# Fresh contemporaneous five-run comparison

Generated2026-11-05T20:38:51.746874+00:00. Same clean GCP VM, same Codex OAuth account, native GPT-6 Luna/low, same32-row ledger task; all native skills/extensions enabled. Seven current clients plus two historical Octet comparators in ONE interleaved cohort. **These are fresh measured results, not previous campaigns or qualifications.**

| Client | Unsampled wall median (s) | Valid/5 | Peak tree RSS median (MiB) | Valid/5 |
|---|---:|---:|---:|---:|
|Octet0.10 source-built candidate|9.466|5/5|38.289|5/5|
|Pi1.5.0|10.954|5/5|144.766|5/5|
|Codex0.166.0|13.021|5/5|272.621|5/5|
|OpenCode2.2.5|10.473|5/5|425.996|5/5|
|fx0.0.16|13.508|5/5|28.762|5/5|
|OMP18.16.0|23.774|5/5|388.410|5/5|
|Prime0.10.2 (all Python skills)|12.654|5/5|125.469|5/5|
|Octet timing reconstruction|8.029|5/5|45.066|5/5|
|Octet exact memory RC|8.544|5/5|46.898|5/5|

Wall and RSS are SEPARATE passes, each with one retained warmup/client followed by five measured runs/client. RSS is common nominal50ms observed process-tree sum, including runtime/children; wait4 maxrss is NOT used for ranking. Primary medians condition on successful independently verified attempts. All attempted rows, warmups, failures, recovered tool errors and outliers remain in raw/ and data/attempts.csv. No retries replace failures.

Independent separate parser/reconstruction: 117/117 original verdicts agree (108 main attempts plus9 qualification02; qualification01 is retained setup diagnostic only). Fixture hashes/reconstructed summaries, unchanged input files, external checker replay, exact final text, native model/effort/session/cwd/checker tools, elapsed arithmetic and RSS sums/gaps checked. Independent verifier does NOT import live.py.

All-client complete-case measured rounds: timing[1, 2, 3, 4, 5]; RSS[1, 2, 3, 4, 5]. Complete-case medians and valid min/max are in data/summary.json.

## Retained failures and qualifications

- No failed measured/warmup attempts in these passes.
- Qualification01 retained two native credential-format failures (Codex/fx); qualification02 all9 valid after adapter corrections. Neither qualification is a measured replicate.

- Recovered native tool errors among all retained main attempts: 13. These attempts remain in the dataset.

## Provenance and limitations

- Octet0.10 is source-built candidate be5bcd39fa26ac024817dd82dfeb78fa99a894e6, SHA23a718110699df99ddbc568c92f2dd6aafda80b18b7a976e3b022d43805c0cdf, NOT an attested official final release. Latest public release visible at setup was0.9.6. Timing reconstruction overlay is incomplete; exact memory RC does not recreate historical DeepSeek study.
- Five schedule rows are exactly the first five of the frozen nine-rotation schedule: each arm occupies5 DIFFERENT positions, position sums23–27; directed adjacent pairs16 once/12 twice/44 absent. Not fully balanced. PROTOCOL-05-ERRATUM.md corrects a stale balance sentence without changing frozen data/order.
- Native prompts/tool schemas/permission policies/context limits/cache behavior differ. Fresh per-pass private client profiles, reused after warmup; OS/package caches stay warm. Unique per-attempt session binding. Provider routing/cache/network variation is uncontrolled.
- fx model/checker command status are native, but provider/effort are accepted explicit argv/config/env and cwd is process-boundary evidence, weaker than other native session exports. Prime has all9 bundled Python skill packages.
- Sampled RSS can miss short peaks; shared pages are double-counted. Observed reparented descendants are tracked by PID/start ticks, but a daemon detaching before first observation can escape. Unsampled timing has no descendant-scope proof. Sampling gaps are recorded per attempt.
- Protocol/fixtures/executables/runtime/package/locks frozen before measured launch; installed/frozen hashes checked before and after. Evidence and private profiles kept separate. Existing campaigns/seals preserved. No B/C/D/F or startup/streaming claims.

## Verification and privacy

- Original VM and independently checked identity-redacted local artifacts agree on all117 verdicts/results. The archive excludes13 Codex session creator_account_id fields; protected VM originals remain unchanged. No credential values or auth databases were transferred/archived. Redaction hashes/paths are in metadata/export-redactions.json.
- Actual measured RSS sample gaps: median 50.000ms, maximum 63.641ms (11500 intervals).
- Checks:6 streaming-verifier fault cases,10 prelaunch harness tests,5 separate-verifier tests; all passed. Previous6 top-level seals and all10,222 indexed files remain unchanged.

## What this does not measure

General coding quality, long-context/multifile work, real-world throughput, concurrency, other models/providers, provider/server RAM, exact unobserved memory peaks, cold-machine startup, statistical equivalence or universal speed superiority. Historical article figures are different workloads/versions and are not directly comparable.