Runtime performance lab publication record
Published product experiments, external contributions, evidence strength, and local-cost boundaries from the first CodeVetter performance campaign.
Runtime performance lab publication record
This page records what left the local performance laboratory on 2026-08-10. It is a publication ledger, not a second copy of the benchmark evidence. The canonical runtime contracts and commands remain in performance.md, while the active OpenSpec qualification records preserve the complete observations and limitations.
Evidence labels
- Confirmed means CodeVetter compared the same workload and metric under a compatible policy and the candidate cleared the materiality threshold.
- Directional means repeated measurements support the candidate, but a retained paired baseline is missing.
- Guardrail means the flow was already fast or the artifact primarily protects against future regression.
- Unverified means the source change is plausible and correctness-tested, but CodeVetter has not measured a defensible before/after effect.
- Rejected means a measured candidate regressed, stayed below policy, or failed to explain the captured workload. Rejected changes were reverted.
These labels separate observed runtime evidence from agent inference. An open pull request is not itself evidence that an optimization works.
Fleet product pull requests
Nine PRs below are merged. Three remain open because merge would cross a recorded safety or verification boundary. The campaign used dedicated branches without a deployment, migration, production database, hosted-product load test, or paid model call.
| Product | Pull request | Publication status | Runtime result |
|---|---|---|---|
| Web Playables | #3 | Open · Confirmed | Reusing the already-computed net revenue improved the 50,000-tick workload from 104.190 to 57.964 ms/op, or 44.37%; the shipping recommendation stayed withheld because the retained baseline had only three samples. Merge remains separate because the connected Cloudflare Pages check may publish main. |
| Anime List | #59 | Merged · Confirmed | Stable bounded ranking improved the 35,000-item workload from 20.575 to 10.386 ms/op, or 49.521%. |
| Significant Hobbies | #74 | Merged · Confirmed | Indexed category lookup improved the 35,000-phase workload from 9.116 to 1.810 ms/op, or 80.145%. |
| App Health | #34 | Merged · Confirmed, small | Reusing the Go response-writer wrapper reduced 30 to 29 allocs/op and 7,361 to 7,334 B/op; latency moved +9.253%, below the recorded rejection threshold. |
| Email Manager | #30 | Merged · Directional | Weekly digest generation measured 16.181 ms/op before the experiment and 7.266/7.534 ms/op afterward, but the original paired capsule was not retained. |
| Starboard | #74 | Open · Unverified | The 50,000-row recommendation flow measured 48.167 ms/op. All 201 CI tests pass, but the changed branches lowered branch coverage below the required 100%; merge is blocked pending focused tests. |
| LoopTV | #35 | Merged · Unverified | The 8,760-row Smart Mix flow measured 4.230 ms/op; the PR caches favorite membership and adds a permanent scale guard. |
| Reader | #36 | Merged · Unverified | RSS parsing measured 8.767 ms/op, but independent profiles disagreed on the source candidate. The PR is a narrow sanitization experiment with markup correctness coverage. |
| RolePatch | #47 | Merged · Unverified | Two profiles repeated calculateATSScore as the owned CPU candidate at 9.345 ms/op; no paired percentage is claimed. |
| Free AI | #51 | Merged · Guardrail | Model selection across all 79 models measured 0.016231 ms/op. CodeVetter refused source optimization and the PR adds only a regression guard. |
| Calorie | #12 | Merged · Confirmed synthetic stress | A one-pass exercise-history scan improved the 35,000-entry synthetic workload from 1.203 to 0.162 ms/op. The absolute saving was 1.041 ms at that stress input; no typical-user or customer-visible latency claim is made. |
| Reddit Insights | #3 | Open · Stacked, unverified | The bounded topic summarizer is based on collector PR #2; no paired speedup is claimed and no PR checks are configured. |
The strongest product results are Significant Hobbies, Anime List, and Web
Playables. App Health is mechanically real but too small to represent the
product’s end game. Merge status does not upgrade evidence strength: the
directional, unverified, guardrail, and synthetic-stress entries are not
independent claims of proven customer impact. The full inactive-project
disposition is retained in
artifacts/performance/old-local-projects-results-2026-08-09.md.
External open-source work
Marked
markedjs/marked#4048 was the first external upstream PR created by this campaign. It optimizes reference-link membership by checking the existing link table directly instead of rebuilding and linearly searching its keys for every queued inline source. The revised patch contains no duplicate lexer state.
Thirty alternating fresh-process pairs with per-process warmup measured:
| References | Baseline | Candidate | Paired median change (95% interval) |
|---|---|---|---|
| 100 | 0.805 ms/op | 0.489 ms/op | -39.2% (-40.3, -38.8) |
| 500 | 11.512 ms/op | 2.265 ms/op | -80.4% (-80.5, -80.2) |
| 2,000 | 164.115 ms/op | 10.095 ms/op | -93.8% (-93.9, -93.8) |
A 2,000-paragraph no-definition control moved +1.3% with a +0.3% to +2.3%
interval. Reintroducing Object.keys or using for...in restored that fast path
but re-enumerated the full link table and erased most of the target improvement,
so neither experiment was retained. Marked’s complete npm test passed: 190
unit tests, 1,779 specification tests, ESM/UMD/CJS and type builds, lint, and
generated-output checks. The PR is open and no longer a draft. Two maintainer
threads on the earlier implementation are outdated. Snyk passes; Vercel’s fork
preview requires upstream authorization and is not a code-test failure.
qs
ljharb/qs#592 adds a guarded fast path
for entirely flat query strings. A fresh ten-pair interleaved publication run
measured 41.706 to 23.183 ms/op at 40,000 parameters, or 44.413%, and returned
shipping_recommended: true with no limitations. The 1,000- and
10,000-parameter inputs improved 32.902% and 47.334%.
All 1,045 upstream tests passed with 100% statement and line coverage. The PR also documents the two unsafe broader candidates that the correctness suite rejected before the final fallback boundary was retained. It is ready for review; upstream maintainers still own the merge decision.
No upstream PR submitted
- go-chi/chi: the profiler found allocation pressure in the router path,
but inspection tied it to required
http.Request.WithContextand context semantics. No unsafe patch was attempted. - Picomatch: the candidate stayed below the materiality policy.
- Pixelmatch: the candidate regressed and was reverted.
- GJSON: the selected benchmark already reported zero allocations and no material bottleneck.
Negative results matter: they show that CodeVetter can withhold a patch when the evidence does not justify one.
What CodeVetter gained
The implementation and all qualification artifacts merged through CodeVetter PR #108. Follow-up PR #109 retains paired Vitest domain metrics, source-anchors campaign hotspot lines while preserving raw coordinates, and records the Calorie evidence boundary. Both post-merge CI runs passed the complete desktop, CLI, MCP, and browser pipeline.
The campaign expanded CodeVetter from a one-shot profiler into a bounded local optimization laboratory:
- Runtime failure capsules capture failing Node/Vitest executions with redaction, source maps, Git-diff relevance, and deterministic evidence versus inference.
- Performance capsules support Node tests, Node scripts, Vitest, and Go benchmarks with repeated metrics, dual-profile source attribution, scale curves, allocation evidence, and absolute-cost refusal.
- Local flow tools expose closed capture, inspect, explain, and verify operations over MCP without accepting arbitrary commands after startup.
- Optimization verification returns confirmed, rejected, inconclusive, or no-confidence and supports alternating paired checkouts.
- Qualification planning discovers safe repository-owned workloads, excludes cloud and remote operations, and emits one bounded next action.
- Durable supervision records resumable runs, budgets, exact revisions, evidence digests, cleanup state, and terminal reasons.
- Autonomous campaigns keep correctness gates and the evaluator immutable while an agent proposes one source experiment at a time.
- Verification receipts ingest project-runner evidence, compare regression state, and explain affected-test blast radius.
- Scaled challenge artifacts provide Node and Go temperature aggregation workloads inspired by the One Billion Row Challenge without claiming an official submission.
- Self-profiling found and fixed repeated source-offset scans in V8 function coverage. The representative 91-document replay improved 18.11%; the 98.808% number belongs only to the adversarial regression fixture.
- Contribution closeout now challenges retained patch complexity, binds local and optional T-Rex evidence to the pull-request head, reads current and outdated review threads, distinguishes failed checks from fork approval, and stops at upstream ownership without posting or assigning maintainer work.
The most important tooling improvements came from false or incomplete early results: startup-dominated tests now fail closed, Vitest names must identify one exact assertion, console metrics use repeated unprofiled medians, TypeScript locations anchor to original source, unsupported scale semantics stay unverified, and an inference must explain the captured workload before it is allowed to guide a patch.
One Billion Row Challenge boundary
The checked-in artifact is a local learning workload, not an official 1BRC result. The Node parser improved from 118.928 to 35.744 ms/op at 800,000 rows, or 69.945%, after CodeVetter selected the parser before source inspection. A later 8.863% micro-optimization was rejected and reverted.
The Go lane adds streaming and parallel parsing experiments plus a recorded 100-million-row result. It does not claim parity with the best public 1BRC implementations, one-billion-row completion, or hardware-independent speed. That distinction is preserved in the artifact README and result record.
Resource and safety boundary
- Product profiling was local and bounded; it did not hit production databases or hosted product endpoints.
- No Fleet product was deployed, migrated, released, or load-tested in the cloud.
- The Email Manager golden-fixture check fetched three small public npm package
artifacts through its existing
pnpm dlxcommand. - Existing package stores and locally generated build/profile artifacts are not evidence of equivalent network transfer or cloud spend.
- Raw owned profiles were deleted after normalization. Durable campaign data is
bounded under
.codevetter/performance-runs/or the campaign’s declared artifact directory.
Publication rule
Future campaign output should enter this ledger only when it has a durable branch, issue, or PR URL and an explicit evidence label. Merge status belongs to GitHub; benchmark truth belongs to the reproducible CodeVetter artifact.