A browser's approval is dated at its own time only where a record of the item began before it, by its jobs (#724, record 0204)
Closes #724. Part 4 of 4: part 1 merged as #751, part 2 as #759, part 3 as #760, and the bound on the pass's asking as #771.
It changes kernel/derive.py, and two protected files: sim/world.py and tests/test_simulator.py, pre-approved by the Maintainer 2026-11-08, ruled by the Factory Builder 2026-11-11. It touches none of a model Lens's own inputs.
What changes
The approval's anchor now reads when a Run began, which parts 1 to 3 put on every run record as began_at.
- The Kernel: a machine user's approval made in the browser is measured from its own time only where a record of the item began at or before it. Until now the anchor read a record's
started_at, which is when its run was made (trials/P-44.json). - What that closes: a Run made while the approving act ran waits in the item's lane and begins after the act ends (
trials/P-64.json). By its start it looked older than the approval, so an edit could pass unseen once the act's own Run was past retention. By its beginning it is younger, and the edit holds the item until a person approves. - A record that says no beginning does not reach back, and never raises: one not begun, never begun, with jobs not read, with a beginning that is no time, or one said to have begun before it was made.
- The act's Run is dated as before, by when it was made.
- The simulator: the world's records say when they began, and the world's reading of an edit is held to the Kernel's from the same records. The fake's rule
a_run_begins_with_its_jobis grounded on P-64, and on P-44 for a Run cancelled while it waited. - A record, 0204, amending five rows of 0168.
Said plainly
- It fails closed: every new case holds an item that passed before; none lets one through.
- What remains, as the record names: an act's Run a person deletes by hand; an approving Run on the same content made in the act's own seconds; a Run that began in the approval's own second; a job outside its item's lane.
facts/1gains no field, so the held-out cases are unchanged.- CI is red on one check until records 0200 to 0203 are on main: records are numbered with no gap. The number is the Factory Builder's of 2026-11-12. I merge main and push once 0203 lands.
- The size is near the budget: 57,669 characters of 60,000 as the Kernel counts a diff.
Checks
make checks: every check OK atc3722bfcbutcontext, which fails on the gap in record numbers alone. It holds main ate719181a.tests.test_decide,tests.test_adapters,tests.test_simulatoron 40 seeds: OK atcaf4762b(460 tests). The commits after it renumber the record and add two mutants. No other module was run at the desk: CI's verify is the answer.- This part's mutants, filtered, on 40 seeds: 79 of 79 killed at
c3722bfc. No mutant is stale. - gitleaks: no leaks.
- Size as the Kernel counts a diff: 57,669 characters, 456 lines.
Reviews
Carbon, Muse, Astra, the Claude review and the security pass, each of this part alone against main at c87fbb45. Every finding is fixed.
| Reviewer | Head | Unread parts | Blocking | Major | Minor | Info |
|---|---|---|---|---|---|---|
| agy | c87fbb4 | 4 | 0 | 2 | 4 | 0 |
| muse | c87fbb4 | none | 1 | 0 | 1 | 0 |
| astra | c87fbb4 | none | 0 | 0 | 0 | 0 |
| claude-review | c87fbb4 | none | 0 | 0 | 2 | 0 |
| claude-security | c87fbb4 | none | 0 | 0 | 0 | 0 |
| Finding | Severity | Criterion | Where | What | Disposition |
|---|---|---|---|---|---|
| agy#1 | major | none | docs/decisions/0195-a-browsers-approval-is-dated-at-its-own-time-only-where-a-record-of-the-item-began-before-it-by-its-jobs.md:25 | No mutant holds a re-run's first beginning against its last attempt's start in approval_anchor | fixed in caf4762b: the mutant "a re-run's first beginning before its last start does not reach back", killed by the re-run case the test already held, and named in the record's row |
| agy#2 | major | PR-3 | tests/test_simulator.py:649 | Grounding parity cases omit a Run whose job began_at precedes its started_at | fixed in caf4762b: the world's anchor takes the Kernel's guard, with the case of a Run made at 16 said to have begun at 15, and the mutants "the world takes a Run said to have begun before it was made as as old as the approval" and "the world takes a Run that began in the second it was made as not as old as the approval" |
| agy#3 | minor | none | tests/test_adapters.py:4351 | Contract check for a_run_begins_with_its_job does not test a Run cancelled while waiting | fixed in c3d06af0, as claude-review#2 |
| agy#4 | minor | PR-3 | tests/test_simulator.py:117 | Lane-start test stops holding began_at once a Run is no longer queued | fixed in caf4762b: the test holds that what a Run's record once says of its beginning it says at every later step, on every seed |
| agy#5 | minor | PR-3 | tools/mutants.py:18067 | Mutant for unread jobs in World.read_at leaves null began_at unheld | fixed in c3722bfc: the test already held a Run whose beginning is null apart from one whose record says nothing; the mutant "the world takes a Run that never began as as old as the approval by its start" now shows it |
| agy#6 | minor | PR-3 | tools/mutants.py:18089 | Mutant for a Run that waited in its lane in World.read_at is killed by unstarted Runs | fixed in c3722bfc: the test already held a Run made before the approval that began after it; the mutant "the world takes a Run that began after the approval as as old as it by its start" changes the comparison alone |
| muse#1 | blocking | none | docs/decisions/0195-a-browsers-approval-is-dated-at-its-own-time-only-where-a-record-of-the-item-began-before-it-by-its-jobs.md:1 | New decision record duplicates main's 0195 and make checks fails at HEAD | fixed in aa54cf20: the record is 0204, the number the Factory Builder gave on 2026-11-12. 0195 was provisional. The check on record numbers stays red until 0200 to 0203 are on main |
| muse#2 | minor | PR-3 | sim/world.py:1428 | World's anchor lacks the Kernel's begun-before-made guard and no test or mutant holds it | fixed in caf4762b, as agy#2 |
| claude-review#1 | minor | none | docs/decisions/0195-a-browsers-approval-is-dated-at-its-own-time-only-where-a-record-of-the-item-began-before-it-by-its-jobs.md:3 | The record's Status says three rows of 0168 are amended and the rest stands, but its own rows also change 0168's "The simulator" and "What it reads" | fixed in c3d06af0: the Status names the five rows of 0168 the record amends, and its row "What it reads" names began_at beside the fields 0168 names |
| claude-review#2 | minor | none | tests/test_adapters.py:4351 | The contract entry for a_run_begins_with_its_job does not read the rule's last clause, a Run cancelled while it waited never began | fixed in c3d06af0: the rule is grounded on P-44 beside P-64, and the contract entry reads P-44's job cancelled while it waited, which the port gives a null beginning |
Carbon's part 4 of 5, the first half of tools/mutants.py, timed out after 15 minutes: it is no read. Carbon is reading the final head again, and what it finds I answer in a comment on this pull request.
Astra read aa54cf20 too, which holds every fix but the two mutants of c3722bfc: no findings. No other reader of mine has read the fixes.
š¤ Generated with Claude Code