they are now.
November: the ports
Then another quiet stretch, from September to November. The Linux ports had started, and the Rogue 5.2 port had been stuck for a while. Claude Fable 5.2 was first used on 11 November and Claude Opus 5.5 on 14 November, and in the week that followed all six Berkeley versions were ported, with more than five hundred separate bug fixes between them, and PC Rogue after them.
What still waits
Two spots in Rogue 5.4 are still patched by hand after compiling: the
exit from do_move() and six bytes of register choice in command().
The likely cause is a compiler build slightly different from the one
reconstructed. If the pattern holds, they are waiting for the next model.