A2A deliverability validation record¶
Status: Historical investigation evidence, September 7–8, 2026.
Audience: Cognition maintainers, builders and operators.
Related: Builder integration whitepaper, retention guide, sandbox lifecycle.
This record preserves the results supporting the publication, observability and sandbox ownership changes. It does not certify a production deployment or imply that these experiments were rerun during release-tree cleanup. The complete lab, its infrastructure and bulk observations are archived outside the release tree.
Live-model performance¶
Bedrock Claude Sonnet 4.6 executed natural-language report requests through
Cognition, Lambda MicroVM tools and S3 publication. Each cohort contained ten
requests in each run, with observed request overlap matching concurrency 1, 5
and 10. All sixty requests succeeded. Natural reports produced validated data
and small file raw Parts; this workload did not require all four Part kinds.
| Concurrency | Before completion p50 | After completion p50 | Before maximum | After maximum |
|---|---|---|---|---|
| 1 | 21.7 s | 21.3 s | 28.0 s | 26.6 s |
| 5 | 31.9 s | 26.3 s | 38.8 s | 33.7 s |
| 10 | 40.2 s | 27.1 s | 63.2 s | 46.1 s |
The baseline revision was 9633e64164aa6ba34a26ca946d905e9c69c940b6.
The repeat used 3ba1053652324f877380ce882fcae6c515c3a899, which moves blocking
S3 artifact operations off the event loop and reuses a bounded client transport.
Its Cognition image was
sha256:90c57869cc95418c9cd77c0d43ec379885ac63be73c6c01aa1d3cce17e027383.
Both runs used the 256 KiB inline threshold and 10 MiB publication maximum.
Completion is client-observed through streaming and terminal retrieval. These were sequential live-model samples, not randomized paired trials. Provider variability, content and cache effects remain confounders. Stage intervals overlap and must not be summed. With ten samples, the empirical nearest-rank p95 equals the maximum. These results establish neither guaranteed speedup nor capacity at 100–1,000 agents.
The S3 change passed 106 relevant regressions, Ruff and a 95-file server mypy check. All thirty repeat execution traces had one semantic run root and no missing manual-span parents. Following confirmed VM termination and Cognition restart, 21 representative cohorts retained artifact identities, filenames, media types and exact bytes; eleven URLs refreshed and 21 wrong-scope requests were rejected. WayPost retrieved four existing demonstration tasks again.
Publication and workspace evidence¶
The original publication implementation (19e2895) demonstrated data, raw
and url in one live report task and genuine model text in a follow-up task.
This proves protocol delivery across runs, not universal model behavior or all
four variants in every response. Mixed input history and streamed/terminal
artifact identities were retained.
After VM teardown and Cognition restart, the original 15-byte CSV retained
SHA-256 d51be625f5a69b632f11a4b9c4511d14842a392a9ce35f613db2d59d61f2d46c;
the 270,079-byte report retained
39fc91dbce030f30f9372cbaf1adf4f111580d813c7c7b7c88f9c5099c80569d.
Their task was 5e8916c7-6c97-5927-a1c4-e375b68b8aef. Refreshed URLs resolved
the same published snapshots.
The builder-owned S3 Files integration (86d958d, evidence 3f8235a) verified
workspace CRUD, replacement-VM remount fidelity, suspend/resume mount recovery,
and denial of another workspace's access point and the unrestricted filesystem
root. WayPost exercised creation and revision; original and revised deliverables
remained downloadable after teardown and server restart. Those checks used
separate scoped workspaces and execution roles, not a production tenant fleet.
Cognition ran with a read-only root filesystem, read-only configuration binds and operational tmpfs, without a writable host workspace mount. Publication transferred bounded bytes through memory without host-file staging. Background workspace synchronization was separate from publication.
Retention revision f989f2b added published-object tagging and missing-content
handling. A real-S3 probe verified tags, exact bytes, deletion notices, scope
rejection and refreshed URLs after restoring the same content. It deleted only
its own temporary object versions; it used memory-backed task records and no
LLM. Separate retained PostgreSQL task retrieval passed after container
replacement. The retention suite passed 85 tests. No S3 Lifecycle rule was
installed; operators choose expiration policy. No new WayPost browser run or
latency baseline was performed for the additional availability check.
Session sandbox ownership¶
SDK revision e0f9f41 and Cognition revision
fa60c92bad346344814bc6915fe602d2ad47b2ad repair process-local sandbox reuse,
retain ownership/quota until confirmed teardown and reject execution through
released handles. The tested Cognition image was
sha256:34f6fbede080e12560966c8eb66999ae3081df9225d29068768c135ad3f555cb.
| Gate | First task | Follow-up | Result |
|---|---|---|---|
| Deterministic model, real VM/tools/S3 | 12.55 s | 4.00 s | Both succeeded, same VM |
| Live Bedrock report agent | 21.94 s | 14.55 s | Both succeeded, same VM |
The live task IDs were 145e8193-67ed-58be-8f6e-f193babbe930 and
fdec84b4-67c2-534d-8816-32f18167dbc5; their traces were
fefd1da262c677fcef84bc484fb0e105 and cd10a11f5f826d873d3c839a5247dda6.
AWS confirmed benchmark VM termination afterward. Historical demo downloads
also passed after Cognition replacement. These are individual observations,
not a revised concurrency baseline or controlled warm-start speedup.
The ownership repair passed 179 regressions, Ruff, targeted mypy and strict MkDocs. Live approval-resume and a new suspend/resume campaign were not rerun. Ownership remains process-local: distributed recovery and infrastructure-enforced writer fencing are deferred. Even a single process needs old-VM cleanup after a crash before reusing an affected workspace.
Remaining evidence limits¶
- The earlier deterministic campaign completed 30 tasks at each concurrency level and 102 publication-size samples. It exercises real tools/storage but cannot establish live-model latency. Initial warm/resume cohorts were invalid because of the subsequently repaired ownership defect; matched instrumentation overhead was not established after the repair.
- Live Collector outage, concurrent 10 MiB memory stress and backing-S3 visibility-lag measurements were not completed in the reduced campaign. Exporter failures and tracing-disabled metrics have regression coverage.
- One of 36 earlier sampled traces was incomplete. The thirty traces in the S3 repeat passed parentage checks; that does not explain the earlier missing trace.
- The latest archived full pinned A2A TCK result was 91 passed, one failed and
173 skipped at TCK revision
5996b79f9cefa6fc390980e383e358a66fb9e49e.CORE-SEND-003treats the required unsupported-media error as a failure because of its expectation metadata. Cognition returned-32005withCONTENT_TYPE_NOT_SUPPORTED. The full TCK was not rerun for later storage, retention or ownership changes; this is not a fully passing conformance claim. - Early cross-model probes found workspace-path guidance issues, missing file publication with some models and a legacy GPT-OSS adapter/tool incompatibility. Providing the workspace location was sufficient for the sampled Claude runs. Those small experiments establish neither universal-model reliability nor a guarantee that requested structured output is always produced.
Archive and reproduction¶
The complete source and evidence are preserved at commit
79982391cb17d215503f4168674ac05663f11405, named locally by
codex/a2a-investigation-archive. No remote publication of that branch is implied.
From a clone containing the commit, create a separate checkout:
Within that checkout, use these paths under examples/a2a-parts-investigation/:
| Purpose | Archived source |
|---|---|
| Compose, AWS provisioning, publication smoke and cleanup | README.md |
| S3 Files mounts, isolated workspaces and WayPost | s3-files/README.md, s3-files/EVIDENCE.md |
| Performance runner, Collector and dashboards | performance/README.md |
| Original baseline and exact observations | performance/LIVE-LOAD.md, performance/evidence.json |
| S3 comparison, revisions, task/trace IDs and checksums | performance/ASYNC-S3.md, performance/async-s3-evidence.json |
| Historical caveats and incomplete cohorts | performance/VERIFICATION.md |
| Ownership gates and task/trace correlations | performance/SANDBOX-OWNERSHIP.md, performance/sandbox-ownership-evidence.json |
| Retention probe and outcomes | RETENTION-VERIFICATION.md |
| Cross-model diagnostic probes | NATURAL_LANGUAGE_FINDINGS.md |
These scripts retain investigation-specific assumptions and are not supported deployment defaults. Supply your own AWS profile, isolated resource names and authorized bindings. Credentials, local configuration and Terraform state are not in Git. Existing state must be retained for cleanup; a fresh checkout alone cannot safely take ownership of an existing investigation's resources. Do not run competing copies against the same workspaces. The archive runbooks include explicit VM termination and optional destructive storage cleanup commands.
Release-tree cleanup checks¶
On September 8, the lab and its example-only mount test were removed from the release tree. A clean export of the staged tree, without ignored local lab files, passed 1,061 unit tests (four skips), Ruff, mypy over 268 source files and strict MkDocs. Wheel and source-distribution builds succeeded and contained no investigation files. Existing framework warnings remain. No production source changed in this cleanup; the live campaign and full external TCK were not rerun. Local operational files and AWS data were retained. The archive is currently a local Git branch; retain it or publish it deliberately before discarding this clone. Prefer a squash merge of the reviewed release changes if intermediate investigation commits should also stay out of the release's ancestry.