cooperative client key-memory results
all 360 registered observations evaluated. exact c25519 key bytes remained observable after close under all; gost post-close zeros do not prove erasure.
cmd/lab -set keymem now has a completed registered series on both cryptographic suites.
a fresh client uses production paths through three authenticated relays and verifies an exact
echo. the assay selects its three 32-byte cell keys and matches only their exact contiguous
bytes in parsed Object payloads of one voluntary go runtime export.
| suite | policy | active | closed-now | closed-1s |
|---|---|---|---|---|
| c25519 | none | 3 | 3 | 3 |
| c25519 | all | 2 | 2 | 2 |
| gost | none | 3 | 3 | 3 |
| gost | all | 1 | 0 | 0 |
these are occurrences per selected key, not a count of separate copies or libraries. every selected key in every process had the displayed count. each condition planned and evaluated 30 fresh processes; all 360 evaluated, with no exclusions. each hop uses 30 processes, not 90 independent observations. any-key presence was 30/30 except gost all at the two post-close points, 0/30. no p-values, confidence intervals or superiority claim are made.
active follows the verified echo with the client open. closed-now follows successful
Client.Close without an intentional wait; closed-1s is at least one monotonic second after
that return. each snapshot uses different keys and a fresh process, so the points do not
trace one key’s lifetime. natural gc and actual export delays belong to this schedule.
measurement used optimized go 1.27.3 on linux/amd64, seed 1, clean commit
eddddb226e7224140f8655b1566a55ce33f3ceda.
binary sha256: a715234e5347d3ade2689dd585cf7c5cc36c82c1e9ec26c2cd29ffedc7067dc7.
the aggregate rows and summaries were independently recomputed. the numbers and provenance
are in the readme,
and commands are in experiment block 4.
c25519 exact bytes remained observable under all after close. gost zeros mean only that the selected exact encoding was not observed in the evaluated Object payloads. transformed or derived gost states, stacks, globals, runtime/system regions and mmap payloads are not searched. protected primary mmap buffers are outside the view even while live; the runtime can silently omit data. none/all changes several measures together. instrumentation-induced compiler/runtime representations cannot be distinguished from library remnants. these counts prove neither erasure, an erasure deadline, hardening efficacy nor resistance to privileged acquisition.
kind was stopped, with no concurrent builds or source writes. whole-host cpu had 30 samples during the series: 3-33%, median 11.5%, mean 12.4%. they include assay work and cannot isolate unrelated load. no row was removed for load, and no performance or load-robustness conclusion is drawn. the passive tables and figures retain their original revision and results.
external acquisition, disk/swap/history, other secret classes and recorded-traffic decryption remain unmeasured beyond v0.4.0 in #192. the remaining full node-observation measurements in #117 are required for v0.4.0, in design and not implemented. the limits describe the boundaries.