The generation-four memory control probe

The speed probe of 17 Sep 2026 (audit/evidence/2026-09-17/gen4-speed-probe/) measured generation four’s jobq at 46 MiB resident after the pairs phase against generation three’s 173 MiB, both with contracts on; with contracts off, 169 and 168. The auditor’s reading counts the 46 MiB as a memory win of the design, a CPU-for-memory trade; Fable’s reading says it is a side effect of change 4’s expensive postcondition (a whole-statement call that walks every finished job on every request, which is also where the region compacts, step 29b) and not of the design. The comparison of 18 Sep (audit/comparisons/2026-09-18-gen4-speed-probe.md, row AUD-COMP-GEN4-RSS-001) leaves the interpretation to Robert and names the test: a cheap contract at the same point. Neither reading is established until this probe runs. Pre-registered here before the run.

Predictions, written before the run

  • Fable’s hypothesis (H1, the compaction point): a cheap ensures on sweep that touches result.board once per request, without walking the finished jobs, keeps the region compacting at the same safe point, so the RSS after pairs stays near generation four’s 46 MiB while the rate returns to near generation three’s. Deleting the postcondition altogether returns the RSS to about 169 MiB (as contracts off does).
  • The auditor’s reading (H2, a design benefit): the low RSS follows from what change 4 keeps and frees, not from the walk; then the cheap ensures variant and the deleted-ensures variant both sit near 46 MiB with contracts on, and the rate is the only thing that changes.
  • A third outcome (H3, the walk itself frees): the low RSS needs the walk over finished jobs (its allocation pattern, not the compaction point); then the cheap variant sits near 169 MiB and only the original postcondition gives 46.

Fable predicts H1 with the numbers: cheap variant 40 to 60 MiB after pairs and 1,300 to 1,700 pairs a second at 32 workers; deleted variant 150 to 180 MiB. The row for Robert is written after the numbers, citing the outcome by its letter.

Orientation

The worktrees ../mo-lang-erosion3-mo and ../mo-lang-erosion4-mo (branches erosion3-mo, erosion4-mo), rebuilt with step 35’s mo; the probe scripts build.sh, run.sh, measure.py, surface.sh under audit/evidence/2026-09-17/gen4-speed-probe/; the diff board-sweep.diff there (the line to vary is the third ensures of sweep in examples/programs/jobq/board.mo, generation four’s line 380). The measurements are measure.py’s: creates a second, pairs a second at 32 workers, RSS after the creates and after the pairs, the restart. uptime and the top of ps before every run; nothing else running on the machine (Robert’s rule, and the orphan of 17 Sep).

Write scope

A new worktree ../mo-lang-erosion4-control from erosion4-mo with one commit per variant on a branch erosion4-control (never merged): variant cheap (the third ensures replaced by ensures looks_due?(result.board, now) or !looks_due?(result.board, now), a call on the board that does not walk finished jobs, in a whole statement), and variant none (the third ensures deleted). Nothing in main’s tree. The outputs under audit/evidence/2026-09-18/gen4-memory-control/ (the run logs as printed with the date and uptime at their head, the diffs of both variants, the scripts used).

Parts

A. Build the two variants with the same mo (toolchain/zig-out/bin/mo on main at the commit named in the log) and rebuild generations three and four with it too, so all four binaries come from one compiler. Confirm each variant’s mo check is green and mo test on board.mo passes (a variant whose tests fail is not a control).

B. Run measure.py (30,000 jobs, MO_CORES=1, as the speed probe did) on each of: gen 3, gen 4, cheap, none, each with contracts on and with MO_CONTRACTS=0: eight runs, each twice, the second run’s numbers on the page beside the first’s. Record the RSS after creates, the RSS after pairs, pairs a second, and restart.

C. One perf record of the cheap variant’s pairs phase, contracts on, and the /slowest surface for it, as the speed probe did, so the reader can see whether mo_disown_in is still on the profile.

Numbers

The eight-cell table (variant × contracts), two trials each, load average beside each; a one-paragraph reading of which of H1, H2, H3 the numbers support, written by the worker as a claim for the lead to read.

Done when

The table with sixteen rows on audit/evidence/2026-09-18/gen4-memory-control/RESULTS.md, every raw log beside it with the date and uptime at its head, the two variant diffs, the branch erosion4-control pushed, no jobq process left (pgrep -x jobq empty), and a report with the numbers and a numbered list “Decisions the brief did not cover”.

Result (18 Sep 2026, 4:45 AM ET; the worker ran 3:57 to 4:33 AM ET; read by Fable)

audit/evidence/2026-09-18/gen4-memory-control/ (RESULTS.md, forty logs, the two diffs, the scripts; branch erosion4-control, not merged). Eight cells, five trials each instead of two (the first sixteen showed one cell moving 56 to 129 MiB), one compiler (mo at 972c872). Medians after the pairs, contracts on / off: gen 3 186 / 233 MiB; gen 4 148 / 95; cheap 95 / 149; none 135 / 174. Pairs a second at 32 workers, contracts on: gen 3 1,893; gen 4 546; cheap 1,834; none 1,906.

None of H1, H2, H3 as written. The low state (33 to 60 MiB after the creates) is intermittent in every generation-four-family cell (gen 4 off 3 of 5, cheap on 3 of 5, cheap off 2 of 5, none on 1 of 5) and absent in generation three (0 of 14); it does not follow the walk (against H3), it does not follow the compaction point the walk sets (against H1: the deleted variant does not return a steady 169), and neither variant sits near 46 with contracts on (against H2 as stated). The 46 MiB of 17 Sep was one draw. The rate is unambiguous: the walk is generation four’s 3.5× loss, and a cheap or absent postcondition restores generation three’s rate. Fable’s rate prediction was low by a core; its RSS predictions held in 1 and 2 of 5 trials, which is to say they were wrong. The open question is the allocator’s resident set outside the region (the surface’s region_bytes 10 to 32 MB under resident_bytes 67 to 190), a runtime probe of its own. The rows of 18 Sep in the decision-log carry the reading for Robert and a measurement rule (a load near 3 is the run itself; timeout --foreground).