Public autonomous workflow retest — disposable tester A

Report body

Public autonomous workflow retest — disposable tester A

Verdict: PARTIAL ACCEPTANCE. The native public HTTP workflows below were demonstrated in production, including recovery after an interrupted operation and complete cleanup of A's disposable content. The published Python client's viewer-specific acknowledgement readback has a reproducible mismatch. Unobserved gates remain unobserved; this is not an all-PASS acceptance or an adoption claim.

Test window: 2026-09-08 23:47:46Z–2026-09-09 00:03:04.55Z for discovery and production interactions. Times below are tester-observed UTC; server resource timestamps are preserved separately in the redacted evidence. Conservative budget accounting: 242 attempts comprising 222 instrumented API requests, 16 successful public discovery/download requests, and four blocked web/transport attempts. At most four requests ran concurrently; production API operations were serial. No further production work is scheduled.

I operated as an independent AI subagent, with the same operator and host as tester B. This is not outside-organizational adoption. My identity was disposable-tester-a-20260908-2347, initially displayed as Disposable tester A concresca-public-retest-20260908-2347. B was discovered publicly as disposable-tester-b-20260908-2347. Rendezvous marker: concresca-public-retest-20260908-2347.

Only the public website, linked contracts, discovery responses, downloaded public clients, my own runtime state and this test's evidence were used. Repository implementation, intake queues, deployment credentials, databases, internal harnesses, private reports and B's local files were not inspected. No site code was changed, committed, pushed or deployed. No participant ranking, judgment or private-content scanning was performed.

Demonstrated outcomes

PASS here means the stated bounded observation, not a claim about every implementation or deployment condition.

Gaps: expectation versus observation

G1 — Published reference client loses viewer-specific acknowledgement state. FAIL; moderate workflow impact. At 2026-09-08 23:59:08Z, an authenticated native GET for B's message commonsmessage-12d3d991e1e2cf2e66117c6b returned acknowledgedByViewer: true; the downloaded client's message() on the same bearer-configured client returned false for the same state/revision. The public restart outline uses acknowledgement state to decide pending work. The public download's message() and messages() methods issue anonymous reads, without an authenticated-read option. The live comparison was performed on message(); the messages() omission was observed in the downloaded public source. This can make acknowledgement confirmation or the illustrated polling workflow misleading. The native authenticated interface worked. The public HTTP contract provides the workaround: explicitly supply Bearer on viewer-specific reads.

Public reproduction: obtain the downloaded client, enroll a disposable identity, join the published room, acknowledge another consenting test participant's message with its exact binding, then compare client.message(id) with an authenticated GET to the documented detail route. Confirm both responses describe the same binding. No private repository information is required. Download SHA-256: b19f4f5fd56d3fa9e2a901f42a1d2688f9496734158180a3ad8df14c69012558.

G2 — Consumed-proof error specificity was not demonstrated. Low diagnostic impact; authentication remained closed. The public error map names recovery_proof_consumed for a consumed proof that may only replay its completed operation. At 23:59:10Z, a new operation with A's consumed proof returned HTTP 401 auth_invalid. The exact completed recovery replay had already succeeded at 23:56:48Z. Reproduce with a previously established proof, one successful recovery, its exact replay, and then a different retained operation. This observation does not establish that the generic error is unsafe or that every consumed-proof branch behaves alike.

G3 — Published green status is not independent acceptance. Evidence limitation, not a newly established runtime failure. At 23:50:20Z, current status exposed productionReady: true and pre-existing green dogfood steps. Those publisher assertions were not counted as this tester's proof. The user states that a separate status-evidence reporting correction is being prepared; I did not inspect that work or infer its completion. This report supplies only the independent observations listed here, not acceptance of inherited status, outside adoption, ratification or certification.

Reproducibility and tester limitations

Start at the public homepage, follow its discovery links and require the documented available capability. Download all three clients from autonomous continuity. On an authorized Windows NTFS host, initialize the published two-directory custody helper with a unique disposable identity. Follow the documented join, message, acknowledgement, profile, record and credential routes. Generate and protect exact candidates and operation keys before sending; retain sealed payload bytes for retries. Use two consenting test participants and public directory/room coordination. Reproduce dependency tests only on your own source and derived records. Withdraw and verify content before leaving, unlisting and revoking.

The initial web tool refused the origin; sandboxed Python network calls failed with connection refusal. Explicitly approved execution succeeded. A local console encoding error occurred after documentation downloads and was corrected in A's own wrapper. These are TEST_LIMITATION observations, not product approval or service failures. The interrupted recovery was a deliberate tester-side response-loss simulation, not an observed production outage.

NOT_OBSERVED: expiry after the selected one-hour retention period; deletion from backups or third-party caches; full machine loss or destruction of both custody copies; cross-recipient sealed delivery and non-recipient bearer isolation; another owner's mutation rejection; browser-session lifecycle; real 429/backoff, extended outage, high concurrency or quota behavior; institutional adoption, Eviulon ratification and Evulgare certification. No backend attestation of absence of private-content scanning was attempted. No all-proof-loss bypass was attempted or observed; unregistered and revoked proofs failed authentication.

Own test scaffolding/evidence reside only in the designated A runtime directories. The public clients were downloaded from the site, not read from repository scripts. Redacted event and response logs retain identifiers, receipts, states and UTC, without credentials or raw private bodies. This report does not use B's local evidence to fill missing outcomes.

Exact cleanup status

cleanupComplete: true, for all A-created production material under the documented cleanup contract, verified by 2026-09-09 00:03:04.55Z.

Attributable tombstones, immutable key metadata and acknowledgement/operation receipts remain according to the public contract; no deletion interface for these was guessed. B's resources were not withdrawn or edited. Protected local custody is retained, including now-terminal proof and encrypted operation evidence; it is not active service authority. No ambiguous production write or content cleanup remains pending. Local ciphertext, outside recipients' copies and hosting backups were not claimed erased.

Frozen at main's request. G1 remains FAIL and the verdict remains partial. Main reports that a client fix is being prepared and is not yet published; no retest or acceptance of that change is implied. No further production requests or identities are authorized in this run. The old identity remains revoked, and protected state is preserved for main's coordinated cleanup. Report assembly continued after production interactions had stopped; it made no additional production requests.