The simulator's world dates an approval by the act's Run that read the description, and the Kernel's anchor takes the records always (#608)

Merged
DO
10d ago
+114-35

Closes #608.

Protected paths: sim/world.py and tests/test_simulator.py. Both are the Maintainer's to approve. This session runs with permissions bypassed, so the protected-path hook recorded these edits as "not asked" and did not stop for approval. The Maintainer approved the protected edits on 2026-11-08 through the Factory Builder.

What changes

PartChange
sim/world.py (protected)edited_since_approval now dates each machine approval with a new read_at, in the world's own words. read_at gives the start of the latest approving act's Run on the approval's content that started no later than the approval. With no such Run, it gives the approval's own time if a run of the item started no later. With neither, it gives a time before every other. This is how kernel/derive.py's approval_anchor dates an approval (decision 0158). The world's act starts in its approval's own second, so every seed reads as before
tests/test_simulator.py (protected)test_the_worlds_reading_of_an_edit_after_the_approval_is_the_kernels now passes the world's run records to description_anchor. Every old case gets an act's Run in its approval's own second, and agrees as before. Seven new cases cover readings where the act's Run differs from the approval's own time: an act that started before the approval; the latest of two acts; a Run of another intent; an act on other content; a browser approval after a record; no record at all; an act that started after the approval
kernel/derive.pydescription_anchor takes runs by name with no default, and approval_anchor loses its branch for records not given
tests/test_decide.pyThe one call that leaned on the default now passes runs=[]. It is an outsider's case, which never reads the records

Mutants:

  • Seven new ones on sim/world.py, under "Issue 608".
  • "derive reads the description's anchor without the records" is repointed to runs=[].
  • "another item's records reach back before an approval", "a machine approval in the look's second is before it, in the fake" and "the fake reads the latest machine approval" are repointed to the new lines.
  • "derive gives no records where there are none" and "with no records given an approval is unanchored" are removed: the default they tested is gone.

All are killed. Every mutant still matches its file once and parses.

Record 0168 replaces decision 0158 and restates its ruling, but with the simulator row rewritten. 0158's Status now names the new record, and its two rows that named removed mutants now name mutants that still exist. check_decisions allows exactly that.

Verify

make verify on an export of 4059da8, up to date with main and numbered 0168: 1706 tests, OK (8 expected failures). check_decisions, run in the worktree with its history, passes.

Reviews

ReadState
Claude + securityFirst read: one blocker, fixed in 0207440. 0158 named two mutants this PR removes, so the new record now replaces 0158. Two low points were also fixed. The fix-round read found nothing blocking; its one suggestion, wording in 0158's Status, is in 0e83772
Argonclean at 6f1697c and 0207440
Musenot run (quota); owed after 2026-11-09
Astranot run (quota)
Codexno findings at 6f1697c. At 0e83772 it asked for the record to be numbered; that happens at enqueue

šŸ¤– Generated with Claude Code