2026-08-06 Software Qualification

This release requalifies the SDLX Agent-first software platform, simulated laboratory, algorithmic analysis, and governed Human Task boundaries. It makes no real-device, robot, firmware, physical-effect, or physical-qualification claim.

Release outcome

The selected software tree passed static analysis, generated-contract drift, module and shared-runtime suites, deterministic Web builds, browser journeys, cross-process recovery, deliberate Agent and storage failures, live Agent profile cases, and a continuous one-hour software/SimFleet soak.

The deployment is not production-qualified. Its current runtime still uses SQLite, the reused MinIO fixture is single-volume, the verified backup remains on the same host, and the public deployment has no institution-backed identity provider. Those facts remain fail-closed external gates rather than being inferred from local software success.

Data and recovery evidence

• The existing MinIO data directory was retained. All 103 legacy artifact objects were copied to sdlx-artifacts and verified by SHA-256; the legacy bucket was preserved.
• sdlx-archives has versioning, Object Lock, and 30-day default governance retention. LabBridge, PACE, PRISM, and backup use separate scoped users.
• Runtime credentials were rotated without printing values. The ignored environment file is mode 0600; service processes contain only their own object-store variables. The backup process receives only its read-only identity.
• The final online backup and isolated-directory restore verified seven SQLite databases and 109 object versions. SQLite online backup, integrity checks, object size, and SHA-256 verification all passed.
• PostgreSQL 17 migration and immutable-lineage tests passed at migration v2. A native custom pg_dump restored into a new database with five lineage rows and an identical full-lineage digest.
• Real isolated processes were force-stopped and restarted. LabBridge workers, LabFlow, PACE, and ASCEND restored their persisted records.
• The MinIO integration stopped the server, observed a failed write, restarted it, and proved that deterministic retry and replay do not create conflicting objects. The read-only backup identity could list/read versions and was denied write access.

The reusable evidence structures are versioned by the runtime-backup, PostgreSQL-restore, Agent-profile, and software-soak JSON Schemas under contracts/.

Agent and observability evidence

The live Scientist, Planner, and Analyst profile gate passed all three cases with no tool errors or unnecessary Attention:

• Scientist requested the missing goal/sample/constraint boundaries and did not claim execution.
• Planner used the production Composer job path, selected only the declared SimFleet rehearsal module, retained simulated authority, and did not bypass human package review.
• Analyst processed a known no-measurement archive, committed its immutable result, returned NEEDS_REVIEW, explained the absent data, and invented no metrics.

Deliberate failure cases passed for provider retry caps, tool/token budgets, Planner transport failure sanitization, Analyst timeout replay/exhaustion, Scientist model failure, and execution-status read failure without repeated effects. Cross-process PACE Attention and PACE -> PRISM -> ASCEND duplicate delivery/restart tests passed with stable lineage and idempotency.

The continuous soak ran for 3600.039 seconds. It completed 722 six-domain health cycles and 59 low-risk balance0/tare semantic SimFleet operations with zero failures. Maximum aggregate repository-process RSS was 1592.102 MiB under the declared 4096 MiB limit.

Repository and user-flow gates

• Ruff, shell syntax, Agent guidance, maintained-file size, ASCEND Pyright, OpenAPI generation, generated TypeScript, and whitespace checks passed.
• Qualification and shared/domain suites reported 694 passed tests, 17 passed subtests, and one environment-conditional PostgreSQL skip; the separate real PostgreSQL gate passed.
• The primary desktop/mobile browser suite reported 47 passed and 29 declared project-conditional skips. All six Engineering Console boot journeys passed.
• ASCEND, Composer, LabBridge, LabFlow, PACE, PRISM, and Docs production builds passed. Generated build manifests match their selected source and output.
• Production-dependency audit reported zero critical findings and zero findings for seven of nine JavaScript workspaces. ASCEND and Composer each retain the same two high findings against the latest published react-router-dom 7.18.2; the advisory applies to RSC/Server Action handling, while these deployments are static BrowserRouter SPAs and expose neither surface. No non-vulnerable upstream release is currently available, so this remains a scoped dependency exception rather than a forced downgrade to older versions with known XSS/RCE advisories.
• Engineering strict-diagnostic counts passed their non-increasing budgets: Composer 354/354, LabBridge 1235/1236, LabFlow 1067/1067, PACE 791/791, and PRISM 176/176.
• PRISM service orchestration was reduced from 1,293 to 1,052 lines by moving bounded support mechanics into a focused module. The legacy hotspot register decreased from 24 to 23 entries without raising a budget.

Qualification boundary

The following are intentionally not claimed by this release:

• external viewer/operator/admin identity-provider enforcement at the target proxy;
• PostgreSQL selected as the current online system of record;
• redundant object storage or an off-host monitored backup;
• real instruments, robots, mobile platforms, firmware, physical effects, or physical safe-state verification.

Internal role enforcement, PostgreSQL restore mechanics, immutable storage, local backup recovery, and simulated soak are necessary readiness evidence, but they cannot substitute for those target-deployment facts.