moca-experiments

Reproducible experiments and benchmark drivers for moca, split out so people deploying moca are not exposed to research clutter in the main repo. Each experiment has a question, a sealed prediction where it applies, a driver, a result and the steps to repeat it.

Experiments

ID	Question	Status	Start here
E1, E3, E4, E6, E7	Knative economics, mobility, recovery, saturation, converge-contention	archived (Kind/OCP cluster, deprecated deployment)	experiments/e01-e07-knative/EXPERIMENTS.md
E2, E5	In-process experiment runners; SWE-bench sandbox sharing	archived	experiments/e02-e05-harness/README.md, experiments/e02-e05-harness/RESULTS.md
E10–E13	microVM sandbox tier: lifecycle, density, vsock egress, restore capacity	done: measured against a pre-set bar, then stopped	experiments/e10-e13-microvm/RESULTS-SUMMARY.md
E14	Stock coding agents with every command in a fresh microVM, 40-run SWE-bench comparison	done: no measurable difference at this sample size	experiments/e14-sbx/RESULTS-SUMMARY.md
E15	Sessions per host: SBX per-Exec microVMs vs resident microVMs, trace replay	done: resident holds as many or more on every host and point	experiments/e15-session-density/results/metal/RESULTS-METAL.md, spec experiments/e15-session-density/SPEC.md
E15b	Per-turn command bundling, SBX vs resident	spec approved, not built	experiments/e15b-bundled/SPEC.md
E16	Does SBX beat resident at duty cycle ≤ 2%?	done: tie at the sealed point, hypothesis falsified	experiments/e16-low-duty/RESULTS.md
E17	Lease mode: does a warm guest with an idle timeout pay?	E17a done: level 1 misses its import/grep bar (short commands gain ~90%); level 2 being built	experiments/e17-lease/results/E17A-REPORT.md, spec experiments/e17-lease/SPEC.md
E18	Crash recovery latency on the Compose trial	done	experiments/e18-crash-recovery/RESULTS.md

Analysis: analysis/tracelab/ re-derives the E15 and E16 tables and figures from the published raw data, and holds the write-up (REPORT.md) and the summary deck. Its fetch script verifies and unpacks the data-v1 and data-v2 archives.

Reproducing

docs/REPRODUCING.md says what each experiment needs: cluster-free tests that run on a laptop, a KVM host for E10–E14, or a bare-metal host for E15–E17. It also covers fetching the raw data and the main-repo checkout the drivers expect.

• The summaries, sealed predictions and analysis code are all in this repo.
• The raw per-run traces for E15 and E16 (about 1.4 GB uncompressed) are release assets, not git content. data/fetch-data.sh downloads and verifies them; see data/MANIFEST.md.
• How an experiment is laid out and what each file is for: docs/CONVENTIONS.md.

Layout

README.md             this file: index of experiments
docs/                 REPRODUCING.md, CONVENTIONS.md, MOVED.md
data/                 MANIFEST.md, fetch-data.sh (raw traces live in release assets)
analysis/tracelab/    E15/E16 analysis: report, deck, tables, fetch script
experiments/          one directory per experiment, `<id>-<slug>`:
                        e01-e07-knative/  e02-e05-harness/  e10-e13-microvm/  e14-sbx/
                        e15-session-density/  e15b-bundled/  e16-low-duty/
                        e17-lease/  e18-crash-recovery/
scripts/              repo tooling (commit-msg hook)

Relationship to the main repo

This repo holds research and benchmark code only. Production code, deployment automation and design documents live in rossoctl/moca. What did not move and why is in docs/REPRODUCING.md.

History

Git history for every file here except experiments/e01-e07-knative/lib.sh (a point-in-time copy) was preserved via git filter-repo from the main repo at the split (2026-09-25).