Research outcome · 2026-10-27 · rejected and rolled back
Faster shutdown.
Rejected candidate.
The V12 experiment reduced full-command latency by 87.40% on the fixed small-pipeline primary case. Three native-wall control intervals exceeded the frozen nonregression limit, so the candidate was rejected and rolled back.
V11 remains the accepted version. The measured V12 improvement and its failed release gate are both retained here. This is an experiment outcome, with no V12 improvement shipped.
- 87.40%
- lower paired full-command latency
small-pipeline primary · experimental - 2,400 / 2,400
- complete second-trial calls
successful with exact outputs - 3 failed bounds
- native-wall controls missed the rule
combined gate failed · V11 restored
A real command-time improvement.
The experiment made the worker heartbeat wait interruptible during shutdown. On the small primary pipeline, the full-command median fell from 1.029 to 0.113 seconds. The paired estimate was -87.40%, with adjusted interval [-89.73%, -84.53%] and exact sign-swap p = 1.907349e-6.
The bars are medians of 100 individual calls per version; the paired estimate comes from 20 blocks and is not the percentage difference between the two medians. Full-command time includes spawning, execution, shutdown, CSV export, reporting and exit.
The join and aggregate operators, reference behavior and timer boundaries stayed unchanged. This measures command responsiveness across the complete lifecycle, rather than isolated worker-reap time or faster query kernels.
Why it was rolled back.
The primary passed, as did all 11 other full-command controls. Nine of 12 native-wall controls passed. The remaining three had adjusted upper bounds above +5%, which failed the combined rule.
| Case | Paired change | Adjusted interval |
|---|---|---|
| Count by product + join · 1M | -0.61% | [-9.18%, +8.78%] |
| Inferred join keys · 1M | +1.52% | [-7.09%, +10.93%] |
| Count + join default control · 1M | -2.06% | [-12.88%, +10.11%] |
Each interval includes zero and extends beyond the allowed upper bound. The evidence did not establish nonregression under the frozen rule; it does not prove a slowdown greater than 5%. All three cases admitted all three workers on every call. Thresholds and observations stayed fixed, and the candidate was rejected.
The rejected source and binaries were archived before rollback. All 107 production files and two test files were restored to the accepted V11 source, and all three rebuilt binaries matched V11 byte for byte.
All 24 endpoints.
Twenty-three controls guarded the full-command primary: 11 other full-command cases and all 12 native-wall cases. Every gate had to pass; gains elsewhere could not compensate for a failed control.
| Workflow | Full command | Gate | Native wall | Gate |
|---|---|---|---|---|
| Count by product + join · 250k | -85.42%[-85.55%, -85.29%] | Pass | -0.12%[-0.67%, +0.43%] | Pass |
| Count by product + join · 1M | -64.81%[-67.56%, -61.82%] | Pass | -0.61%[-9.18%, +8.78%] | Fail |
| Filter + aggregate + join · 250k | -77.58%[-78.78%, -76.32%] | Pass | +0.13%[-0.56%, +0.82%] | Pass |
| Filter + aggregate + join · 1M | -66.00%[-67.59%, -64.34%] | Pass | +0.03%[-1.40%, +1.49%] | Pass |
| Small pipelineFull-command primary | -87.40%[-89.73%, -84.53%] | Pass | +0.27%[-0.53%, +1.08%] | Pass |
| Duplicate join keys · 64 | -88.37%[-90.45%, -85.85%] | Pass | +0.11%[-0.94%, +1.18%] | Pass |
| Empty right input · 64 | -89.23%[-90.42%, -87.90%] | Pass | +0.19%[-0.65%, +1.03%] | Pass |
| Empty left input · 64 | -87.36%[-90.24%, -83.64%] | Pass | +0.42%[-0.43%, +1.27%] | Pass |
| Pipeline + sort · 1M | -65.44%[-67.14%, -63.65%] | Pass | -0.02%[-1.65%, +1.63%] | Pass |
| Inferred join keys · 1M | -64.11%[-67.04%, -60.92%] | Pass | +1.52%[-7.09%, +10.93%] | Fail |
| Count + join default control · 1M | -64.39%[-67.13%, -61.43%] | Pass | -2.06%[-12.88%, +10.11%] | Fail |
| Sum + sort · 1M | -88.96%[-89.08%, -88.84%] | Pass | +0.23%[-0.50%, +0.96%] | Pass |
Native wall retains its earlier runLead boundary; worker reap and later export/reporting are outside it. No timer was shortened to obtain the full-command result.
The complete experiment history.
The first trial was paused at the user’s request after 655 completed calls, with invocation 656 interrupted. Its raw evidence remains archived, without a performance verdict. Before further measurement, a resumption addendum declared one complete fresh trial using the same source, binaries, seed, ordering, settings and gates.
The complete second trial contains all 2,400 calls. Its statistics use none of the interrupted trial’s observations. There was one administrative whole-trial restart; within the second trial there were no retries, exclusions or early success stopping.
Apple M5 Max, 18 cores, macOS. Three requested loopback workers, four threads each, on one shared Mac. The second trial recorded 920 contended calls and 34 calls with fewer than three workers admitted before completion. Every observation remains included.
Each of 12 cases ran in 20 paired blocks, with five matched repetitions per version and block. The five values became one geometric mean, leaving 20 inference units. All 24 intervals use paired log-latency t calculations with 19 degrees of freedom and Bonferroni alpha 0.05 / 24. The primary also required an exact two-sided sign-swap test over all 2²⁰ block assignments.
Every complete-trial call returned exact complete CSV and printed/native-JSON/journal checksum matches against preserved old outputs. File identity reads warmed the OS cache symmetrically. This is a shared-host internal lifecycle comparison; no Photon, Rust, Zig, Swift, Go or market comparison was run.
Source and evidence identifiers
- accepted V11 source commit
- 7d94e68c043c5111f52a624bda8e347b9a9da34b
- rejected V12 source commit
- 981b0ad393410fa21993049e9ad6dcea88fb67c7
- accepted V11 cast binary sha256
- 0f3dd00e29bd3a2ee2ce80aa57256e0bfc5496c78c8670a7c3de32d535ca7198
- rejected V12 cast binary sha256
- 6c5bbc35b50880e296fcc03ec33d1cc75414820a42c3c30c8ec4d9f893abdcff
- second trial protocol sha256
- acb7044f21c5439876aeb4373bc9d6ae46f1e06893a48baa99b909c5ef875b92
- second trial records sha256
- 8c8b187b280d50567e6d23dc7a6ac872dcd651db3a526a50e1bfd70deaf7ada3
- formal analysis sha256
- 9efd139f0df84e4a235a091cd3ec83f09c7a06f6c36d02455bf0b5b32e2e52ae
- independent statistics review sha256
- 61b4836eec7b45340cc6fb5c1f9f2c831dead255bbb437969929d26a2aa50303
- resumption addendum sha256
- ed5b318c54558d906383eb1a9cc74c7487c61deba68937337bfeab1db4d34640
- first trial records sha256
- 0e2a79dfefb1f6d0eabacba1a8528f5b453fa87ebe62f114b324ed805680d6e5
- raw backup commit
- 56a54e1939fbb983d483d43d598e46c2b5ff470a
- rollback verification sha256
- 9ee47dd3a8933a3f4bc0f630f5294c22d95c6700aa4c27649c78698059dc093f
- final rejection decision sha256
- c547fc5c0a8ba2efd3e1168c229b9a79f94eb9b45d89f641fa9e491f998a2bd9
- independent postrollback review sha256
- 5366d8c6a0b8129ef037151007cc5fe0d13f28b96f8aa2cdb2dbdc7bb7c243d3
The accepted V11 study and existing language tables are unchanged. No experimental V12 gain was applied to those measurements. Summary data and hashes are public; engine code, binaries and raw operational logs remain private.